KI-ULTRA
Leitfäden
Hier finden Sie die beiden KI-ULTRA Leitfäden. Sie umfassen Handlungsempfehlungen und Referenzwerkzeuge wie z. B. Arbeitsvorlagen für beteiligte Personen im Rahmen der KI-Einführung.
Werfen Sie jetzt einen Blick in die Leitfäden
Leitfaden zu Strategie und Wandel für KI-Einsatz
Leitfaden lesen
Leitfaden zur Durchführung
von KI-Projekten
Leitfaden lesen
Leitfaden zu Strategie und Wandel für KI-Einsatz
Menschenzentrierte Erschließung der Potenziale Künstlicher Intelligenz
Einleitung
Künstliche Intelligenz (KI) wird als ein unternehmerischer Erfolgsfaktor betrachtet, dessen Bedeutung zunimmt. Allerdings setzt der KI-Einsatz ein differenziertes und abgestimmtes Vorgehen voraus. KI-Anwendungen können z. B. dazu beitragen, neue Geschäftsfelder zu erschließen, effizienter zu arbeiten, die Produktqualität zu verbessern, die Kundenbindung zu stärken und die menschliche Arbeit sicherer zu gestalten. Dies fördert eine nachhaltige Entwicklung von Organisationen. Der mit dem KI-Einsatz erfahrungsgemäß einhergehende technologische, organisationale und kulturelle Wandel stellt eine strategische Gestaltungsaufgabe dar, die von allen beteiligten Akteuren in Unternehmen, Verbänden, Politik und Bildungseinrichtungen erhebliche Anstrengungen erfordert. Wenn Fach- und Führungskräfte, Beschäftigte und deren Vertretungen vor Ort den Wandel partnerschaftlich gestalten, werden betriebliche Erfolgspotenziale erschlossen und in der Folge der Wirtschaftsstandort und die soziale Marktwirtschaft gestärkt. In Einzelfällen werden Probleme frühzeitig erkannt und unerwünschte Auswirkungen vermieden.
Die KI-ULTRA Leitfäden
Im Rahmen des vom Bundesministerium für Arbeit und Soziales (BMAS) geförderten Projekts KI-ULTRA (»Unternehmenslabore für Transformation und Change im Zusammenhang mit Künstlicher Intelligenz in der Arbeitswelt«) wurden zusammen mit den partizipierenden Unternehmen, Verwaltungen und Organisationen zwei Leitfäden entwickelt: Der vorliegende Leitfaden zu Strategie und Wandel für den KI-Einsatz und ein weiterer Leitfaden zur Durchführung von KI-Projekten:
1. Leitfaden zur Durchführung von KI-Projekten:
Dieser Leitfaden enthält ein Vorgehensmodell, mit dessen Hilfe Projekte zur Einführung von KI-Anwendungen im Arbeitsumfeld geplant und durchgeführt werden können. Lesen Sie diesen Leitfaden, wenn Sie planen, ein konkretes KI-Entwicklungs- oder Einführungsprojekt umzusetzen.
2. Leitfaden zu Strategie und Wandel für den KI-Einsatz:
Dieser Leitfaden unterstützt dabei, günstige Rahmenbedingungen für den KI-Einsatz über ein (pilothaftes) Einzelprojekt hinaus zu schaffen. Lesen Sie diesen Leitfaden, wenn Sie Ihre Organisation grundsätzlich fit für den Einsatz von KI machen möchten, und Sie KI als betriebliche Querschnittstechnologie etablieren wollen.
Die Trennung zwischen der operativen Durchführung des KI-Einführungsprojekts (»Wie?«) und dem strategisch angelegten Wandel (»Auf welcher Basis? In welchem Kontext?«) erhöht die bersichtlichkeit der Darstellungen. Tatsächlich aber hängen beide Ebenen stark miteinander zusammen: Die übergeordnete Strategie (z. B. ethische Richtlinien) beeinflusst die Aktivitäten in konkreten Projekten.
Introduction
Artificial intelligence (AI) is considered to be an increasingly important success factor for businesses. However, the use of AI requires a differentiated, coordinated approach. The potential for AI applications is great: They can help to develop new business areas, boost efficiency, improve product quality and customer relations, and make human work safer. This promotes sustainable development of organizations. Experience has shown that the technological, organizational and cultural changes associated with the use of AI pose a complex strategic challenge. They require extensive effort from all relevant stakeholders in businesses, associations, government, and educational institutions. If specialists, managers, as well as employees and their local representatives work together to shape the change, organizations can exploit opportunities for success, thereby strengthening the social market economy and country’s position as a business location. In individual cases, problems can be detected at an early stage, avoiding any undesirable effects.
The KI-ULTRA guides
Two guides were developed together with the participating companies, administrative entities and organizations as part of the KI-ULTRA (Company laboratories for transformation and change related to artificial intelligence in the world of work) project funded by the German Federal Ministry of Labor and Social Affairs (BMAS). The first is this Guide to strategy and change toward the use of AI and the second the Guide to the implementation of AI projects:
1. Guide to the implementation of AI projects:
This guide contains a procedural model that can be used to plan and implement projects relating to the introduction of AI applications in the work environment. Read this guide if you are planning to implement a specific AI development project or introduction project.
2. Guide to strategy and change toward the use of AI:
This guide helps to establish favorable general conditions for the use of AI beyond the implementation of an individual (pilot) project. Read this guide if you want to ensure that your organization is fundamentally ready for the use of AI and if you want to establish AI as a cross-sectional technology within your organization.
The distinction between the operational implementation of the AI introduction project (“How?“) and the strategic change (“On what basis? In what context?“) has been made for the purposes of clarity. In reality, however, the two levels are closely interrelated: The overarching strategy (e.g., ethical guidelines) influences activities in specific projects, while the results of specific projects (e.g., the division of functions between humans and AI) influence the strategic dimension, such as the organizational culture.
Die Ergebnisse konkreter Projekte (z. B. die Funktionsteilung zwischen Mensch und KI-Anwendung) beeinflussen zudem die strategische Dimension wie beispielsweise die Organisationskultur.
In den KI-ULTRA Leitfäden benennen wir allgemeine rechtliche Aspekte, die beim Einsatz von KI-Lösungen beachtet werden müssen. Bitte beachten Sie, dass Rechtsfragen in den Leitfäden nur überblicksartig dargestellt werden können. Die Ausführungen stellen keine Rechtsberatung dar und können eine rechtliche Beratung im Einzelfall nicht ersetzen.
In the KI-ULTRA guides, we highlight general legal aspects that must be considered when implementing AI solutions. Please note that the guides only provide an overview of the legal issues. The recommendations do not constitute legal advice and cannot replace this.
About this guide
This guide to strategic change will help you to identify potential applications of artificial intelligence and leverage these on a company-specific basis. It will also detail the challenges associated with the holistic introduction of Al applications. One key focus is the human-centered design of Al-supported work. It is possible to optimize existing business processes or develop novel business models only if humans and artificial intelligence mutually reinforce each other’s capabilities. This guide summarizes various work results and the outcomes of discussions held within the framework of the KI-ULTRA project. However, it only provides basic guidance with respect to relevant issues, interdependencies and operational practices. It is important to reflect on these ideas in the respective business context; experience suggests that it is best to do this in a project team involving experts, management and the works council. Employees should also be involved in the design and implementation processes in order to address their needs and utilize their practical knowledge, thereby ensuring the success of the change process.
Such discussions with experts and stakeholders will allow you to make an informed judgement on the extent to which the use of artificial intelligence will affect the operational business model and corporate strategy. This applies in particular to legal issues, which must be examined on a case-by-case basis before any Al is used. This guide provides an overview of such issues. However, this guide is not a set of step-by-step instructions to be followed in a specific sequence. Instead, it is divided into six action areas that you can implement separately:
- Operational Al strategy
- Human-centered design of changes
- Work design strategies
- Social conditions
- Skills development and learning
- Data strategy and technology transition.
The “Operational Al strategy” action area provides the general basis for change management. This guide describes the topic areas that are particularly important based on project experience and in which the solutions are not yet generally known — at least with respect to the application of Al.
MG – management, e.g., general management, team leaders, etc.
EM – employees
OS – occupational safety specialist, work designer (including productivity and quality management)
SE – IT security specialist
Al – Al experts, people who deal with data, algorithms and models
HR – Human Resources
WC – works council
Target audience
This guide is aimed at private-sector companies, non-governmental organizations (NGOs) and public institutions that intend to implement and use artificial intelligence applications. The main target audience is comprised of decision-makers at the level of operational management, project management and employee representation. In addition, the guide is aimed at all stakeholders who work on the development and application of artificial intelligence in a work context. Icons are used throughout the document to indicate which action aids are of particular interest to which stakeholder group. Experience shows that effective, sustainable solutions are developed through dialog. This enables a wide range of perspectives to be considered. Speak to experts or join inter-company working groups.
Reading guide
This guide is divided into six action areas. Each action area comprises several action aids that include clear and specific instructions for a particular topic in order to improve your decision-making ability or ability to act. In some cases, however, the guide can only give you general advice and ask questions that you will need to consider if you are to find suitable answers for your organization. In other cases, you should bring in external consultants to help you successfully apply the methods. The action aids include the following elements:
Title
For example: Defining strategic goals
Icons
Description
The description is a text that briefly illustrates the problem and solution.
Guiding questions
The guiding questions provide additional help and supplement the description. Try to answer the guiding questions yourself. The answers may vary depending on the organization and industry.
Measures and tools
Where possible, we refer to methods and tools developed by us or by third parties that you can use to answer the guiding questions. These methods are freely available.
When is it worth getting involved with AI?
What do we mean by “artificial intelligence”?
There is no clear and comprehensive definition of “artificial intelligence.” In this guide, the term refers to technologies that can imitate certain types of behavior that would be impossible for conventional software algorithms. We use the term to refer to standardized Al applications that have a narrow scope in terms of function (e.g., text or voice recognition methods) as well as solutions that need to be developed or adapted for specific applications.
The European Commission understands that the use of artificial intelligence can contribute to improving forecasts, optimizing operations and resource allocation, and personalizing service delivery. This can support socially and environmentally beneficial outcomes and provide key competitive advantages to companies and the economy.
From a legal perspective, there is also no clear-cut definition of “Al.” The draft of a regulation on artificial intelligence at EU level (EU Al Act), which was discussed in the EU Parliament on June 14, 2023, provides a broad definition of Al systems; Article 3 (1) defines such a system as “a machine-based system that is designed to operate with varying levels of autonomy and that can, for explicit or implicit objectives, generate outputs such as predictions, recommendations or decisions that influence the physical or virtual environments.” Since it can be assumed that the expected entry into force of the EU AI Act will have a significant impact on the use of Al, it is important to keep a close eye on the legislative process. It is likely that Al systems will be governed following a risk-based approach similar to that of the General Data Protection Regulation (GDPR), whereby the requirements for creating and using Al systems will gradually become more rigorous as the level of risk associated with the use of the respective Al system increases.
Artificial intelligence (Al) is, by definition, a sequence of mathematical models running on computing machines. With the help of special interfaces, these models record human and machine behavior using extensive data in order to mathematically process and imitate it. In the training process, the data patterns are submitted to experienced Al instructors who determine how the artificial intelligence should respond to an input. Once trained, the Al machine uses its heuristic set of methods to respond to data patterns that it identifies by itself. Al is therefore capable of handling human tasks with a level of adequacy that is at least similar to that of humans.
The potential of using AI
The action aids documented here should provide you with a better understanding of why and how you might use Al in your organization, whether in a pilot project or on a larger scale. This requires a basic understanding of the possibilities and limitations of Al applications. The use of Al should help to provide a better solution to a specific problem in an organization. Practical examples offer guidance; these can be found on the KI-ULTRA website or in the knowledge pool of the “Zukunftszentren”, for example.
We recommend a human-centered design approach for Al in order to improve its effectiveness when it is deployed in practice. The human brain and artificial intelligence should not be pitted against each other but rather used in tandem. No Al system can compete with the ease and efficiency with which humans learn, understand and act. Humans have cognitive skills that also include creative inspiration and human empathy. For example, they can differentiate between animate and inanimate objects. They can pick up new terms after hearing them just once or twice, and learn by trial and error.
Conversely, artificial intelligence is capable of many things that humans simply cannot do or only with great difficulty. It can identify patterns in huge data volumes, manage complex manufacturing processes, analyze satellite images to help farmers increase their crop yields, detect financial fraud and determine user preferences. More importantly, however, Al can maximize productivity within the process of dividing tasks between humans and machines.
Compared to other digitalization solutions, Al solutions cannot usually be bought “off the shelf.” They often need to be adapted to the organization in question. For this reason, using Al (compared to other technologies) often takes more time and costs more money before the technology can show its strengths. Its use is also associated with significant uncertainty: It is often only possible to identify whether a project will be a success at a late stage. A process of trial and error is required in order to determine which approach will achieve the objective.
Al can help to achieve the following goals:
- offering new services or products, or achieving better quality than the competition (e.g., targeted product suggestions for existing customers; simplifying predictive machine maintenance);
- making processes more efficient by automating sub-steps;
- relieving employees of monotonous, menial tasks;
- providing targeted support for inexperienced employees so that a larger number of applicants can be considered for a role in view of the (regional) shortage of skilled workers.
Al systems open up new potential for assistance and automation. If you are facing the challenge of making business processes in your organization more flexible, then engaging with the topic of Al can open up some new perspectives and complement other organizational approaches for improvement. In addition to these specific benefits, Al can provide fundamental value by supporting operational decision-making in turbulent markets (e.g., demand-based production planning) on the basis of extensive datasets.
To be able to make the most of Al and obtain meaningful results, companies first need a high-quality database. Al systems can, however, work with incomplete datasets, provided that certain compromises are accepted in terms of the quality of the results. This makes the technical system highly adaptable to changing conditions.
The limitations of using AI
The limitations of using Al systems mainly result from the existing technological possibilities, regulatory frameworks and what is deemed acceptable by employees. For example, Al systems have a limited capacity to imitate emotional skills or complex thought processes, such as when selecting employees or resolving social conflicts:
- Al systems have no emotions. This means that they are incapable of acting with empathy.
- Al systems have no capacity to make moral judgments. The ability to make complex decisions on moral grounds cannot be simulated using machines. Only humans can make moral decisions.
- Al systems have no awareness of problems: Although they are capable of imitating certain thought processes, they have no underlying rational understanding, intuition or insight.
Al systems must only be used within the scope of the relevant legal framework. Labor and data protection laws as well as the (future) Al Act may restrict the use of Al. When the EU Al Act enters into force, certain Al practices will be banned; high-risk Al systems will have to meet certain requirements. The use of an Al system might depend on the approval of employee representatives, i.e., works council or staff council (see “Human-centered development of Al solutions” below). As a result, the use of Al will be restricted in cases where, for example, the changes to working conditions or performance monitoring would be unfavorable for employees, thereby precluding approval.
Al should not make critical decisions autonomously that could affect people’s lives (e.g., customers, patients), such as a medical diagnosis. Rather, Al should assist the people who are making the decisions. It should also be noted that Article 22 (1) GDPR prohibits automated decision-making that produces legal effects concerning the data subject or that similarly significantly affects the data subject (e.g., economic and social effects). This prohibition is subject to a few exceptions (Article 22 (2) GDPR). For example, if an Al application is used to identify unprofitable contracts that are to be terminated, the decision must not be made immediately but only after (decision-relevant) human interaction and in line with transparent criteria (based on the automated decision) in order to comply with data protection law. It is therefore important that such human interaction is not just a formality and a person actually influences the content of the decision.
If your intended application comes up against these limitations, there are two options:
- You realize that the use of Al will not make a significant contribution to achieving your organization’s goals. As such, there is no need for further action in this area.
- You aim for a holistic design approach that takes into account human skills and contributions. In this way, any possible unwanted side effects resulting from the use of Al can be offset and the skills of employees can be enhanced through the use of Al.
Legal requirements for the use of AI
In addition to the technical and organizational limitations of the use of Al, it is important to always bear in mind that the specific use of Al systems must be in line with the applicable legal provisions (which are currently undergoing some substantial changes), taking into consideration not only the EU Al Act previously mentioned. Particular attention should be paid to the following areas in the strategic planning of the use of Al:
General use of Al:
- rights clearance in the procurement of training data
- use of the Al by the user / monitoring data entry (e.g., data protection law, trade secrets)
Decision making:
- labor law
- data protection legislation
- general right of personality
Productive use of Al
- protectability of data produced using Al
- statutory intellectual property rights to the data
- patents / utility models
- design (in the legal sense)
- copyrights / ancillary copyrights
- contractual data protection
The following sections will provide a basic overview of more specific legal restrictions or requirements. This information will not be separated from the rest of the text but integrated at the relevant points. Due to the wide variety of possible uses of Al, the legal information provided herein is not exhaustive and is not a substitute for individual legal advice. As such, you should always clarify individual questions with legal experts.
Action area: Operational Al strategy
Defining strategic goals
If your company is considering introducing Al, you need to establish in advance what you expect to gain from the use of Al, what projects would be suitable as pilot projects and whether your workforce has the necessary expertise for the long term. If you are faced with your very first Al project, it will probably seem very daunting to you. Once you know what you will need to focus on during this process, you will see the project in a new light.
Developing an Al strategy is a good idea if any of the following points apply to your company:
- You are planning to introduce Al applications on a large scale.
- Different departments will need to collaborate to implement the planned use of Al.
- You have secured the support of relevant decision-makers (e.g., management, data protection officer, other stakeholders) to ensure that the use of Al is a success (e.g., for launching follow-up projects).
In the Al strategy, you document the way in which the use of Al applications is intended to help achieve the operational goals. This highlights the significance of the Al applications for all stakeholders.
The Al strategy always starts with the organization’s overarching vision. The vision describes a desirable future state of the organization. The vision should clearly show how the organization’s core business relates to its market environment, for example, customers, suppliers and competitors (see Figure 2); from experience, a vision that is purely inwardly focused (“Our administrative processes are particularly lean”) can lead to bad decisions. Once the vision has been formulated, it is linked to the organization’s strategic goals and the contribution made by Al applications.
Rather than being specified at management level, the Al strategy should be drawn up in a participative decision-making process. In other words, all of the relevant stakeholders should be involved in drawing up the strategy (i.e., business management, managers of specialist departments, domain experts, Al experts, works council, affected employees). The Al strategy serves as a basis for change management (see “Change management as a means of harnessing the potential of Al” below): The strategic measures are developed based on the target state documented in the Al strategy. The process of drawing up the Al strategy therefore helps to identify the importance of organizational and cultural changes (beyond the purely technical or process-related aspects) for the introduction of Al.
These measures can be better coordinated between departments or areas of responsibility if all the parties involved understand the goal behind each measure and how important the goal is to the organization as a whole. If the Al strategy links the measures to the organizational goals and is developed in a joint process, the parties involved will be more willing to support the measures. Even if a pilot project is canceled because it is unlikely to succeed, the Al strategy serves to ensure that the corporate goals in the field of Al will be achieved in the medium to long term. The joint strategy discussion may highlight which forms of Al expertise are already available in the company and where there are gaps that need to be filled.
Smaller organizations may be able to list some relevant Al applications straight away (see “Systematic search for Al use cases” below); for larger organizations, it may be advisable to clarify the following questions:
- What are our commercial objectives with regard to using Al (e.g., improving market position, reducing costs, gaining the ability to cooperate with customers)?
- How does the Al application help to achieve these objectives and how will the achievement of objectives be tracked (e.g., exploiting new market potential, effects of automation)?
It is important to clarify the operational benefit of using Al. One of the first questions you should ask is: “Why do we think that this investment will pay off?” To be able to answer this question, you need to know which problems you want to tackle (e.g., bottlenecks in the operational process) or where you want to improve efficiency and innovative strength (such as in the digitalization of distribution channels). Al projects should focus on processes that have a tangible impact on expenditure and resources and that significantly influence operational profitability.
The importance of the strategic use of Al for the organization is illustrated by the example of three levels:
- Level 1: Al potential is (currently) of minor importance to the success of the organization. The core of the strategy is therefore to drive digitalization in order to make efficiency gains and, where applicable, to establish better conditions for the use of Al in the future.
- Level 2: Al is used to make improvements in individual application areas (e.g., in production and assembly); all other possible fields of application (e.g., customer relations management) continue to operate as before.
- Level 3: The future business model is largely based on the use of Al; the widespread implementation of Al is seen as a critical success factor.
This classification can also be helpful when answering the question “Make or buy?” (see “Strategic approach to ‘make or buy'” below). The more important Al is for the core business, the more worthwhile it is for the organization to establish its own Al expertise. You will need to answer the question “Make or buy?” for each individual Al project, but it is advisable to specify a strategic direction.
Guiding questions
- What overall benefit do you expect Al applications to bring to the organization?
- What strategic goals are to be achieved through the use of Al applications?
- What specific tasks do you want to address to gain experience with Al? Where do you want to improve efficiency or innovative strength, for example?
- What types of Al applications can be considered in your organization and to what extent?
- What expertise should be available for the introduction and implementation of Al applications?
- Should Al applications be purchased or developed in-house (over the long term)? (See “Strategic approach to ‘make or buy'” below)
- How will you determine whether Al is actually helping to solve pressing problems? Under what conditions will Al projects be canceled?
- How can it be ensured that the use of Al will have a positive impact on working people and their circumstances? Will jobs be lost?
- Is there a need to obtain further qualifications? Will the way employees communicate change?
- How will the Al strategy be implemented in a way that complies with data protection law?
- How will it be ensured that all of the relevant stakeholders are involved?
- Which core business processes will be affected by the planned use of Al?
- Will the introduction of Al involve a change to the business organization that requires co-determination?
- What resources will be made available for implementing the Al strategy?
- What will the available time frame be for doing this?
Measures and tools
- With the involvement of relevant stakeholders, define the contribution made by Al to achieving the vision or goals of the organization.
- Involve a high-ranking decision-maker who will support the implementation of the Al strategy and enable it to be implemented by approving the use of resources (e.g., project funds).
- Include the works council or staff council at an early stage to minimize potential fears and ensure timely communication for maximum transparency.
- Involve the data protection officer and, if necessary, schedule a data protection impact assessment from the start.
- Involve the legal department to ensure that processes are legally compliant.
- Review the vision and strategy on a regular basis and adapt them as necessary.
- Visualize and communicate the Al strategy using a strategy map (see Figure 2).
- When formulating a corporate strategy, an external coach can broaden perspectives, reflect on the discussion and ask critical questions to ensure a better understanding.
Determining the transformation requirements
In order to develop change measures based on the Al strategy, it is a good idea to first take stock of the current situation, i.e., which favorable conditions for harnessing the potential of Al already exist and which do not. The KI-ULTRA Evaluation Toolkit³ is available as a guiding evaluation instrument. The purpose of the evaluation is to identify the strategic action areas with the greatest potential for improvement. Each of the result fields in the Evaluation Toolkit links to relevant action aids in this document, which can be used to improve the conditions for the use of Al. This guide therefore constitutes a maturity model for the use of Al; other maturity models can be used too.
Maturity evaluations primarily refer to the following main topics:
- Organization and work methods
- Infrastructure
- Datasets and data availability
- Culture and collaboration
- Ethical, legal and social implications (ELSI).
Guiding questions
- In light of the general conditions, is your company currently capable of addressing the topic of Al in an operational context?
- Are the appropriate resources and expertise available?
- Has upper management made a clear commitment to implementing Al?
- What legal issues need to be clarified? Is the necessary legal expertise available or has it been acquired? What needs to be done to establish the necessary general conditions?
- What needs to be done to establish the necessary general conditions?
- Who needs to be involved if the general conditions in the individual action areas are to be improved?
- Which works council co-determination rights require consideration? Is there a plan of work procedures or processes that the works council must be notified of and consulted about?
Measures and tools
- KI-ULTRA Evaluation Toolkit
https://www.ergebnisse.ki-ultra.iao.fraunhofer.de
Strategic approach to “make or buy”
Should you purchase your Al applications or develop them yourself? This question does not need to be addressed straight away because you will probably bring in external support for your first Al project. However, the question of whether to make or buy will arise in the medium term if Al is to be used on a larger scale or if Al solutions are likely to be a unique selling point of the organization. When making this decision, it is important to consider the guiding questions listed below. However, the weighting of these aspects in the decision-making process should be defined specifically for the organization in question on the basis of the Al strategy.
Guiding questions
- Are appropriate skills and structures (e.g., data scientists, data engineers, application designers) available so that a realistic estimate of the level of work and the costs for Al development can be made and assigned to a cost center?
- Which alternative promises to better incorporate proven findings from the field of industrial science in the human-centered work design?
- Which option (purchasing Al or developing it in-house) will better address any possible concerns of the works council regarding the impact on employees?
- If a solution is purchased, does this create a unilateral dependence on an external service provider, for example, due to migration of expertise or disclosure of datasets?
- What are the follow-up costs when purchasing a solution from external service providers (e.g., updates, maintenance)?
- Is the service provider known to be reliable?
- Can you trust that they will respond quickly to any problems and charge a reasonable price for this?
- Are there GDPR-compliant solutions for the tasks to be handled? What agreements can be made with the service provider in terms of responsibility for the datasets (e.g., fulfilling information requirements, conducting data protection impact assessments, ensuring the technical security of the datasets, etc.)?
- Where will the data be processed?
Systematic search for AI use cases
The procedural model for project implementation starts with a defined use cases for Al. A key question for many organizations is how they can identify appropriate use cases in the first place in order to pick low-hanging fruit.
The following are suitable candidates:
- activities that are time- and labor-intensive (such as reading extensive documentation to identify where action is required).
- work processes that require complex image recognition.
- processes that can be complemented by speech analysis (such as in customer service).
- predictions about customer behavior in sectors such as insurance, banking or advertising.
The larger the organization and the more diverse the business processes, the more important it becomes to incorporate expert knowledge from the specialist departments via appropriate processes and structures. Therefore, the search for use cases should be approached as a cross-departmental strategic measure.
There are two possible approaches:
- based on problems identified in specialist departments or by technical specialists (e.g., employees, quality assurance and scheduling)
- based on the potential of the Al application (i.e., identified by people with technological expertise, people with Al expertise)
The following gives a brief overview of the problem-induced approach and the technology-induced approach:
Approach 1: Identifying AI technology for a given problem
The problem-oriented approach (“market pull”) ensures that the specific need for an Al application is identified at an early stage, which can simplify the subsequent process. Solutions can often be found in issues already identified in a business process. For example, the existing issue could be the catalyst to contact an internal specialist department or service provider to examine the use of Al. The availability of any potentially suitable Al solutions in the company, which could be adapted for the given issue, should also be assessed.
Guiding questions
- Have Al technologies already been used for similar tasks? If so, which ones?
- (Are there any known comparable applications from other organizations?)
- In what way is Al better or different compared to previous technologies? What is the difference?
- Will the use of Al improve work processes and workflows for the benefit of the employees?
- Which optimization goals will be achieved by using Al technologies?
- Which business process data is available in appropriate quality? Which additional data has to be collected?
- Under what conditions can the chosen technology be used for the known problem in compliance with the law?
- Taking into account all intellectual property rights (see “Minimizing legal risks” below), is suitable training data available?
Approach 2: Identifying a problem suitable for AI technology
Alternatively, you can check whether any business process issues can be optimized through Al technologies and, if so, which ones (data-driven approach, “technology push”). Is it possible to tap into new business potential using the availability of (large) datasets? For example, is it possible to identify customer needs more effectively and produce target group-specific products? Can the purchasing behavior trends be more easily predicted, such as over Christmas? Can the wear behavior of machinery be more easily predicted inside the organization and can maintenance be planned more efficiently?
In this approach, people with Al expertise should continuously identify and evaluate the potential benefits of Al technologies. During concrete projects, they then define requirements for promising technologies. These should be specific but not limit the potential solution in any way. A sample question could be how to automate quality assurance through image recognition. However, it is important to bear in mind that Al will only ever improve individual phases of a process; it is therefore difficult to calculate how it will impact the overall result.
Guiding questions
- What competitive advantage would it offer the organization if recurring patterns could be detected in a process, if it was possible to predict process behavior, if statistical methods could be used to support decision-making processes?
- Are there any recurring, essential tasks that require a lot of time or effort?
- Are there any tasks where an analysis of large data volumes would reduce mistakes?
- Can you demonstrate what Al is capable of (and what it is not) for specific applications?
- Are there areas in which work processes and workflows could be improved for the benefit of employees?
For most organizations, Al is a means to an end and not an end in itself. In other words, its main purpose is to solve known problems or achieve improvements. Al should be used when it is the most promising of all the solutions available. Remember: It usually takes some time before the impact of a new technology on a work system becomes apparent.
Measures and tools
Companies can use their suggestion scheme to bring together problems and Al technology. Topic-specific workshops have also proven to be useful. If data scientists and data experts work in the company, you should consider forming an “Al hub.” This would do the following:
- Promote the exchange of expert knowledge and expertise between different parts of the company, for example, to obtain tips and help when facing problems in a project or to share the latest technical information (e.g., new technologies, discussion of experiences).
- Members of the hub can approach specialist departments directly to discuss daily working routines and identify areas with potential to remove tasks or provide support.
- Offer Al training for employees so that they gain basic knowledge and core areas of expertise in relation to Al.
The following aspects are critical to the success of Al hubs, etc.:
- providing resources (e.g., digital communication platform, internal forum)
- providing sufficient capacities
- Internal communications: Employees should be familiar with the hub, its members and responsibilities.
AI map: This lists AI providers, users and development projects in the form of a map and in a list. It helps to spark ideas.4
Please note: If an Al hub is set up, this does not replace measures for employee involvement, such as open feedback discussions about the implementation of Al in the organization. Irrespective of the existence of an Al hub, co-determination rights of employee representatives must be respected (in particular the works council’s right to be kept informed and consulted in the planning of work procedures and work processes). Through training measures on the basics of Al, it should be ensured that the people involved can evaluate the extent to which Al would actually be useful for solving a problem.
AI map: https://cta4.plattform-lernende-systeme.de/map-on-ai-map.html
Action area: Human-centered design of changes
Change management as a means of harnessing the potential of AI
The introduction of Al applications may require operational changes in an organization. As part of change management, these changes are implemented in a targeted way and affected employees are involved at an early stage so that the organization adapts to the new requirements. If a fundamental change is required, an agreement must be reached with the works council to reconcile their interests; this process might involve drawing up a social compensation plan. Employee transfers must be covered by the employment contract and are only permitted with the approval of the works council.
If you are initially planning an Al pilot project to gain experience, then the Guide to the implementation of Al projects provides a basis for ensuring that the most important change requirements are taken into account over the course of the introduction project. In relation to change management and with regard to a specific Al application, you can also use the “Introduction of Al Systems in Companies” guide from “Plattform Lernende Systeme.”
The following procedural guide is aimed at organizations that want to fully harness the potential of Al and therefore implement Al as a cross-sectional technology. This requires different areas of the company (e.g., HR, production and customer services) to work together.
The starting point for change management is the Al strategy (see “Defining strategic goals” above). It sets out what contribution Al makes to the success of the organization and how this contribution is created. The objectives that are relevant to the use of Al are derived from the following strategic action areas (work design, competence management and learning, data strategy) and the corresponding procedural guides.
Further objectives may include:
- strengthening the role of humans as experts (in the sense of dividing functions between humans and Al in a complementary manner);
- further developing the organizational structure;
- creating new jobs or departments to pool Al activities;
- providing infrastructure
Humans and artificial intelligence complement each other. No Al-controlled machine can compete with the ease and efficiency with which humans learn, understand and identify correlations. Conversely, Al is capable of many things that would be difficult for even the most talented person: identifying patterns in large data volumes, supporting decision-making processes, managing data, etc. When humans and Al collaborate, data can be used for sophisticated and differentiated purposes. However, the use of Al requires a high level of commitment from employees. This is the focus and objective of human-centered change management.
Guiding questions
Answer the following questions within the scope of change management:
- Has a development goal been defined that recognizes the potential benefits of human and Al collaboration?
- Does the planned use of Al result in changes to the business model that require changes in the organization?
- Does an operational change require co-determination?
- Does communication help informed decisions to be made within a reasonable time frame in order to utilize the benefits of Al?
For example, you may be using Al in different departments to support corporate decision-making (e.g., predicting demand and adapting production, pricing products and services). However, the benefits of the technology fizzle out because decision-making channels are long and involve different levels of the hierarchy. One solution may be to introduce interdepartmental teams so that decision-making is accelerated.
- Is our working method suitable for the planned introduction of Al? How can we accelerate the development and introduction of Al while respecting the co-determination rights of employee representatives?
Al development and introduction projects usually involve a great deal of uncertainty. At the start of a project, it is unclear what the application should look like in the end and how complicated the development process will be. Agile working methods can help to complete Al projects faster and ensure they are a success, or to ensure that unprofitable and unfeasible projects are wound up quickly. It is not necessary for the entire organization to transition to agile working methods right away. As long as the Al project teams include members with experience of agile working methods, then, for example, employees from specialist departments that employ lean management or conventional project structures can adapt to agile working methods effectively. When applying an agile approach, the continuously changing requirements must be legally reassessed in order to avoid unnecessary expenditure in unusable Al functionalities.
- Do we need to define new roles or business processes?
Where personal data is being handled, it is necessary, for example, to appoint a data protection officer. If the intention is for data to be accessible across departments, processes should be defined to ensure that this access is compliant with data protection legislation.
The evaluation should take account of the relevant human-related impact as well as purely functional target criteria: How does the change affect employees’ understanding of their role and their level of satisfaction? Are any particular people in the organization placed at a disadvantage by the change? Are findings on human-centered work design sufficiently considered? Are any particular people in the organization placed at a disadvantage by the change? Does any change result in prohibited discrimination? Does a social compensation plan need to be negotiated in order to compensate for disadvantages related to an operational change?
Human-centered development of AI solutions
When a human-centered approach is used to design systems, the focus is placed on the actual users and other affected parties, actively including them in the design process.
The number of decisions that a company needs to make is increasing. Key decisions require problem-solving, analysis, strategic thinking or the ability to prioritize. People must be able to evaluate the potential impact of such decisions and to weigh up alternatives. Al systems can be of great benefit in this respect. If data required to make decisions is shared on platforms, more people can be involved in decision-making processes, thereby increasing their acceptance.
The following decision-making scenarios are possible:
- Several Al applications ensure that the final expert opinion of decision-makers is given greater weight. One example is the use of imaging procedures in medicine.
- In other cases, Al bundles the decision-making process without changing the decision-makers. Personnel selection is one example.
- Several Al applications centralize the decision-making process. Credit card validation is a typical example.
AI projects are only effective if they respect employees’ decision-making authority. A human-centered approach to Al design fosters a strong, mutual relationship between humans and machines, combining their competences.
The effects of Al on employees can be difficult to predict. For this reason, we recommend following a consistent human-centered approach when introducing Al: As a rule, the affected employees and those responsible for representing their interests should be involved in change processes. If the works council needs to assess the introduction or application of Al in order to perform its duties, it can appoint an expert.
The purpose of the human-centered approach is to incorporate the knowledge and creativity of employees in the early stages of the design process. This ensures that the resulting Al applications and work processes result in a noticeable improvement for employees. This approach promotes practical knowledge, motivation and identification with the task and the technology.
Employee representatives (works council/staff council) may have co-determination rights relating to the introduction of Al systems. Software is often a technical tool that can be used to monitor the behavior or performance of employees, which is why the works council must co-determine its introduction and use. If Al leads to a fundamental alteration of work organization, this constitutes a work alteration that requires co-determination. In this case, an agreement on a reconciliation of interests should be reached and a social compensation plan may need to be drawn up. Even at the planning stage for applying Al to work procedures and processes, the works council must be consulted and its recommendations and concerns taken into account. If the work situation for employees is significantly altered, this may constitute a transfer that requires approval by the works council. If Al is used to draw up guidelines on employee selection in recruitment, transfers, pay group reassignment or even termination, the works council must be involved.
To ensure that employees can actively participate in the design process, appropriate conditions must be created. Basic knowledge of Al (see “Making Al knowledge available” below) must first be acquired so that everyone involved has the same understanding and uses the correct terms when discussing possible solutions and approaches. Basic knowledge of Al helps us to reflect on possible reservations about Al applications. In particular, the employees’ understanding of their own professional role needs to be taken into account: Does the new, altered task fit with this self-image? For example, how will employees’ decision-making authority, flexibility with regard to the completion of their own tasks and interaction with colleagues change?
Guiding questions
- Which employee skills are indispensable for customer focus and business success? How could Al be used to complement these in an appropriate way?
- What attitudes do employees have towards Al and what are the experiences and assumptions underlying these attitudes?
- How can we create a basic understanding of Al across the whole organization?
- Which formats and media are suitable for providing internal information on the topic of AI?
- What experience does the company have of involving employees and their representatives? How important is it in relation to the organizational concept? Are there agreements with the works council on the introduction and application of programs?
- Employee participation requires company rules. This is the only way to ensure that accountability, trust and responsibility form the basis of daily actions. To what extent is this taken into account in the organizational culture?
- Is it assumed that affected parties and employees (data subjects) consent to their data being processed (insofar as this is necessary)? If personal data is to be processed: Is it ensured that there is a legal basis for processing this data? Can the data processing be based on the valid, voluntary consent of the respective data subject? How can the data subjects be clearly informed in advance about what will happen to their data and for what purpose? (See “Protecting personal data and other sensitive data” below; see “Data access” in the Guide to the implementation of Al projects.)
Measures and tools
- Make information on the planned use of Al transparent and publicly accessible.
- Facilitate co-determination of the works council.
- Discuss the planned use of Al with the works council at an early stage.
- Establish communication formats for all affected employees.
Action area: Work design strategies
Despite the use of innovative technology, human contribution will continue to play a decisive role in shaping work in the future. A human-centered design approach creates favorable conditions for ensuring that people are motivated and qualified to work and learn. The purpose of human-centered work design is to combine the potential for human achievement with the functional potential of technology. The way in which this is implemented in an operational sense is always decided in the specific project at hand. However, you should reflect on several principles at the strategic design level; this is the only way to ensure that Al applications are selected appropriately and implemented effectively.
Incorporating human and machine capabilities
The way in which functions are appropriately divided between humans and machines is decided on a case-by-case basis at an operational project level. However, you should create the necessary conditions at a strategic level. The introduction highlighted the potential and the limitations of Al applications (see “The potential of using AI”). You should also consider the strengths of humans when compared with Al systems, for example, as shown in the model of the missing middle (see Figure 3).
In principle, Al has the ability to replace human work (i.e., automation) and to provide assistance and augmentation (i.e., to enhance human skills). A central task in the division of functions is to assign individual tasks to technology or to humans. Here, it is important to ensure that humans end up with enough scope for actions and decisions. Together with the requirement level of the activity, this scope determines which human learning and adaptation capabilities are to be utilized so that the parts of work that cannot be planned are managed with flexibility. A division of functions based on human strengths creates favorable conditions for an agile working method.
In particular, Al applications should not be used to make decisions that have a direct impact on people; instead, such decisions should require interpersonal negotiation processes. Examples include shift or vacation planning. In many areas, using Al solutions to (completely) automate decision-making processes is only permitted by law to a very limited extent. In addition, Al applications should not be used to monitor and control human behavior; this approach is known as New Taylorism. This is where IT systems specify ideal conduct for people (“one best way”) and monitor compliance. No evidence has been found to indicate that this optimization approach actually results in any improvements. However, experience suggests that viewing human performance primarily as a technical challenge results in many problems. Instead, work design should foster human capabilities, for example through understanding and meaningful engagement.
In Al systems, the interaction between humans and machines must be designed in such a way that the work content and communication tasks that motivate people and promote learning do not end up being replaced. This may occur when, for example, interesting programming tasks are reassigned to a robot, but employees are still responsible for routinely inserting workpieces into a punch machine. It is imperative to ensure that the interaction between humans and machines is designed in a transparent manner. This task is specified in the model of the “missing middle.”
![Figure 3: Model of the missing middle [2]](https://websites.fraunhofer.de/ergebnisse-ki-ultra/wp-content/uploads/2025/04/2025-04-30-11_32_54-2023_Leuteritz_Leitfaden_zu_Strategie_fuer_den_KI-Einsatz.docx-Kompatibilita.png)
Guiding questions
- Which tasks can be executed reliably by people (and thus assigned to them)? How will automation in the form of Al impact the potential for people to gain experience, as well as the diversity of tasks available and the motivation level of working people?
- What are the legal requirements and the ethical aspects of using Al? Is there a risk model with suitable criticality levels to identify and minimize potential risks of the Al application? Which co-determination rights of employee representatives require consideration?
- Will decisions be made based on automated processing that is not subject to human control?
- What functions can generally be assigned to Al applications? To what extent can they intervene in the social fabric of the company? What impact does this have on the Al strategy?
- Does Al actually offer enough flexibility to adapt as necessary to the needs of the company and of people? Or would conventional automation methods be sufficient?
- If Al is replacing human work: Does a social compensation plan need to be drawn up? Are transfers required?
Measures and tools
- Define how functions will be divided between humans and Al, and coordinate with all people involved.
- Provide information on strategic frameworks for all developers.
- Derive measures for human-centered work design. You can make use of the appropriate literature for this purpose ([2], [3], [4], [5]).
- In smaller companies, an employee workshop can be held for discussion of the extent to which working conditions are in line with the design goals and what changes need to be pursued as part of a joint process.
Minimizing safety risks
Technical systems must be safely and reliably implemented as a matter of principle; relevant safety strategies are governed in laws, regulations and standards, for example. In the context of high-risk robotic Al systems, accident prevention is of paramount importance. On the one hand, accidents can be traced back to technical parameters. On the other hand, incorrect human behavior must be avoided – this can be more common if tasks are monotonous or cause mental exhaustion.
When implementing Al applications, the legal requirements pertaining to occupational safety must be taken into account. In particular, a comprehensive risk assessment may be required and the works council must be involved in matters of occupational health and safety. It is equally important to take into account the risk-based regulation of Al systems expected to take effect with the entry into force of the EU AI Act.
A fundamental criterion of human-centered design relates to the absence of harm in the interaction between people and machines. The safety risks of a machine must be reduced to a generally accepted level. Safety controllers play a significant role in this safety concept.
In Al systems, it is no longer possible to understand algorithms and decision-making processes in detail (i.e., “black box” behavior). As a consequence, people do not always judge processes correctly and sometimes make inappropriate interventions. Therefore, overarching safety functions and emergency strategies should always be in place, too. The extent to which Al algorithms can make safe decisions is generally not known in sufficient detail. For this reason, the safety level of Al systems should be verified reliably on the basis of strong indicators. This includes criteria such as the explainability and comprehensibility of decisions, applicability to different situations, monitoring during operating time and transparent forms of human-machine interaction.
A risk assessment (as would be implemented at an operational level) usually forms the basis of a safe design. At a strategic level, the structural requirements for occupational safety must be met by involving qualified personnel, providing testing equipment or training employees. In individual cases, failure to follow the appropriate provisions may mean that Al applications need to be put out of service.
The Data Ethics Commission (“Datenethikkommission” or DEK) of the German federal government has developed a criticality pyramid that evaluates Al-based decision systems and considers the probability and severity of harm. It defines requirements and control methods and offers guidance for assessing the general conditions for Al applications. In the future, the “positive list” of high-risk Al systems in the upcoming EU AI Act will also need to be taken into account.
Guiding questions
- What safety risks need to be considered when using Al?
- Does the product safety and operational safety of the Al system comply with the applicable laws?
- How should operational occupational safety be organized when using Al systems?
- Do employee representatives need to be involved due to matters of occupational health and safety?
Measures and tools
- Identify the applicable laws and regulations. In particular, check whether the Al systems to be implemented meet the legal requirements in terms of product safety.
- Check whether existing insurance policies cover the use of Al systems in the specific application area or whether additional insurance cover is required.
- Establish a risk assessment process (information is provided in, e.g., the guidelines of the German Federal Institute for Occupational Safety and Health (BAuA)).
Protecting personal data and other sensitive data
The key principle of all Al systems is to accumulate and analyze large quantities of data. The process of recording, storing and analyzing this data is becoming more and more efficient. In a work context, data analyses, inter alia, can be used to prevent errors – but also for monitoring and performance control. If personal data is collected and analyzed, the works council must be consulted.
When developing and using Al, compliance with the laws on the protection of employees’ personal rights is required, from the GDPR to labor law, in particular the co-determination rights of employee representatives. Furthermore, personal data of third parties (e.g., customers) may only be processed in accordance with the applicable data protection requirements.
To determine the legal requirements, it is necessary to differentiate between the different data types and the data subjects whose data is to be processed in specific individual cases during training of the Al and during its subsequent practical use. If personal data is processed, there must be a legal basis for doing so. Personal data refers to all information relating to an identified or identifiable natural person. This includes, for example, the name, address and contact details of employees or other persons as well as information that can be combined with other data to identify a person, such as an IP address or employee number. Processing this data is only permitted if there is an explicit legal basis for this or if consent has been obtained from the data subject. The latter is often critical in an employment relationship as there are strict requirements regarding the effectiveness of voluntary consent due to the hierarchical relationship between employer and employee.
Even stricter rules apply to datasets comprising special categories of personal data (known as “sensitive data”). With regard to employees, this includes data on ethnic origin, religious beliefs, trade union membership and health as well as genetic and biometric data. Processing this data is prohibited unless one of the very limited exceptions stipulated in the GDPR applies.
In addition, it is important to take into account general data protection requirements as early as possible during project planning. In particular, the data subjects must be informed of the processing operations. If data was collected in the past for specific purposes (e.g., when recruiting an employee), it is necessary to check whether this data may be used for any other purpose (e.g., training of Al solutions or performance control) in accordance with the purpose limitation principle.
When using new technologies, such as Al solutions, the controller must conduct a data protection impact assessment in accordance with the GDPR prior to processing. This is particularly the case with regard to employees, as the rights of employees require special protection compared with those of the employer.
Particularly strict requirements apply if Al solutions are used for the preparation of decisions that produce legal effects (e.g., in personnel recruitment or development); the use of Al solutions is usually not permitted in employee contexts such as these.
Particular attention should also be paid to Al systems that independently gather data via the internet, for example, by “crawling” through databases and websites (text and data mining) and by using search engines. The extent to which this is permissible under data protection law and whether this poses additional risks must be evaluated and documented in each specific case.
If only data that cannot be attributed to individual employees or other persons is used (e.g., purely business-related or aggregated datasets), consent is not required in accordance with data protection law. It can therefore be useful to limit the use of Al solutions from the outset to aggregated, non-personally identifiable data or to anonymize data before it is used in an Al system. If only data that cannot be attributed to individual employees or other persons is used (e.g., purely business-related or aggregated datasets), consent is not required in accordance with data protection law. Moreover, the processing of personal data is not necessary and therefore prohibited if the purpose of the processing operation can also be achieved using anonymous datasets.
Consequently, it is essential to check as a matter of priority whether anonymous data would be sufficient for the intended purpose. If personal data is used, it is generally less intrusive to anonymize this data before it is used in an Al system. From a legal point of view, even anonymizing data can constitute processing that requires a legal basis.
If data is not separated into aggregated and personal data, individual performance control may become possible. Data that can be restored back to its original personally identifiable form is not anonymized but pseudonymized, meaning that it still qualifies as personal data under data protection law.
Al that facilitates performance control requires an agreement with the works council.
Using predictive analytics methods (i.e., methods used to predict future events based on analyses of historical data) runs the risk of illegal discrimination and may influence conduct. Violations of the German General Act on Equal Treatment (AGG) must be avoided completely. If Al is used in establishing selection guidelines, the works council must co-determine this use.
Guiding questions
- Which datasets are to be used for developing and training the Al? Which datasets will then be processed during its subsequent practical use? What legal requirements must be taken into consideration?
- Is the use of anonymous data sufficient for the purpose of the respective application?
- Is there a sufficient legal basis for processing personal data?
- Who has access to the data? How are the organization’s datasets protected against misuse or industrial espionage?
- Do the risks to personal data need to be evaluated and documented in advance by means of a data protection impact assessment? Who will complete the assessment?
- How are co-determination rights exercised with regard to the protection of personal data?
- How are the criteria of data minimization and purpose limitation of data use implemented?
- How are employees and other data subjects informed about personal rights and data protection issues? Is the use of Al solutions compatible with the processing purposes that were communicated to employees?
- Does the Al autonomously prepare or make decisions that produce legal effects concerning employees or other persons (e.g., by means of predictive analytics)?
Ensuring ethical compatibility
The human criteria also include legal and ethical requirements (ELSI). ELSI is an acronym that stands for “ethical, legal and social implications.” Technology has no capacity to make moral judgments. To ensure that human autonomy with regard to moral decisions and actions is retained, ethical principles must be followed during the system design process.
The design principles of Floridi and Cowls [6] provide a basis for discussion:
- Beneficence: This principle emphasizes the importance of human welfare, and of preserving human dignity and protecting the ecosphere.
- Non-maleficence: This principle warns against the negative consequences of excessive use or misuse of Al systems. It is particularly important to avoid any violations of privacy and to respect human autonomy.
- Autonomy: When using Al systems, humans yield some of their decision-making authority to a technical artifact. The possibility of Al systems assuming decision-making authority and thereby undermining human autonomy must be avoided.
- Justice: The principle of justice should ensure equal access to the benefits of Al systems with the aim of avoiding discrimination of groups in society. It also includes unbiased use of datasets that are used for Al systems.
- Explainability: This criterion refers to the transparency and comprehensibility of Al-based decision-making processes in an epistemological sense, as well as to accountability with regard to the decisions made.
The criterion of ethical compatibility comprises those design aspects that reinforce human confidence in the use of Al by assigning humans an appropriate amount of decision-making and control capabilities to avoid non-transparent, discriminating or non-traceable decision-making processes.
Ethical compatibility is assessed using the following criteria:
- Decision-making authority: Decisions that have ethical relevance are made by humans and not intelligent machines, which clearly defines responsibility and the agency to act.
- Transparency: Functionality is clear, consistent and transparent; non-transparent decision-making processes using machine learning are avoided.
- Explainability: The ability to explain the functionality in understandable terms.
- Explainability: The ability to explain the functionality in understandable terms
- Personal data: The scope and nature of use of personal data has been clarified; all parties involved have expressed their consent.
The law may impose strict limits on decisions made by Al systems that require ethical justification. Al systems are not generally granted a legal personality as they are neither a natural person nor a legal person. Depending on the situation, the manufacturer, provider and/or user of the Al system bears legal responsibility for Al-based decisions. While the manufacturer of an Al system may be liable from a product liability perspective, for example, for any damages caused to third parties, the operators of an Al system are, in principle, responsible for ensuring legal compliance when using the Al system.
It is important to differentiate between damages that are directly caused by the use of Al systems (e.g., irreversible deletion or incorrect modification of datasets) and those arising from unfavorable decisions made on the basis of the results of data processing carried out by an Al system (e.g., conclusion of unconscionable contracts) or from infringing training data collected and processed by the user.
Guiding questions
- Do the people involved understand the impact of the ethical design of Al applications, or which functions may not be possible to implement?
- Has it been clarified who is responsible for the use of an Al system and to what extent?
- Is all possible damage covered by the conclusion of appropriate insurance contracts?
- What are the opportunities for employees and their representatives, among others, to have a say and shape the design, including for experimental and test phases?
Measures and tools
- Define an ethical guideline for the Al application (in coordination with employees or their representatives).
- Create a risk assessment (e.g., in accordance with [7]) for the use of Al systems in certain use cases.
- Safeguard the use of Al systems through certifications and the conclusion of insurance contracts.
Reference documents
- Opinion of the Data Ethics Commission of the German federal government on the use of algorithms and artificial intelligence
- European Ethics Guidelines for Trustworthy Artificial Intelligence (AI),
- “Hambach Declaration on Artificial Intelligence.”
Minimizing legal risks
Given that the key principle of Al systems is to accumulate and analyze large quantities of data, as described above, particular attention must be paid to the lawfulness of procuring such (training) data.
Caution: Training data may be legally protected
Regardless of whether training data must be considered as personal data and thus subject to the strict requirements of data protection law, particularly with regard to its processing, the training data itself may be protected from several legal perspectives and, as such, consent from a rights holder may be required in order to collect and process this data. Moreover, the databases from which the training data originates may also be read-protected (even if the data content itself is not protected).
Training data may be protected if it constitutes copyrighted work or a protected design, whether registered or unregistered. The following are some relevant examples of types of works that are protected by law:
- Literary works: novels, poems, plays, essays, articles, etc.
- Musical works: compositions, songs, lyrics, individual sound sequences
- Visual arts: paintings, sculptures, drawings, photographs, graphics, etc.
- Architectural works: buildings, structures, sketches, plans
- Cinematographic works: movies, animations, etc.
- Computer programs: software code, applications, algorithms
The actual training of an Al system, i.e., the automated evaluation of the patterns and correlations concealed in the training data, is in fact similar to “enjoyment” of a work, which is why it is generally possible to analyze training data without requiring explicit permission under copyright law. However, the training data needs to be saved beforehand, which is classified as reproduction and, as such, generally requires permission from the respective copyright holder.
Relevant examples of protected designs include:
- external appearance of all types of consumer products;
- all types of fabric patterns and pattern designs.
A prerequisite for protection as a design could be registering the design at a national (governed by the German Act on the Legal Protection of Designs (DesignG)) or European (governed by Regulation (EU) No 6/2002) level, which is beneficial as it makes it possible to verify whether any protection exists by simply searching the register. However, there is also the possibility that unregistered designs pursuant to Regulation (EU) No 6/2002 are afforded design protection without being registered, which generally makes researching existing rights extremely difficult or even impossible.
Reading databases?
In addition to, or irrespective of, the protection of (the content of) training data, the databases from which the training data originates may also be protected by copyright. This means that lists of poems or medical lexicons as well as websites with comprehensively organized content and a search function, and even databases in general, may be subject to copyright or copyright-like protection in cases where the acquisition, verification or presentation of such databases entails investment that is substantial in nature or scope. If copyright protection for a database exists as such, the protectability of the database content itself (i.e., the individual datasets that do not need to be subject to legal protection) no longer makes a difference.
If the training data or databases from which the training data originates are protected by copyright, the admissibility of extracting data and, where applicable, the database structure depends on the copyright exception provisions for text and data mining. This right, which is also afforded to commercial companies, significantly limits the rights of the copyright holder in favor of the Al user. However, it is important to note that copyright holders can exclude their published works from the admissibility of text and data mining if they have reserved their rights “in machine-readable form” — which can just be simple information in an imprint on a website or information in GT&Cs.
Purchasing training data?
There are now many services on the market offering training datasets. If procuring such training datasets is being considered, attention must be paid to the contract design and negotiations. The properties of the datasets being procured should be described in the contract in as specific a manner as possible. The provider should also be able to provide a binding statement on the legal protectability and, in particular, the legal situation, and contractually agree to a certain level of quality. In particular, it is necessary to include indemnification clauses that stipulate that the provider of the training datasets shall be solely liable for third-party claims for damages if the use of their training data violates the rights of such third parties.
Guiding questions
- Is the employee responsible for procuring training data aware of the fact that the legal situation regarding all training data being procured needs to be clearly defined before the training data is procured?
- Does the training data being procured fall under one of the categories of legal protection: Does it constitute personal data? If not, could the training data be protected by copyright because it constitutes images, movies, texts, etc.?
- Which people (e.g., in-house legal department, external legal consultants) must be included when checking which intellectual property rights apply to the data being procured and how this data may be used?
Action area: Social conditions
When it comes to using Al, the following recommendations are indispensable for establishing appropriate framework conditions. Without these general conditions, Al may fail or its desirable potential will go unharnessed. The advantages of the type of organizational culture recommended here are always countered by the work required and costs. For this reason, a joint decision must be made — based on the Al strategy — regarding which cultural aspects are to be promoted.
AI-promoting organizational culture
What is an organizational culture?
The organizational culture shapes the perception, reasoning, feelings and ultimately the actions of members of an organization. The organizational culture can be considered at three levels [8]:
- “Artifacts” are the visible effects of the organizational culture. They include dress code and etiquette, how work is organized, what people discuss and which form of exchange they use and which roles and processes exist for different tasks in the organization.
- “Values and standards” are the implicit basis of the artifacts. The artifact of “height-adjustable desks,” for example, reflects the standard that the health of employees should be protected.
- “Basic assumptions” are the basis of values and standards. They provide an implicit description of the world and human beings. An example of a basic assumption could be: “People are, in principle, motivated to achieve goals.” This results in the standard of giving people autonomy over how they allocate their work, which results in the artifact of flexible working hours.
The organizational culture influences how successful an organization is at implementing individual Al solutions and the overall utilization of the potential of Al.
General measures for cultural change
The members of an organization should always jointly agree to enshrine certain values and embody the corresponding culture. One thing is particularly important: Support from business management has to be visible in that its actions reflect the chosen values. This includes creating the necessary structures and approving the required resources.
- Definition of joint target values, e.g., in workshops or through surveys
- Moderated transformation process
- Provision of resources by leaders to establish artifacts (e.g., further training courses)
- Values embodied by leaders (e.g., when making strategic decisions)
- Regular progress evaluations
General measures and tools
- My perfect day: In a moderated workshop, groups of around five to eight employees describe a “perfect” daily routine. Their descriptions are used to determine which basic assumptions and values are predominant and which need to be established in order to bring about improvements. This results in a roadmap for cultural change.
- Use the KI-ULTRA Evaluation Toolkit to evaluate the predominant organizational culture.
- The values of the respective culture dimension should be embodied by management.
![Figure 4: Overview of the features of an AI-friendly organizational culture. Original illustration based on [8]](https://websites.fraunhofer.de/ergebnisse-ki-ultra/wp-content/uploads/2025/04/2025-04-30-11_34_23-2023_Leuteritz_Leitfaden_zu_Strategie_fuer_den_KI-Einsatz.docx-Kompatibilita.png)
Compared to the technical literature, we have chosen a simple and pragmatic description of the aspects that are relevant to the use of Al (see Figure 4). These cultural dimensions may be harmonized and in some cases may even be mutually dependent or support each other.
Establishing a humanistic organizational culture
A humanistic organizational culture puts the well-being of people at the center of decisions. It is human-centered and based on equal opportunities. Human-centered means that actions are guided by the extent to which decisions impact employees, customers and the social environment. This focus on people makes it easier to involve the relevant stakeholders, particularly employees, in the Al development and introduction process and thereby identify requirements and solutions at an early stage. When there is no such focus, there is the risk that legitimate interests will be ignored during the introduction of Al. This erodes trust in the change process. This can result in people having reservations about Al solutions.
A humanistic organizational culture is also based on equal opportunities, i.e., it gives everybody the same rights. Employees of diverse backgrounds, of any gender and with or without disabilities are not structurally disadvantaged by measures such as the use of Al. A participative organizational culture of this type makes it easier to implement ethical guidelines that promote acceptance of Al applications. A non-discriminatory and inclusive implementation of Al applications provides benefits for the organization, its employees and wider society. In addition, the concrete implementation of Al as well as other processes in the organization (e.g., recruitment processes) have to comply with the legal framework for non-discrimination. Diversity in the organization can also boost innovative capabilities.
Measures and tools
- Promote communication between members of the organization.
- Actively involve affected people in relevant decisions and take their concerns seriously.
- Take into consideration the mental and physical health of employees in the workplace.
- Involve employee representatives in the planning of work procedures and work processes that use Al.
- Define organizational objectives that take into account social values, such as sustainability, social justice, etc.
- Appoint an officer for equal opportunities / diversity.
- Continuously involve employee representatives (e.g., works council/staff council).
- Use anonymized recruitment processes.
- Involve the works council if Al is used to draw up guidelines on employee selection in recruitment, transfers, pay group reassignment or termination.
- Offer diversity training.
- Institutionalize knowledge transfer (e.g., via mentoring programs).
Establishing a culture of failure
A culture of failure is characterized by the fact that the practice of employees taking calculated risks to achieve the organization’s goals (e.g., Al development, even if the success of the project is not guaranteed) is generally accepted, or even encouraged. Mistakes are not punished but openly admitted and used as a learning opportunity. A culture of failure is necessary to foster a strong culture of innovation; the key to developing innovative solutions is openly acknowledging problems and tackling them in a constructive manner.
Measures and tools
- Introduce formats for dealing with mistakes in a constructive manner, e.g., “fuck-up nights”.
- Involve the works council in the creation of rules on how to handle mistakes.
- Organize meetings where employees can discuss failed projects and the causes of failure.
- If required, offer further training so that identified mistakes are not repeated in the future.
Establishing a culture of trust, cooperation, and innovation
These three dimensions of an organizational culture are mutually dependent and can very rarely be established independently of one another. The culture of trust and the culture of cooperation/collaboration form the foundation for an effective culture of innovation.
A culture of trust is characterized by the basic assumption that employees are generally motivated to act in the best interests of the organization and its members. Important information is shared rather than held back. Employees have the freedom to organize their own work to attain the goals of the organization.
Sharing information is a prerequisite for successful collaboration in teams and thus for a culture of cooperation. The basic assumption behind a culture of cooperation is that strong team cohesion leads to better work results than if everyone works for themselves. Such a culture of cooperation does not need to be implemented in an absolute form. Many organizations successfully use the concept of “coopetition.” These organizations provide performance incentives both for cooperation as well as for competition within the organization (e.g., through a cost center structure).
Common concepts behind a culture of innovation include aspects of the other culture dimensions mentioned here (e.g., trust, collaboration, change). It emphasizes the aspect of promoting innovation within the company. Innovation describes the process by which organizations turn ideas into new products, services or processes in order to gain an advantage in their market. Thinking and acting innovatively is not something that can be prescribed, but general conditions that are conducive to innovation can be established. A culture of innovation promotes the search for new ideas and their transformation into something that really exists and works. The basic assumption behind it is that innovation is important to the success of the organization. It is also based on values and standards that allow creative solutions to be found and implemented in everyday work.
Together, these three culture dimensions enable the potential of Al to be harnessed: Al applications can automate routine tasks and free up employees so that they can use their capacities for more demanding activities. A culture of trust allows for autonomy, providing freedom for creative ideas. A culture of cooperation helps to implement these creative ideas as innovation is primarily the result of teamwork. A culture of cooperation is also an important catalyst for organizational learning: Not only do individual employees learn new skills, but whole team structures adapt to new requirements. This type of organizational learning is key to an organization’s innovative capacity. A culture of innovation helps to identify new use cases in a creative process, as well as confront the necessary risks and implement an agile development process with the goal of creating a functioning solution. This also involves rearranging existing processes.
Measures and tools
- Implement flexible and agile working models.
- Transfer decision-making authority to employees or self-organized teams, e.g., when selecting tasks.
- Involve the works council in the design of working hours and the creation of ethical guidelines.
- Managers have the role of supporters so that employees can reach their potential.
- Use of team targets instead of, or in addition to, complementing individual targets.
- Create incentives through works council agreements on the wage structure.
- Transfer decision-making authority to employees or self-organized teams, e.g., when selecting tasks.
- Manager as a supporter and mediator who promotes team cohesion.
- Set up a process for evaluating ideas for innovations and select the most promising ones for an innovation project.
- Provide funding and personnel capacity to enable the implementation of new ideas.
- Network with customers, project partners or even competitors so that you can work together to predict developments and trends, and actively help to shape them.
Establishing a culture of change
The basic assumption behind a culture of change is that continuous change is a positive thing, which benefits the organization more than it harms it, or which may even be essential to its survival. This assumption does not exist in every organization. After all, every company has its own (seemingly) proven recipes for success and processes, and change always costs time and money.
It is often the case that innovations in the company can only be implemented successfully if staff actively support them, e.g., by familiarizing themselves with new software and new processes and engaging in further training.
Measures and tools
- Hold regular meetings where employees can discuss innovations and offer praise or criticism without facing any negative consequences.
- Introduce a company suggestion scheme, taking into account the co-determination right of the works council.
- Enable further training for individual employees so that they can prepare for change in the workplace independently and actively help to shape this change.
Establishing a data culture
The basic assumption behind a data culture is that data is extremely valuable to an organization and its success depends on how effectively this data is used. Data drives decision-making processes. Employees feel responsible for the quality and availability of their data and ensure that it is freely accessible within the organization, provided that this complies with data protection law. They also undertake continuing professional development to be better equipped to work with data. The organization supports this professional development.
Measures and tools
- Offer further training measures on the following topics: statistical bases for data-driven decision-making, the potential and limitations of Al, dealing with results from Al applications and new tools, such as software for data analysis.
- Provide data protection guidelines and processes, including role and rights concepts for data access.
- Provide infrastructure and software for data management and analysis.
Empowerment-oriented leadership
In addition to the cultural dimensions mentioned above, leadership behavior — usually exercised by line managers or project managers — also influences whether employees are able to harness the potential freed up by Al, or whether the effects of introducing Al will fizzle out in this area. Particularly if the organizational culture is based on trust and collaboration, the leadership style practiced in the organization should be in harmony with this. We therefore recommend practicing an empowerment-oriented leadership style in the areas relevant to the use of Al.
Empowerment means enabling employees to act and make decisions independently. There are two main reasons why this is advantageous:
- Al introduction projects are generally characterized by a high level of uncertainty and complexity, even if the company prioritizes the use cases with low complexity and high expected earnings. This complexity relates not only to the Al application itself, but also to the work processes and business processes associated with its use. To ensure that projects can be implemented efficiently and successfully, skilled workers with specialist knowledge from different domains must work together. Rather than being restricted by rigid and detailed instructions, they should be able to act independently using their expertise for the benefit of the project.
- Al applications offer the potential to reorganize work in such a way that people and machines can each do what they do best. (See: “Incorporating human and machine capabilities” above). It is therefore a good idea for any routine tasks or calculations to be delegated to the Al application if they are repeated or can be standardized and are too complex for human intelligence. Meanwhile, employees are given the freedom to do what they do best: making decisions based on experience or taking account of ethical principles, being creative, interacting with people on a personal level and advising or inspiring them. This calls for an appropriate level of autonomy.
Transformational leadership is one example of empowerment-oriented leadership. In this leadership style, managers are characterized by four behaviors:
- Idealized influence: The leader acts as a role model with their ethical behavior, and thereby gains respect, trust and admiration.
- Inspiring motivation: The leader motivates employees with a meaningful and positive vision of the future that people can work toward.
- Intellectual stimulation: The leader encourages employees to find their own solutions to problems.
- Individualized consideration: The leader acts as a coach or mentor who supports employees in their personal development. Employees’ individual needs (e.g., relating to flexible working hours) are identified and taken into account.
Employees should be able to decide for themselves how to complete their tasks at work. Instead of writing detailed work instructions, the goals that individuals or teams must achieve are defined. These goals can also be agreed mutually within the team. The goals help employees to understand the work results expected of them. Goals should be SMART (“specific, measurable, achievable, relevant and time-bound”).
The manager should actively ensure that goals are always up to date and should remove any obstacles if necessary (e.g., by setting up contacts, providing resources). If a manager adopts a leadership style based on achieving goals but only intervenes when there are problems, this is inadequate; this leadership style is called “passive management by objectives” and has a negative impact on the health of employees.
In all of these endeavors, the leader should pay particular attention to ensuring that the scope for each person to make decisions remains within that person’s desired parameters so that nobody becomes stressed through overwork.
Measures and tools
- Diagnosis: You can measure the level of empowerment-oriented leadership free of charge with the KI-ULTRA Evaluation Toolkit.
- Diagnosis: You can measure employee autonomy with the Job Diagnostic Survey (available from the German Federal Institute for Occupational Safety and Health (BAuA)).
- Intervention: Select managers who adopt an empowerment-oriented leadership style.
- Intervention: Based on the diagnostic data, managers should be given preventative training so that they know how to take an empowerment-oriented approach. If available: Include empowerment-oriented leadership in the training catalog or list of values.
Action area: Skills development and learning
Employees are indispensable even in digitalized work systems: They initiate, train and control technical systems as well as map datasets or support technical functions that technology cannot execute on its own. The use of innovative technologies requires employees to be trained in new tasks and may also necessitate the recruitment of new employees. It also opens up new career and development paths as the basis for personal growth and corporate flexibility. The need for qualification not only depends on a specific use case, but reflects the overarching Al strategy of the organization. Organizational learning complements the process of organizational development.
Strategic skills management
To ensure Al introduction projects are a success, employees must develop the relevant skills in a systematic way and these skills need to be made available at the right time. Compared to other digitalization projects, the use of Al can cause working methods to change fundamentally and may require special expertise to be developed (e.g., for handling probabilistic system outputs).
In this context, it can be helpful to implement a systematic skills management process based on the company’s Al strategy. One example of a systematic approach to strategy-based skills management is the Fraunhofer “Kompetenz-Kompass” (Skills Compass) [9]. This was developed on the basis of DIN PAS 1093:2009-07 and comprises five steps:
- Link to strategy
- Skills model
- Skills measurement
- Skills development
- Skills review
Step 1 “Link to strategy” and step 5 “Skills review” should generally ensure that suitable tools are used to systematically take into account the impact of strategic business decisions on skills requirements and that the effectiveness of activities aimed at developing skills is reviewed.
With regard to the use of Al, it is usually best if the specialist departments involved work with the Al experts or the IT department to identify the work requirements that arise for employees, then use this as a basis for drawing up skills requirements. The Fraunhofer Skills Compass can be applied to steps 2 “skills model” and 3 “skills measurement”: In step 2, you identify and describe the skills required for the relevant department so that the new tasks or the tasks changed by Al can be completed. You also define the skill level required to complete the task (basic, advanced or expert level). In step 3, this target skills profile is compared with the actual skills profile (i.e., the skills currently available), in each case at the relevant skill level (Figure 5). This enables you to detect at an early stage which skills need to be developed in step 4 and to what extent.
![Figure 5: Example illustration of skills measurement [9]](https://websites.fraunhofer.de/ergebnisse-ki-ultra/wp-content/uploads/2025/04/2025-04-30-11_35_53-2023_Leuteritz_Leitfaden_zu_Strategie_fuer_den_KI-Einsatz.docx-Kompatibilita.png)
Guiding questions
- How will the skills requirements for employees change if the Al strategy is implemented as planned?
- Will the use of Al change work procedures and processes in such a way that the works council will need to be involved in the planning stage?
- When will the skills be needed and how long is it likely to take for these to be made available in the organization (through further training or new hires)?
- Do you know how the existing (actual) and required (target) skills can be measured and how the entire skills management process can be evaluated?
Measures and tools
- Fraunhofer Skills Compass “Vorgehensweise zur Kompetenzbedarfsermittlung in Organisationen” (“Procedure for identifying skills requirements in organizations”) [9] with guiding questions and tools for strategic skills management.
Helping employees learn how to use AI
The use of Al should improve the agility and resilience of value creation processes. Agility and resilience are the result of using the dedicated abilities of workers. The ability to work in a motivated and self-organized manner can only develop in conjunction with work processes where these abilities are required. Employees’ abilities to self-motivate (i.e., goal setting), to efficiently pursue goals and to self-evaluate their work are crucial here. Self-motivated learning and working should not be taught in abstract terms.
Work activities have a motivational potential, which comes to the fore by introducing work design measures that enhance skills. This promotes intrinsic motivation, reinforces attitudes and leads to the development of appropriate coping strategies. Self-motivation requires specific work content that encourages motivation and learning.
Promoting the capacity to learn is a prerequisite for the development of employees in digitalized work processes. However, this is not an ad-hoc solution. Instead, it has to be a component of HR development and work design.
Guiding questions
- Which learning requirements can be implemented in Al-supported work processes, depending on job requirements and individual performance prerequisites?
- How can a systematic change in activities contribute to maintaining the capacity to learn? How does this affect staff deployment and career and development paths? Do the changes in activities involve transfers that require co-determination?
- How can the capacity to learn be promoted, e.g., through hands-on learning in the use of digital components or through retraining?
- How can operational training measures be developed and carried out with the involvement of the works council?
- How can participative ways of working help to maintain the capacity and willingness to learn, e.g., by allocating organizational tasks to affected employees in quality circles?
- Can tandem learning be established where employees contribute their digital expertise and specialists reciprocate with their technological know-how, i.e., by sharing experiences?
Open Lab Days: http://s.fhg.de/old
Promoting the organizational capacity to learn
Growing the organizational capacity to learn is recommended so that organizations and individuals can quickly adapt to newly identified needs and reallocate capacities. Organizational learning is a response to the challenges of organizational change, such as volatility, uncertainty, complexity and ambiguity. It involves establishing scope for action and refraining from making one-sided optimizations of operational efficiency (without an appropriate reference system for innovation and change).
Organizational learning refers to permanent, purposefully introduced behavioral changes in the organization, which are supported by changed processes or structures. A basic requirement for an organization’s capacity to learn is continuous communication about changes.
Guiding questions
- How can changes in an organization’s environment be predicted by using Al, for example?
- How can open communication on organizational developments be institutionalized and supported through technical measures?
Action area: Data strategy and technology transition
Accuracy of AI
The type of Al method used, the available data and the task itself can determine the accuracy of the results. The quality of the results in turn influences a project’s return on investment. Setting reasonable expectations of the success of an Al project is only possible if you understand which factors will affect its accuracy. For example, image recognition applications are generally more reliable than forecasting applications. Just because an Al solution worked well for one task does not mean that it will also be suitable for other tasks. As a rule, Al models cannot be easily transferred to other areas. This means that Al projects must always be guided by specific processes and the available datasets.
Establishing the data strategy
Data forms the foundation of Al projects. While the data required for certain use cases may be available or can be collected during a project, there are many use cases that need a significantly long data history. This raises the question of how the organization will handle the process of collecting data in the medium and long term.
In the same way that processes evolve, datasets that you need for operating an Al model may also change at short notice. Whenever more up-to-date data becomes available, the effectiveness and accuracy of the Al application are affected.
With a growing volume of processed data and associated application software, the question of a unified strategy for data management becomes increasingly important. Growing systems normally result in a significant amount of technical and organizational work on data integration and maintenance. In the medium and long term, this uses up too many resources. There is therefore a need to decide on data consistency and planned structures.
Data storage can either be in the form of large data warehouses (i.e., large volumes of preprocessed data stored for their application purposes) or data lakes (i.e., large volumes of data consistently stored in their raw form) or in the form of many individual databases. In addition to the technical storage of data and data access, there are also organizational challenges: implementing the appropriate role and rights concepts, avoiding redundancies, complying with data protection requirements and implementing suitable documentation processes (especially for the processing of personal data).
A successful data strategy also takes into account legal aspects. When designing the data strategy, it is important to know which data may be used for which purposes as well as to what extent the data to be used for an Al system needs special protection. It must always be ensured that the data can be used in compliance with the law. In terms of personal data, this is only possible if the relevant legal bases exist for the data use and if the data subjects are fully informed of the processing purposes. If data was collected in the past for different purposes (e.g., collecting the master data of an employee for the purpose of handling the employment relationship), the processing purposes can only be changed subsequently if certain requirements are met (purpose limitation principle). For this reason, the data protection officer should be involved in the development of the data strategy as soon as possible.
Aside from data protection considerations, it is possible, in principle, to use data without restrictions. This applies to both self-generated data and data procured from third parties as there is no legal recognition of data ownership, and copyright (or copyright-like law) on raw data, such as measurement data or statistical data, only applies in very few cases. However, this may need to be assessed if data comprises texts or images (see “Minimizing legal risks” above).
If data from publicly available sources or data providers is used, all applicable data licensing regulations must be followed, particularly those that may preclude the use of the respective datasets in Al systems or require additional license fees to do so. To avoid “under-licensing,” it is important to document (in the context of a comprehensive data strategy) whether data that was not collected in-house is subject to such restrictions. Appropriate documentation processes also strengthen data management and provide a clearer overview of the data available in the respective organization.
It should also be noted that the data available in an organization may be subject to confidentiality agreements with third parties or may even be classified as trade secrets that must be actively protected pursuant to the German Trade Secrets Protection Act (GeschGehG). Missing documentation on the confidentiality status of the respective data may mean that such data is processed during the use of Al systems and consequently made available to third parties because an Al system cannot readily differentiate between “ordinary” and sensitive data. Disclosing such data may constitute a violation of confidentiality agreements and thus result in contractual penalties. It may also lead to the loss of the organization’s own trade secrets.
As the results generated using Al systems are generally not protected by copyright law but can infringe the rights of third parties (see “Minimizing legal risks” above), it is important in every use case to establish how the results obtained using Al systems will be used. As Al results fundamentally lack copyright protection, if they will be available to third parties, it is important to establish whether transferal to third parties will occur under a licensing agreement and what rights and obligations the parties are subject to.
Guiding questions
- Which datasets are absolutely necessary for the Al application to be useful and profitable?
- In which operational units and to what extent is data collected for medium- and long-term purposes?
- How is data quality ensured when collecting data?
- What are the requirements for data storage and data processing?
- Is the usability of data restricted, for example, due to data licensing regulations? Is the data subject to confidentiality agreements with third parties or does it need to be actively protected because it constitutes trade secrets?
- Is there documentation on the data available in the organization?
- Which employees and organizational units can access which data, both now and in the future, using which means of access?
- Which type of data storage fits the organization’s requirements?
- Are there different requirements for different locations or divisions, based on the structure and size of the organization?
- For which processing purposes was personal data collected in the past? Which processing purposes were data subjects informed about? Are these processing purposes compatible with the proposed new use of Al?
Measures and tools
- Document the data available in the organization together with possible restrictions based on missing rights or confidentiality requirements.
- Create a role and rights concept that is as broad as possible.
- Conclude a works council agreement on the introduction and application of the programs required for Al.
- Make a conscious decision on the type and scope of data warehousing and how they are integrated into the organizational infrastructure
- Define target processes for general data processing application cases (e.g., using the Unified Modeling Language)
- Define how results generated using Al systems may be used.
- Use (data) licenses for transferring results generated using Al systems.
Technology transition
In order to be able to work with Al applications in a sustainable manner, it is important to regularly maintain and update the technical infrastructure. It may even be necessary to introduce new technologies to the organization (technology transition). These may include data collection systems (sensors) and data storage systems (e.g., dedicated databases, data warehouses or data lakes), as well as new software for data analysis (e.g., statistical programs).
One typical technological change is the transition from traditional on-premises IT to the cloud paradigm, which involves a complete overhaul of working methods, responsibilities and technologies. While these technical changes are necessary to remain operational in the long term, they do initially incur costs and tie up personnel resources. In particular, employees must be trained in the new technologies, and processes and roles may need to be redefined.
The transition from one technology to another should be planned in a centralized manner where possible. Important stakeholders — particularly future users and employee representatives — should be involved in this planning process at an early stage. The objective should be to create an integrated infrastructure, rather than establishing incompatible isolated solutions. This is the only way that data from across the organization can be combined and used efficiently. Having said that, such changes involve costs and can be tiring for employees over time. Some technological advances can be skipped or introduced at a later date.
Guiding questions
- Is the technology transition planned centrally so that incompatible isolated solutions are avoided?
- How will we decide which technological changes are important/indispensable and which can be postponed or avoided?
- Have all processes, roles and responsibilities been defined for the new technology and have the people affected been informed of the changes? Is compliance with data protection law still guaranteed even after the transition (e.g., will personal data that is no longer required be deleted or anonymized)?
- Have the necessary maintenance processes been included in the planning and are the tools required for this available?
- Are the relevant stakeholders, especially users, included in the planning at an early stage?
- What co-determination rights does the works council have? When should the works council be involved?
- Is it clear how a planned technology transition is expected to contribute to achieving the organizational goals?
- Have the expected benefits of a technology transition project been documented and communicated openly?
- Have resources (working hours, budget) been allocated for training employees on how to work with the new technology?
- Is there a clear timeline for the implementation?
- What difficulties should be expected as a result of process dependencies, downtime and data migration?
- What is the expected lifetime of the new system?
- Have you defined how the technology transition process will be monitored and who will ensure any necessary adjustments are made if unexpected problems arise?
Measures and tools
- Carry out a SWOT analysis to weigh up the benefits and risks associated with a technology transition.
Literature
- Proposal of the European Commission for an EU Regulation laying down harmonized rules on artificial intelligence (AI Act) (2021/0106).
- Daugherty, P. R.; Wilson, H. J.: Human + Machine. Reimagining Work in the Age of AI. Harvard Business Review Press, 2018.
- Huchler, N.; Adolph, L.; André, E.; Bauer, W.; Bender, N.; Müller, N.; Neuburger, R.; Peissner, M.; Steil, J.; Stowasser, S.; Suchy, O. (ed.): Kriterien für die menschengerechte Gestaltung der Mensch-Maschine-Interaktion bei Lernenden Systemen (Criteria for the human-centered design of human-machine interaction in learning systems). Munich: 2020. https://www.plattform-lernende-systeme.de/files/Downloads/Publikationen/AG2_Whitepaper2_220620.pdf.
- Huchler, N.: Komplementäre Arbeitsgestaltung. Grundrisse eines Konzepts zur Humanisierung der Arbeit mit KI (Complementary work design. Outline of a concept for the humanization of work with AI). Zeitschrift für Arbeitswissenschaft 76 (2022) pp. 158–175.
- Pokorni, B.; Braun, M.; Knecht, C.: Menschzentrierte KI-Anwendungen in der Produktion. Praxiserfahrungen und Leitfaden zu betrieblichen Einführungsstrategien (Human-centered Al applications in production. Practical experiences and guideline for operational introduction strategies). Project report in the Advanced Al Center for Learning Systems. Stuttgart: Fraunhofer IAO: 2021.
- Floridi, L.; Cowls, J.: A Unified Framework of Five Principles for AI in Society. Harvard Data Science Review. https://doi.org/10.1162/99608f92.8cd550d1, published June 23, 2019.
- Krafft, T. D.; Zweig, K. A.: Transparenz und Nachvollziehbarkeit algorithmenbasierter Entscheidungsprozesse (Transparency and traceability of algorithm-based decision-making processes). https://www.vzbv.de/sites/default/files/downloads/2019/05/02/19-01-22_zweig_krafft_transparenz_adm-neu.pdf
- Schein, E. H.: What is culture. In Frost, P.; Martin, J., Moore, L.; Lundberg, C.; Louis, M. (ed.): Reframing organizational culture. Washington: Sage, 1991, pp. 243–253.
- Schnalzer, K., Karapidis, A., Dworschak, B., Mozer, P. (2022). Vorgehensweise zur Kompetenzbedarfsermittlung in Organisationen (Procedure for identifying skills needs in organizations), Stuttgart. https://publica.fraunhofer.de/handle/publica/419707
Imprint
Fraunhofer Institute for Industrial Engineering IAO
Nobelstraße 12
70569 Stuttgart
Germany
www.iao.fraunhofer.de
Authors
Martin Braun, Jan-Paul Leuteritz, Maike Link, Helmut Zaiser, Damian Kutzias
Editors
Oliver Riedel, Katharina Hölzle, Wilhelm Bauer, Matthias Peissner
Contact
Press and Public Relations Fraunhofer IAO
Nobelstraße 12
70569 Stuttgart
Germany
presse@iao.fraunhofer.de
Fraunhofer Publica
http://dx.doi.org/10.24406/publica-4419
© Fraunhofer IAO, 2025
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License: https://creativecommons.org/licenses/by-sa/4.0/legalcode.en
The legal expertise was provided by the law firms BHO Legal and Oppenländer.
About this guide
This guide to strategic change will help you to identify potential applications of artificial intelligence and leverage these on a company-specific basis. It will also detail the challenges associated with the holistic introduction of Al applications. One key focus is the human-centered design of Al-supported work. It is possible to optimize existing business processes or develop novel business models only if humans and artificial intelligence mutually reinforce each other's capabilities. This guide summarizes various work results and the outcomes of discussions held within the framework of the KI-ULTRA project. However, it only provides basic guidance with respect to relevant issues, interdependencies and operational practices. It is important to reflect on these ideas in the respective business context; experience suggests that it is best to do this in a project team involving experts, management and the works council. Employees should also be involved in the design and implementation processes in order to address their needs and utilize their practical knowledge, thereby ensuring the success of the change process.
Such discussions with experts and stakeholders will allow you to make an informed judgement on the extent to which the use of artificial intelligence will affect the operational business model and corporate strategy. This applies in particular to legal issues, which must be examined on a case-by-case basis before any Al is used. This guide provides an overview of such issues. However, this guide is not a set of step-by-step instructions to be followed in a specific sequence. Instead, it is divided into six action areas that you can implement separately:
- Operational Al strategy
- Human-centered design of changes
- Work design strategies
- Social conditions
- Skills development and learning
- Data strategy and technology transition.
The "Operational Al strategy” action area provides the general basis for change management. This guide describes the topic areas that are particularly important based on project experience and in which the solutions are not yet generally known — at least with respect to the application of Al.
MG - management, e.g., general management, team leaders, etc.
EM - employees
OS - occupational safety specialist, work designer (including productivity and quality management)
SE - IT security specialist
Al - Al experts, people who deal with data, algorithms and models
HR - Human Resources
WC - works council
Target audience
This guide is aimed at private-sector companies, non-governmental organizations (NGOs) and public institutions that intend to implement and use artificial intelligence applications. The main target audience is comprised of decision-makers at the level of operational management, project management and employee representation. In addition, the guide is aimed at all stakeholders who work on the development and application of artificial intelligence in a work context. Icons are used throughout the document to indicate which action aids are of particular interest to which stakeholder group. Experience shows that effective, sustainable solutions are developed through dialog. This enables a wide range of perspectives to be considered. Speak to experts or join inter-company working groups.
Reading guide
This guide is divided into six action areas. Each action area comprises several action aids that include clear and specific instructions for a particular topic in order to improve your decision-making ability or ability to act. In some cases, however, the guide can only give you general advice and ask questions that you will need to consider if you are to find suitable answers for your organization. In other cases, you should bring in external consultants to help you successfully apply the methods. The action aids include the following elements:
Title
For example: Defining strategic goals
Icons
Description
The description is a text that briefly illustrates the problem and solution.
Guiding questions
The guiding questions provide additional help and supplement the description. Try to answer the guiding questions yourself. The answers may vary depending on the organization and industry.
Measures and tools
Where possible, we refer to methods and tools developed by us or by third parties that you can use to answer the guiding questions. These methods are freely available.
When is it worth getting involved with AI?
What do we mean by "artificial intelligence"?
There is no clear and comprehensive definition of "artificial intelligence." In this guide, the term refers to technologies that can imitate certain types of behavior that would be impossible for conventional software algorithms. We use the term to refer to standardized Al applications that have a narrow scope in terms of function (e.g., text or voice recognition methods) as well as solutions that need to be developed or adapted for specific applications.
The European Commission understands that the use of artificial intelligence can contribute to improving forecasts, optimizing operations and resource allocation, and personalizing service delivery. This can support socially and environmentally beneficial outcomes and provide key competitive advantages to companies and the economy.
From a legal perspective, there is also no clear-cut definition of "Al." The draft of a regulation on artificial intelligence at EU level (EU Al Act), which was discussed in the EU Parliament on June 14, 2023, provides a broad definition of Al systems; Article 3 (1) defines such a system as "a machine-based system that is designed to operate with varying levels of autonomy and that can, for explicit or implicit objectives, generate outputs such as predictions, recommendations or decisions that influence the physical or virtual environments." Since it can be assumed that the expected entry into force of the EU AI Act will have a significant impact on the use of Al, it is important to keep a close eye on the legislative process. It is likely that Al systems will be governed following a risk-based approach similar to that of the General Data Protection Regulation (GDPR), whereby the requirements for creating and using Al systems will gradually become more rigorous as the level of risk associated with the use of the respective Al system increases.
Artificial intelligence (Al) is, by definition, a sequence of mathematical models running on computing machines. With the help of special interfaces, these models record human and machine behavior using extensive data in order to mathematically process and imitate it. In the training process, the data patterns are submitted to experienced Al instructors who determine how the artificial intelligence should respond to an input. Once trained, the Al machine uses its heuristic set of methods to respond to data patterns that it identifies by itself. Al is therefore capable of handling human tasks with a level of adequacy that is at least similar to that of humans.
The potential of using AI
The action aids documented here should provide you with a better understanding of why and how you might use Al in your organization, whether in a pilot project or on a larger scale. This requires a basic understanding of the possibilities and limitations of Al applications. The use of Al should help to provide a better solution to a specific problem in an organization. Practical examples offer guidance; these can be found on the KI-ULTRA website or in the knowledge pool of the "Zukunftszentren", for example.
We recommend a human-centered design approach for Al in order to improve its effectiveness when it is deployed in practice. The human brain and artificial intelligence should not be pitted against each other but rather used in tandem. No Al system can compete with the ease and efficiency with which humans learn, understand and act. Humans have cognitive skills that also include creative inspiration and human empathy. For example, they can differentiate between animate and inanimate objects. They can pick up new terms after hearing them just once or twice, and learn by trial and error.
Conversely, artificial intelligence is capable of many things that humans simply cannot do or only with great difficulty. It can identify patterns in huge data volumes, manage complex manufacturing processes, analyze satellite images to help farmers increase their crop yields, detect financial fraud and determine user preferences. More importantly, however, Al can maximize productivity within the process of dividing tasks between humans and machines.
Compared to other digitalization solutions, Al solutions cannot usually be bought "off the shelf." They often need to be adapted to the organization in question. For this reason, using Al (compared to other technologies) often takes more time and costs more money before the technology can show its strengths. Its use is also associated with significant uncertainty: It is often only possible to identify whether a project will be a success at a late stage. A process of trial and error is required in order to determine which approach will achieve the objective.
Al can help to achieve the following goals:
- offering new services or products, or achieving better quality than the competition (e.g., targeted product suggestions for existing customers; simplifying predictive machine maintenance);
- making processes more efficient by automating sub-steps;
- relieving employees of monotonous, menial tasks;
- providing targeted support for inexperienced employees so that a larger number of applicants can be considered for a role in view of the (regional) shortage of skilled workers.
Al systems open up new potential for assistance and automation. If you are facing the challenge of making business processes in your organization more flexible, then engaging with the topic of Al can open up some new perspectives and complement other organizational approaches for improvement. In addition to these specific benefits, Al can provide fundamental value by supporting operational decision-making in turbulent markets (e.g., demand-based production planning) on the basis of extensive datasets.
To be able to make the most of Al and obtain meaningful results, companies first need a high-quality database. Al systems can, however, work with incomplete datasets, provided that certain compromises are accepted in terms of the quality of the results. This makes the technical system highly adaptable to changing conditions.
The limitations of using AI
The limitations of using Al systems mainly result from the existing technological possibilities, regulatory frameworks and what is deemed acceptable by employees. For example, Al systems have a limited capacity to imitate emotional skills or complex thought processes, such as when selecting employees or resolving social conflicts:
- Al systems have no emotions. This means that they are incapable of acting with empathy.
- Al systems have no capacity to make moral judgments. The ability to make complex decisions on moral grounds cannot be simulated using machines. Only humans can make moral decisions.
- Al systems have no awareness of problems: Although they are capable of imitating certain thought processes, they have no underlying rational understanding, intuition or insight.
Al systems must only be used within the scope of the relevant legal framework. Labor and data protection laws as well as the (future) Al Act may restrict the use of Al. When the EU Al Act enters into force, certain Al practices will be banned; high-risk Al systems will have to meet certain requirements. The use of an Al system might depend on the approval of employee representatives, i.e., works council or staff council (see "Human-centered development of Al solutions" below). As a result, the use of Al will be restricted in cases where, for example, the changes to working conditions or performance monitoring would be unfavorable for employees, thereby precluding approval.
Al should not make critical decisions autonomously that could affect people's lives (e.g., customers, patients), such as a medical diagnosis. Rather, Al should assist the people who are making the decisions. It should also be noted that Article 22 (1) GDPR prohibits automated decision-making that produces legal effects concerning the data subject or that similarly significantly affects the data subject (e.g., economic and social effects). This prohibition is subject to a few exceptions (Article 22 (2) GDPR). For example, if an Al application is used to identify unprofitable contracts that are to be terminated, the decision must not be made immediately but only after (decision-relevant) human interaction and in line with transparent criteria (based on the automated decision) in order to comply with data protection law. It is therefore important that such human interaction is not just a formality and a person actually influences the content of the decision.
If your intended application comes up against these limitations, there are two options:
- You realize that the use of Al will not make a significant contribution to achieving your organization's goals. As such, there is no need for further action in this area.
- You aim for a holistic design approach that takes into account human skills and contributions. In this way, any possible unwanted side effects resulting from the use of Al can be offset and the skills of employees can be enhanced through the use of Al.
Legal requirements for the use of AI
In addition to the technical and organizational limitations of the use of Al, it is important to always bear in mind that the specific use of Al systems must be in line with the applicable legal provisions (which are currently undergoing some substantial changes), taking into consideration not only the EU Al Act previously mentioned. Particular attention should be paid to the following areas in the strategic planning of the use of Al:
General use of Al:
- rights clearance in the procurement of training data
- use of the Al by the user / monitoring data entry (e.g., data protection law, trade secrets)
Decision making:
- labor law
- data protection legislation
- general right of personality
Productive use of Al
- protectability of data produced using Al
- statutory intellectual property rights to the data
- patents / utility models
- design (in the legal sense)
- copyrights / ancillary copyrights
- contractual data protection
The following sections will provide a basic overview of more specific legal restrictions or requirements. This information will not be separated from the rest of the text but integrated at the relevant points. Due to the wide variety of possible uses of Al, the legal information provided herein is not exhaustive and is not a substitute for individual legal advice. As such, you should always clarify individual questions with legal experts.
Action area: Operational Al strategy
Defining strategic goals
If your company is considering introducing Al, you need to establish in advance what you expect to gain from the use of Al, what projects would be suitable as pilot projects and whether your workforce has the necessary expertise for the long term. If you are faced with your very first Al project, it will probably seem very daunting to you. Once you know what you will need to focus on during this process, you will see the project in a new light.
Developing an Al strategy is a good idea if any of the following points apply to your company:
- You are planning to introduce Al applications on a large scale.
- Different departments will need to collaborate to implement the planned use of Al.
- You have secured the support of relevant decision-makers (e.g., management, data protection officer, other stakeholders) to ensure that the use of Al is a success (e.g., for launching follow-up projects).
In the Al strategy, you document the way in which the use of Al applications is intended to help achieve the operational goals. This highlights the significance of the Al applications for all stakeholders.
The Al strategy always starts with the organization's overarching vision. The vision describes a desirable future state of the organization. The vision should clearly show how the organization's core business relates to its market environment, for example, customers, suppliers and competitors (see Figure 2); from experience, a vision that is purely inwardly focused ("Our administrative processes are particularly lean") can lead to bad decisions. Once the vision has been formulated, it is linked to the organization's strategic goals and the contribution made by Al applications.
Rather than being specified at management level, the Al strategy should be drawn up in a participative decision-making process. In other words, all of the relevant stakeholders should be involved in drawing up the strategy (i.e., business management, managers of specialist departments, domain experts, Al experts, works council, affected employees). The Al strategy serves as a basis for change management (see "Change management as a means of harnessing the potential of Al" below): The strategic measures are developed based on the target state documented in the Al strategy. The process of drawing up the Al strategy therefore helps to identify the importance of organizational and cultural changes (beyond the purely technical or process-related aspects) for the introduction of Al.
These measures can be better coordinated between departments or areas of responsibility if all the parties involved understand the goal behind each measure and how important the goal is to the organization as a whole. If the Al strategy links the measures to the organizational goals and is developed in a joint process, the parties involved will be more willing to support the measures. Even if a pilot project is canceled because it is unlikely to succeed, the Al strategy serves to ensure that the corporate goals in the field of Al will be achieved in the medium to long term. The joint strategy discussion may highlight which forms of Al expertise are already available in the company and where there are gaps that need to be filled.
Smaller organizations may be able to list some relevant Al applications straight away (see "Systematic search for Al use cases" below); for larger organizations, it may be advisable to clarify the following questions:
- What are our commercial objectives with regard to using Al (e.g., improving market position, reducing costs, gaining the ability to cooperate with customers)?
- How does the Al application help to achieve these objectives and how will the achievement of objectives be tracked (e.g., exploiting new market potential, effects of automation)?
It is important to clarify the operational benefit of using Al. One of the first questions you should ask is: "Why do we think that this investment will pay off?" To be able to answer this question, you need to know which problems you want to tackle (e.g., bottlenecks in the operational process) or where you want to improve efficiency and innovative strength (such as in the digitalization of distribution channels). Al projects should focus on processes that have a tangible impact on expenditure and resources and that significantly influence operational profitability.
The importance of the strategic use of Al for the organization is illustrated by the example of three levels:
- Level 1: Al potential is (currently) of minor importance to the success of the organization. The core of the strategy is therefore to drive digitalization in order to make efficiency gains and, where applicable, to establish better conditions for the use of Al in the future.
- Level 2: Al is used to make improvements in individual application areas (e.g., in production and assembly); all other possible fields of application (e.g., customer relations management) continue to operate as before.
- Level 3: The future business model is largely based on the use of Al; the widespread implementation of Al is seen as a critical success factor.
This classification can also be helpful when answering the question "Make or buy?" (see "Strategic approach to 'make or buy'" below). The more important Al is for the core business, the more worthwhile it is for the organization to establish its own Al expertise. You will need to answer the question "Make or buy?" for each individual Al project, but it is advisable to specify a strategic direction.
Guiding questions
- What overall benefit do you expect Al applications to bring to the organization?
- What strategic goals are to be achieved through the use of Al applications?
- What specific tasks do you want to address to gain experience with Al? Where do you want to improve efficiency or innovative strength, for example?
- What types of Al applications can be considered in your organization and to what extent?
- What expertise should be available for the introduction and implementation of Al applications?
- Should Al applications be purchased or developed in-house (over the long term)? (See "Strategic approach to 'make or buy'" below)
- How will you determine whether Al is actually helping to solve pressing problems? Under what conditions will Al projects be canceled?
- How can it be ensured that the use of Al will have a positive impact on working people and their circumstances? Will jobs be lost?
- Is there a need to obtain further qualifications? Will the way employees communicate change?
- How will the Al strategy be implemented in a way that complies with data protection law?
- How will it be ensured that all of the relevant stakeholders are involved?
- Which core business processes will be affected by the planned use of Al?
- Will the introduction of Al involve a change to the business organization that requires co-determination?
- What resources will be made available for implementing the Al strategy?
- What will the available time frame be for doing this?
Measures and tools
- With the involvement of relevant stakeholders, define the contribution made by Al to achieving the vision or goals of the organization.
- Involve a high-ranking decision-maker who will support the implementation of the Al strategy and enable it to be implemented by approving the use of resources (e.g., project funds).
- Include the works council or staff council at an early stage to minimize potential fears and ensure timely communication for maximum transparency.
- Involve the data protection officer and, if necessary, schedule a data protection impact assessment from the start.
- Involve the legal department to ensure that processes are legally compliant.
- Review the vision and strategy on a regular basis and adapt them as necessary.
- Visualize and communicate the Al strategy using a strategy map (see Figure 2).
- When formulating a corporate strategy, an external coach can broaden perspectives, reflect on the discussion and ask critical questions to ensure a better understanding.
Determining the transformation requirements
In order to develop change measures based on the Al strategy, it is a good idea to first take stock of the current situation, i.e., which favorable conditions for harnessing the potential of Al already exist and which do not. The KI-ULTRA Evaluation Toolkit³ is available as a guiding evaluation instrument. The purpose of the evaluation is to identify the strategic action areas with the greatest potential for improvement. Each of the result fields in the Evaluation Toolkit links to relevant action aids in this document, which can be used to improve the conditions for the use of Al. This guide therefore constitutes a maturity model for the use of Al; other maturity models can be used too.
Maturity evaluations primarily refer to the following main topics:
- Organization and work methods
- Infrastructure
- Datasets and data availability
- Culture and collaboration
- Ethical, legal and social implications (ELSI).
Guiding questions
- In light of the general conditions, is your company currently capable of addressing the topic of Al in an operational context?
- Are the appropriate resources and expertise available?
- Has upper management made a clear commitment to implementing Al?
- What legal issues need to be clarified? Is the necessary legal expertise available or has it been acquired? What needs to be done to establish the necessary general conditions?
- What needs to be done to establish the necessary general conditions?
- Who needs to be involved if the general conditions in the individual action areas are to be improved?
- Which works council co-determination rights require consideration? Is there a plan of work procedures or processes that the works council must be notified of and consulted about?
Measures and tools
- KI-ULTRA Evaluation Toolkit
https://www.ergebnisse.ki-ultra.iao.fraunhofer.de
Strategic approach to “make or buy”
Should you purchase your Al applications or develop them yourself? This question does not need to be addressed straight away because you will probably bring in external support for your first Al project. However, the question of whether to make or buy will arise in the medium term if Al is to be used on a larger scale or if Al solutions are likely to be a unique selling point of the organization. When making this decision, it is important to consider the guiding questions listed below. However, the weighting of these aspects in the decision-making process should be defined specifically for the organization in question on the basis of the Al strategy.
Guiding questions
- Are appropriate skills and structures (e.g., data scientists, data engineers, application designers) available so that a realistic estimate of the level of work and the costs for Al development can be made and assigned to a cost center?
- Which alternative promises to better incorporate proven findings from the field of industrial science in the human-centered work design?
- Which option (purchasing Al or developing it in-house) will better address any possible concerns of the works council regarding the impact on employees?
- If a solution is purchased, does this create a unilateral dependence on an external service provider, for example, due to migration of expertise or disclosure of datasets?
- What are the follow-up costs when purchasing a solution from external service providers (e.g., updates, maintenance)?
- Is the service provider known to be reliable?
- Can you trust that they will respond quickly to any problems and charge a reasonable price for this?
- Are there GDPR-compliant solutions for the tasks to be handled? What agreements can be made with the service provider in terms of responsibility for the datasets (e.g., fulfilling information requirements, conducting data protection impact assessments, ensuring the technical security of the datasets, etc.)?
- Where will the data be processed?
Systematic search for AI use cases
The procedural model for project implementation starts with a defined use cases for Al. A key question for many organizations is how they can identify appropriate use cases in the first place in order to pick low-hanging fruit.
The following are suitable candidates:
- activities that are time- and labor-intensive (such as reading extensive documentation to identify where action is required).
- work processes that require complex image recognition.
- processes that can be complemented by speech analysis (such as in customer service).
- predictions about customer behavior in sectors such as insurance, banking or advertising.
The larger the organization and the more diverse the business processes, the more important it becomes to incorporate expert knowledge from the specialist departments via appropriate processes and structures. Therefore, the search for use cases should be approached as a cross-departmental strategic measure.
There are two possible approaches:
- based on problems identified in specialist departments or by technical specialists (e.g., employees, quality assurance and scheduling)
- based on the potential of the Al application (i.e., identified by people with technological expertise, people with Al expertise)
The following gives a brief overview of the problem-induced approach and the technology-induced approach:
Approach 1: Identifying AI technology for a given problem
The problem-oriented approach ("market pull") ensures that the specific need for an Al application is identified at an early stage, which can simplify the subsequent process. Solutions can often be found in issues already identified in a business process. For example, the existing issue could be the catalyst to contact an internal specialist department or service provider to examine the use of Al. The availability of any potentially suitable Al solutions in the company, which could be adapted for the given issue, should also be assessed.
Guiding questions
- Have Al technologies already been used for similar tasks? If so, which ones?
- (Are there any known comparable applications from other organizations?)
- In what way is Al better or different compared to previous technologies? What is the difference?
- Will the use of Al improve work processes and workflows for the benefit of the employees?
- Which optimization goals will be achieved by using Al technologies?
- Which business process data is available in appropriate quality? Which additional data has to be collected?
- Under what conditions can the chosen technology be used for the known problem in compliance with the law?
- Taking into account all intellectual property rights (see "Minimizing legal risks" below), is suitable training data available?
Approach 2: Identifying a problem suitable for AI technology
Alternatively, you can check whether any business process issues can be optimized through Al technologies and, if so, which ones (data-driven approach, "technology push"). Is it possible to tap into new business potential using the availability of (large) datasets? For example, is it possible to identify customer needs more effectively and produce target group-specific products? Can the purchasing behavior trends be more easily predicted, such as over Christmas? Can the wear behavior of machinery be more easily predicted inside the organization and can maintenance be planned more efficiently?
In this approach, people with Al expertise should continuously identify and evaluate the potential benefits of Al technologies. During concrete projects, they then define requirements for promising technologies. These should be specific but not limit the potential solution in any way. A sample question could be how to automate quality assurance through image recognition. However, it is important to bear in mind that Al will only ever improve individual phases of a process; it is therefore difficult to calculate how it will impact the overall result.
Guiding questions
- What competitive advantage would it offer the organization if recurring patterns could be detected in a process, if it was possible to predict process behavior, if statistical methods could be used to support decision-making processes?
- Are there any recurring, essential tasks that require a lot of time or effort?
- Are there any tasks where an analysis of large data volumes would reduce mistakes?
- Can you demonstrate what Al is capable of (and what it is not) for specific applications?
- Are there areas in which work processes and workflows could be improved for the benefit of employees?
For most organizations, Al is a means to an end and not an end in itself. In other words, its main purpose is to solve known problems or achieve improvements. Al should be used when it is the most promising of all the solutions available. Remember: It usually takes some time before the impact of a new technology on a work system becomes apparent.
Measures and tools
Companies can use their suggestion scheme to bring together problems and Al technology. Topic-specific workshops have also proven to be useful. If data scientists and data experts work in the company, you should consider forming an "Al hub." This would do the following:
- Promote the exchange of expert knowledge and expertise between different parts of the company, for example, to obtain tips and help when facing problems in a project or to share the latest technical information (e.g., new technologies, discussion of experiences).
- Members of the hub can approach specialist departments directly to discuss daily working routines and identify areas with potential to remove tasks or provide support.
- Offer Al training for employees so that they gain basic knowledge and core areas of expertise in relation to Al.
The following aspects are critical to the success of Al hubs, etc.:
- providing resources (e.g., digital communication platform, internal forum)
- providing sufficient capacities
- Internal communications: Employees should be familiar with the hub, its members and responsibilities.
AI map: This lists AI providers, users and development projects in the form of a map and in a list. It helps to spark ideas.4
Please note: If an Al hub is set up, this does not replace measures for employee involvement, such as open feedback discussions about the implementation of Al in the organization. Irrespective of the existence of an Al hub, co-determination rights of employee representatives must be respected (in particular the works council's right to be kept informed and consulted in the planning of work procedures and work processes). Through training measures on the basics of Al, it should be ensured that the people involved can evaluate the extent to which Al would actually be useful for solving a problem.
AI map: https://cta4.plattform-lernende-systeme.de/map-on-ai-map.html
Action area: Human-centered design of changes
Change management as a means of harnessing the potential of AI
The introduction of Al applications may require operational changes in an organization. As part of change management, these changes are implemented in a targeted way and affected employees are involved at an early stage so that the organization adapts to the new requirements. If a fundamental change is required, an agreement must be reached with the works council to reconcile their interests; this process might involve drawing up a social compensation plan. Employee transfers must be covered by the employment contract and are only permitted with the approval of the works council.
If you are initially planning an Al pilot project to gain experience, then the Guide to the implementation of Al projects provides a basis for ensuring that the most important change requirements are taken into account over the course of the introduction project. In relation to change management and with regard to a specific Al application, you can also use the "Introduction of Al Systems in Companies" guide from "Plattform Lernende Systeme."
The following procedural guide is aimed at organizations that want to fully harness the potential of Al and therefore implement Al as a cross-sectional technology. This requires different areas of the company (e.g., HR, production and customer services) to work together.
The starting point for change management is the Al strategy (see "Defining strategic goals" above). It sets out what contribution Al makes to the success of the organization and how this contribution is created. The objectives that are relevant to the use of Al are derived from the following strategic action areas (work design, competence management and learning, data strategy) and the corresponding procedural guides.
Further objectives may include:
- strengthening the role of humans as experts (in the sense of dividing functions between humans and Al in a complementary manner);
- further developing the organizational structure;
- creating new jobs or departments to pool Al activities;
- providing infrastructure
Humans and artificial intelligence complement each other. No Al-controlled machine can compete with the ease and efficiency with which humans learn, understand and identify correlations. Conversely, Al is capable of many things that would be difficult for even the most talented person: identifying patterns in large data volumes, supporting decision-making processes, managing data, etc. When humans and Al collaborate, data can be used for sophisticated and differentiated purposes. However, the use of Al requires a high level of commitment from employees. This is the focus and objective of human-centered change management.
Guiding questions
Answer the following questions within the scope of change management:
- Has a development goal been defined that recognizes the potential benefits of human and Al collaboration?
- Does the planned use of Al result in changes to the business model that require changes in the organization?
- Does an operational change require co-determination?
- Does communication help informed decisions to be made within a reasonable time frame in order to utilize the benefits of Al?
For example, you may be using Al in different departments to support corporate decision-making (e.g., predicting demand and adapting production, pricing products and services). However, the benefits of the technology fizzle out because decision-making channels are long and involve different levels of the hierarchy. One solution may be to introduce interdepartmental teams so that decision-making is accelerated.
- Is our working method suitable for the planned introduction of Al? How can we accelerate the development and introduction of Al while respecting the co-determination rights of employee representatives?
Al development and introduction projects usually involve a great deal of uncertainty. At the start of a project, it is unclear what the application should look like in the end and how complicated the development process will be. Agile working methods can help to complete Al projects faster and ensure they are a success, or to ensure that unprofitable and unfeasible projects are wound up quickly. It is not necessary for the entire organization to transition to agile working methods right away. As long as the Al project teams include members with experience of agile working methods, then, for example, employees from specialist departments that employ lean management or conventional project structures can adapt to agile working methods effectively. When applying an agile approach, the continuously changing requirements must be legally reassessed in order to avoid unnecessary expenditure in unusable Al functionalities.
- Do we need to define new roles or business processes?
Where personal data is being handled, it is necessary, for example, to appoint a data protection officer. If the intention is for data to be accessible across departments, processes should be defined to ensure that this access is compliant with data protection legislation.
The evaluation should take account of the relevant human-related impact as well as purely functional target criteria: How does the change affect employees' understanding of their role and their level of satisfaction? Are any particular people in the organization placed at a disadvantage by the change? Are findings on human-centered work design sufficiently considered? Are any particular people in the organization placed at a disadvantage by the change? Does any change result in prohibited discrimination? Does a social compensation plan need to be negotiated in order to compensate for disadvantages related to an operational change?
Human-centered development of AI solutions
When a human-centered approach is used to design systems, the focus is placed on the actual users and other affected parties, actively including them in the design process.
The number of decisions that a company needs to make is increasing. Key decisions require problem-solving, analysis, strategic thinking or the ability to prioritize. People must be able to evaluate the potential impact of such decisions and to weigh up alternatives. Al systems can be of great benefit in this respect. If data required to make decisions is shared on platforms, more people can be involved in decision-making processes, thereby increasing their acceptance.
The following decision-making scenarios are possible:
- Several Al applications ensure that the final expert opinion of decision-makers is given greater weight. One example is the use of imaging procedures in medicine.
- In other cases, Al bundles the decision-making process without changing the decision-makers. Personnel selection is one example.
- Several Al applications centralize the decision-making process. Credit card validation is a typical example.
AI projects are only effective if they respect employees' decision-making authority. A human-centered approach to Al design fosters a strong, mutual relationship between humans and machines, combining their competences.
The effects of Al on employees can be difficult to predict. For this reason, we recommend following a consistent human-centered approach when introducing Al: As a rule, the affected employees and those responsible for representing their interests should be involved in change processes. If the works council needs to assess the introduction or application of Al in order to perform its duties, it can appoint an expert.
The purpose of the human-centered approach is to incorporate the knowledge and creativity of employees in the early stages of the design process. This ensures that the resulting Al applications and work processes result in a noticeable improvement for employees. This approach promotes practical knowledge, motivation and identification with the task and the technology.
Employee representatives (works council/staff council) may have co-determination rights relating to the introduction of Al systems. Software is often a technical tool that can be used to monitor the behavior or performance of employees, which is why the works council must co-determine its introduction and use. If Al leads to a fundamental alteration of work organization, this constitutes a work alteration that requires co-determination. In this case, an agreement on a reconciliation of interests should be reached and a social compensation plan may need to be drawn up. Even at the planning stage for applying Al to work procedures and processes, the works council must be consulted and its recommendations and concerns taken into account. If the work situation for employees is significantly altered, this may constitute a transfer that requires approval by the works council. If Al is used to draw up guidelines on employee selection in recruitment, transfers, pay group reassignment or even termination, the works council must be involved.
To ensure that employees can actively participate in the design process, appropriate conditions must be created. Basic knowledge of Al (see "Making Al knowledge available" below) must first be acquired so that everyone involved has the same understanding and uses the correct terms when discussing possible solutions and approaches. Basic knowledge of Al helps us to reflect on possible reservations about Al applications. In particular, the employees' understanding of their own professional role needs to be taken into account: Does the new, altered task fit with this self-image? For example, how will employees' decision-making authority, flexibility with regard to the completion of their own tasks and interaction with colleagues change?
Guiding questions
- Which employee skills are indispensable for customer focus and business success? How could Al be used to complement these in an appropriate way?
- What attitudes do employees have towards Al and what are the experiences and assumptions underlying these attitudes?
- How can we create a basic understanding of Al across the whole organization?
- Which formats and media are suitable for providing internal information on the topic of AI?
- What experience does the company have of involving employees and their representatives? How important is it in relation to the organizational concept? Are there agreements with the works council on the introduction and application of programs?
- Employee participation requires company rules. This is the only way to ensure that accountability, trust and responsibility form the basis of daily actions. To what extent is this taken into account in the organizational culture?
- Is it assumed that affected parties and employees (data subjects) consent to their data being processed (insofar as this is necessary)? If personal data is to be processed: Is it ensured that there is a legal basis for processing this data? Can the data processing be based on the valid, voluntary consent of the respective data subject? How can the data subjects be clearly informed in advance about what will happen to their data and for what purpose? (See "Protecting personal data and other sensitive data" below; see "Data access" in the Guide to the implementation of Al projects.)
Measures and tools
- Make information on the planned use of Al transparent and publicly accessible.
- Facilitate co-determination of the works council.
- Discuss the planned use of Al with the works council at an early stage.
- Establish communication formats for all affected employees.
Action area: Work design strategies
Despite the use of innovative technology, human contribution will continue to play a decisive role in shaping work in the future. A human-centered design approach creates favorable conditions for ensuring that people are motivated and qualified to work and learn. The purpose of human-centered work design is to combine the potential for human achievement with the functional potential of technology. The way in which this is implemented in an operational sense is always decided in the specific project at hand. However, you should reflect on several principles at the strategic design level; this is the only way to ensure that Al applications are selected appropriately and implemented effectively.
Incorporating human and machine capabilities
The way in which functions are appropriately divided between humans and machines is decided on a case-by-case basis at an operational project level. However, you should create the necessary conditions at a strategic level. The introduction highlighted the potential and the limitations of Al applications (see "The potential of using AI"). You should also consider the strengths of humans when compared with Al systems, for example, as shown in the model of the missing middle (see Figure 3).
In principle, Al has the ability to replace human work (i.e., automation) and to provide assistance and augmentation (i.e., to enhance human skills). A central task in the division of functions is to assign individual tasks to technology or to humans. Here, it is important to ensure that humans end up with enough scope for actions and decisions. Together with the requirement level of the activity, this scope determines which human learning and adaptation capabilities are to be utilized so that the parts of work that cannot be planned are managed with flexibility. A division of functions based on human strengths creates favorable conditions for an agile working method.
In particular, Al applications should not be used to make decisions that have a direct impact on people; instead, such decisions should require interpersonal negotiation processes. Examples include shift or vacation planning. In many areas, using Al solutions to (completely) automate decision-making processes is only permitted by law to a very limited extent. In addition, Al applications should not be used to monitor and control human behavior; this approach is known as New Taylorism. This is where IT systems specify ideal conduct for people ("one best way") and monitor compliance. No evidence has been found to indicate that this optimization approach actually results in any improvements. However, experience suggests that viewing human performance primarily as a technical challenge results in many problems. Instead, work design should foster human capabilities, for example through understanding and meaningful engagement.
In Al systems, the interaction between humans and machines must be designed in such a way that the work content and communication tasks that motivate people and promote learning do not end up being replaced. This may occur when, for example, interesting programming tasks are reassigned to a robot, but employees are still responsible for routinely inserting workpieces into a punch machine. It is imperative to ensure that the interaction between humans and machines is designed in a transparent manner. This task is specified in the model of the "missing middle."
Guiding questions
- Which tasks can be executed reliably by people (and thus assigned to them)? How will automation in the form of Al impact the potential for people to gain experience, as well as the diversity of tasks available and the motivation level of working people?
- What are the legal requirements and the ethical aspects of using Al? Is there a risk model with suitable criticality levels to identify and minimize potential risks of the Al application? Which co-determination rights of employee representatives require consideration?
- Will decisions be made based on automated processing that is not subject to human control?
- What functions can generally be assigned to Al applications? To what extent can they intervene in the social fabric of the company? What impact does this have on the Al strategy?
- Does Al actually offer enough flexibility to adapt as necessary to the needs of the company and of people? Or would conventional automation methods be sufficient?
- If Al is replacing human work: Does a social compensation plan need to be drawn up? Are transfers required?
Measures and tools
- Define how functions will be divided between humans and Al, and coordinate with all people involved.
- Provide information on strategic frameworks for all developers.
- Derive measures for human-centered work design. You can make use of the appropriate literature for this purpose ([2], [3], [4], [5]).
- In smaller companies, an employee workshop can be held for discussion of the extent to which working conditions are in line with the design goals and what changes need to be pursued as part of a joint process.
Minimizing safety risks
Technical systems must be safely and reliably implemented as a matter of principle; relevant safety strategies are governed in laws, regulations and standards, for example. In the context of high-risk robotic Al systems, accident prevention is of paramount importance. On the one hand, accidents can be traced back to technical parameters. On the other hand, incorrect human behavior must be avoided – this can be more common if tasks are monotonous or cause mental exhaustion.
When implementing Al applications, the legal requirements pertaining to occupational safety must be taken into account. In particular, a comprehensive risk assessment may be required and the works council must be involved in matters of occupational health and safety. It is equally important to take into account the risk-based regulation of Al systems expected to take effect with the entry into force of the EU AI Act.
A fundamental criterion of human-centered design relates to the absence of harm in the interaction between people and machines. The safety risks of a machine must be reduced to a generally accepted level. Safety controllers play a significant role in this safety concept.
In Al systems, it is no longer possible to understand algorithms and decision-making processes in detail (i.e., "black box" behavior). As a consequence, people do not always judge processes correctly and sometimes make inappropriate interventions. Therefore, overarching safety functions and emergency strategies should always be in place, too. The extent to which Al algorithms can make safe decisions is generally not known in sufficient detail. For this reason, the safety level of Al systems should be verified reliably on the basis of strong indicators. This includes criteria such as the explainability and comprehensibility of decisions, applicability to different situations, monitoring during operating time and transparent forms of human-machine interaction.
A risk assessment (as would be implemented at an operational level) usually forms the basis of a safe design. At a strategic level, the structural requirements for occupational safety must be met by involving qualified personnel, providing testing equipment or training employees. In individual cases, failure to follow the appropriate provisions may mean that Al applications need to be put out of service.
The Data Ethics Commission ("Datenethikkommission" or DEK) of the German federal government has developed a criticality pyramid that evaluates Al-based decision systems and considers the probability and severity of harm. It defines requirements and control methods and offers guidance for assessing the general conditions for Al applications. In the future, the “positive list” of high-risk Al systems in the upcoming EU AI Act will also need to be taken into account.
Guiding questions
- What safety risks need to be considered when using Al?
- Does the product safety and operational safety of the Al system comply with the applicable laws?
- How should operational occupational safety be organized when using Al systems?
- Do employee representatives need to be involved due to matters of occupational health and safety?
Measures and tools
- Identify the applicable laws and regulations. In particular, check whether the Al systems to be implemented meet the legal requirements in terms of product safety.
- Check whether existing insurance policies cover the use of Al systems in the specific application area or whether additional insurance cover is required.
- Establish a risk assessment process (information is provided in, e.g., the guidelines of the German Federal Institute for Occupational Safety and Health (BAuA)).
Protecting personal data and other sensitive data
The key principle of all Al systems is to accumulate and analyze large quantities of data. The process of recording, storing and analyzing this data is becoming more and more efficient. In a work context, data analyses, inter alia, can be used to prevent errors – but also for monitoring and performance control. If personal data is collected and analyzed, the works council must be consulted.
When developing and using Al, compliance with the laws on the protection of employees' personal rights is required, from the GDPR to labor law, in particular the co-determination rights of employee representatives. Furthermore, personal data of third parties (e.g., customers) may only be processed in accordance with the applicable data protection requirements.
To determine the legal requirements, it is necessary to differentiate between the different data types and the data subjects whose data is to be processed in specific individual cases during training of the Al and during its subsequent practical use. If personal data is processed, there must be a legal basis for doing so. Personal data refers to all information relating to an identified or identifiable natural person. This includes, for example, the name, address and contact details of employees or other persons as well as information that can be combined with other data to identify a person, such as an IP address or employee number. Processing this data is only permitted if there is an explicit legal basis for this or if consent has been obtained from the data subject. The latter is often critical in an employment relationship as there are strict requirements regarding the effectiveness of voluntary consent due to the hierarchical relationship between employer and employee.
Even stricter rules apply to datasets comprising special categories of personal data (known as "sensitive data"). With regard to employees, this includes data on ethnic origin, religious beliefs, trade union membership and health as well as genetic and biometric data. Processing this data is prohibited unless one of the very limited exceptions stipulated in the GDPR applies.
In addition, it is important to take into account general data protection requirements as early as possible during project planning. In particular, the data subjects must be informed of the processing operations. If data was collected in the past for specific purposes (e.g., when recruiting an employee), it is necessary to check whether this data may be used for any other purpose (e.g., training of Al solutions or performance control) in accordance with the purpose limitation principle.
When using new technologies, such as Al solutions, the controller must conduct a data protection impact assessment in accordance with the GDPR prior to processing. This is particularly the case with regard to employees, as the rights of employees require special protection compared with those of the employer.
Particularly strict requirements apply if Al solutions are used for the preparation of decisions that produce legal effects (e.g., in personnel recruitment or development); the use of Al solutions is usually not permitted in employee contexts such as these.
Particular attention should also be paid to Al systems that independently gather data via the internet, for example, by "crawling" through databases and websites (text and data mining) and by using search engines. The extent to which this is permissible under data protection law and whether this poses additional risks must be evaluated and documented in each specific case.
If only data that cannot be attributed to individual employees or other persons is used (e.g., purely business-related or aggregated datasets), consent is not required in accordance with data protection law. It can therefore be useful to limit the use of Al solutions from the outset to aggregated, non-personally identifiable data or to anonymize data before it is used in an Al system. If only data that cannot be attributed to individual employees or other persons is used (e.g., purely business-related or aggregated datasets), consent is not required in accordance with data protection law. Moreover, the processing of personal data is not necessary and therefore prohibited if the purpose of the processing operation can also be achieved using anonymous datasets.
Consequently, it is essential to check as a matter of priority whether anonymous data would be sufficient for the intended purpose. If personal data is used, it is generally less intrusive to anonymize this data before it is used in an Al system. From a legal point of view, even anonymizing data can constitute processing that requires a legal basis.
If data is not separated into aggregated and personal data, individual performance control may become possible. Data that can be restored back to its original personally identifiable form is not anonymized but pseudonymized, meaning that it still qualifies as personal data under data protection law.
Al that facilitates performance control requires an agreement with the works council.
Using predictive analytics methods (i.e., methods used to predict future events based on analyses of historical data) runs the risk of illegal discrimination and may influence conduct. Violations of the German General Act on Equal Treatment (AGG) must be avoided completely. If Al is used in establishing selection guidelines, the works council must co-determine this use.
Guiding questions
- Which datasets are to be used for developing and training the Al? Which datasets will then be processed during its subsequent practical use? What legal requirements must be taken into consideration?
- Is the use of anonymous data sufficient for the purpose of the respective application?
- Is there a sufficient legal basis for processing personal data?
- Who has access to the data? How are the organization's datasets protected against misuse or industrial espionage?
- Do the risks to personal data need to be evaluated and documented in advance by means of a data protection impact assessment? Who will complete the assessment?
- How are co-determination rights exercised with regard to the protection of personal data?
- How are the criteria of data minimization and purpose limitation of data use implemented?
- How are employees and other data subjects informed about personal rights and data protection issues? Is the use of Al solutions compatible with the processing purposes that were communicated to employees?
- Does the Al autonomously prepare or make decisions that produce legal effects concerning employees or other persons (e.g., by means of predictive analytics)?
Ensuring ethical compatibility
The human criteria also include legal and ethical requirements (ELSI). ELSI is an acronym that stands for "ethical, legal and social implications." Technology has no capacity to make moral judgments. To ensure that human autonomy with regard to moral decisions and actions is retained, ethical principles must be followed during the system design process.
The design principles of Floridi and Cowls [6] provide a basis for discussion:
- Beneficence: This principle emphasizes the importance of human welfare, and of preserving human dignity and protecting the ecosphere.
- Non-maleficence: This principle warns against the negative consequences of excessive use or misuse of Al systems. It is particularly important to avoid any violations of privacy and to respect human autonomy.
- Autonomy: When using Al systems, humans yield some of their decision-making authority to a technical artifact. The possibility of Al systems assuming decision-making authority and thereby undermining human autonomy must be avoided.
- Justice: The principle of justice should ensure equal access to the benefits of Al systems with the aim of avoiding discrimination of groups in society. It also includes unbiased use of datasets that are used for Al systems.
- Explainability: This criterion refers to the transparency and comprehensibility of Al-based decision-making processes in an epistemological sense, as well as to accountability with regard to the decisions made.
The criterion of ethical compatibility comprises those design aspects that reinforce human confidence in the use of Al by assigning humans an appropriate amount of decision-making and control capabilities to avoid non-transparent, discriminating or non-traceable decision-making processes.
Ethical compatibility is assessed using the following criteria:
- Decision-making authority: Decisions that have ethical relevance are made by humans and not intelligent machines, which clearly defines responsibility and the agency to act.
- Transparency: Functionality is clear, consistent and transparent; non-transparent decision-making processes using machine learning are avoided.
- Explainability: The ability to explain the functionality in understandable terms.
- Explainability: The ability to explain the functionality in understandable terms
- Personal data: The scope and nature of use of personal data has been clarified; all parties involved have expressed their consent.
The law may impose strict limits on decisions made by Al systems that require ethical justification. Al systems are not generally granted a legal personality as they are neither a natural person nor a legal person. Depending on the situation, the manufacturer, provider and/or user of the Al system bears legal responsibility for Al-based decisions. While the manufacturer of an Al system may be liable from a product liability perspective, for example, for any damages caused to third parties, the operators of an Al system are, in principle, responsible for ensuring legal compliance when using the Al system.
It is important to differentiate between damages that are directly caused by the use of Al systems (e.g., irreversible deletion or incorrect modification of datasets) and those arising from unfavorable decisions made on the basis of the results of data processing carried out by an Al system (e.g., conclusion of unconscionable contracts) or from infringing training data collected and processed by the user.
Guiding questions
- Do the people involved understand the impact of the ethical design of Al applications, or which functions may not be possible to implement?
- Has it been clarified who is responsible for the use of an Al system and to what extent?
- Is all possible damage covered by the conclusion of appropriate insurance contracts?
- What are the opportunities for employees and their representatives, among others, to have a say and shape the design, including for experimental and test phases?
Measures and tools
- Define an ethical guideline for the Al application (in coordination with employees or their representatives).
- Create a risk assessment (e.g., in accordance with [7]) for the use of Al systems in certain use cases.
- Safeguard the use of Al systems through certifications and the conclusion of insurance contracts.
Reference documents
- Opinion of the Data Ethics Commission of the German federal government on the use of algorithms and artificial intelligence
- European Ethics Guidelines for Trustworthy Artificial Intelligence (AI),
- "Hambach Declaration on Artificial Intelligence."
Minimizing legal risks
Given that the key principle of Al systems is to accumulate and analyze large quantities of data, as described above, particular attention must be paid to the lawfulness of procuring such (training) data.
Caution: Training data may be legally protected
Regardless of whether training data must be considered as personal data and thus subject to the strict requirements of data protection law, particularly with regard to its processing, the training data itself may be protected from several legal perspectives and, as such, consent from a rights holder may be required in order to collect and process this data. Moreover, the databases from which the training data originates may also be read-protected (even if the data content itself is not protected).
Training data may be protected if it constitutes copyrighted work or a protected design, whether registered or unregistered. The following are some relevant examples of types of works that are protected by law:
- Literary works: novels, poems, plays, essays, articles, etc.
- Musical works: compositions, songs, lyrics, individual sound sequences
- Visual arts: paintings, sculptures, drawings, photographs, graphics, etc.
- Architectural works: buildings, structures, sketches, plans
- Cinematographic works: movies, animations, etc.
- Computer programs: software code, applications, algorithms
The actual training of an Al system, i.e., the automated evaluation of the patterns and correlations concealed in the training data, is in fact similar to "enjoyment” of a work, which is why it is generally possible to analyze training data without requiring explicit permission under copyright law. However, the training data needs to be saved beforehand, which is classified as reproduction and, as such, generally requires permission from the respective copyright holder.
Relevant examples of protected designs include:
- external appearance of all types of consumer products;
- all types of fabric patterns and pattern designs.
A prerequisite for protection as a design could be registering the design at a national (governed by the German Act on the Legal Protection of Designs (DesignG)) or European (governed by Regulation (EU) No 6/2002) level, which is beneficial as it makes it possible to verify whether any protection exists by simply searching the register. However, there is also the possibility that unregistered designs pursuant to Regulation (EU) No 6/2002 are afforded design protection without being registered, which generally makes researching existing rights extremely difficult or even impossible.
Reading databases?
In addition to, or irrespective of, the protection of (the content of) training data, the databases from which the training data originates may also be protected by copyright. This means that lists of poems or medical lexicons as well as websites with comprehensively organized content and a search function, and even databases in general, may be subject to copyright or copyright-like protection in cases where the acquisition, verification or presentation of such databases entails investment that is substantial in nature or scope. If copyright protection for a database exists as such, the protectability of the database content itself (i.e., the individual datasets that do not need to be subject to legal protection) no longer makes a difference.
If the training data or databases from which the training data originates are protected by copyright, the admissibility of extracting data and, where applicable, the database structure depends on the copyright exception provisions for text and data mining. This right, which is also afforded to commercial companies, significantly limits the rights of the copyright holder in favor of the Al user. However, it is important to note that copyright holders can exclude their published works from the admissibility of text and data mining if they have reserved their rights "in machine-readable form" — which can just be simple information in an imprint on a website or information in GT&Cs.
Purchasing training data?
There are now many services on the market offering training datasets. If procuring such training datasets is being considered, attention must be paid to the contract design and negotiations. The properties of the datasets being procured should be described in the contract in as specific a manner as possible. The provider should also be able to provide a binding statement on the legal protectability and, in particular, the legal situation, and contractually agree to a certain level of quality. In particular, it is necessary to include indemnification clauses that stipulate that the provider of the training datasets shall be solely liable for third-party claims for damages if the use of their training data violates the rights of such third parties.
Guiding questions
- Is the employee responsible for procuring training data aware of the fact that the legal situation regarding all training data being procured needs to be clearly defined before the training data is procured?
- Does the training data being procured fall under one of the categories of legal protection: Does it constitute personal data? If not, could the training data be protected by copyright because it constitutes images, movies, texts, etc.?
- Which people (e.g., in-house legal department, external legal consultants) must be included when checking which intellectual property rights apply to the data being procured and how this data may be used?
Action area: Social conditions
When it comes to using Al, the following recommendations are indispensable for establishing appropriate framework conditions. Without these general conditions, Al may fail or its desirable potential will go unharnessed. The advantages of the type of organizational culture recommended here are always countered by the work required and costs. For this reason, a joint decision must be made — based on the Al strategy — regarding which cultural aspects are to be promoted.
AI-promoting organizational culture
What is an organizational culture?
The organizational culture shapes the perception, reasoning, feelings and ultimately the actions of members of an organization. The organizational culture can be considered at three levels [8]:
- "Artifacts" are the visible effects of the organizational culture. They include dress code and etiquette, how work is organized, what people discuss and which form of exchange they use and which roles and processes exist for different tasks in the organization.
- "Values and standards" are the implicit basis of the artifacts. The artifact of "height-adjustable desks," for example, reflects the standard that the health of employees should be protected.
- "Basic assumptions" are the basis of values and standards. They provide an implicit description of the world and human beings. An example of a basic assumption could be: "People are, in principle, motivated to achieve goals." This results in the standard of giving people autonomy over how they allocate their work, which results in the artifact of flexible working hours.
The organizational culture influences how successful an organization is at implementing individual Al solutions and the overall utilization of the potential of Al.
General measures for cultural change
The members of an organization should always jointly agree to enshrine certain values and embody the corresponding culture. One thing is particularly important: Support from business management has to be visible in that its actions reflect the chosen values. This includes creating the necessary structures and approving the required resources.
- Definition of joint target values, e.g., in workshops or through surveys
- Moderated transformation process
- Provision of resources by leaders to establish artifacts (e.g., further training courses)
- Values embodied by leaders (e.g., when making strategic decisions)
- Regular progress evaluations
General measures and tools
- My perfect day: In a moderated workshop, groups of around five to eight employees describe a "perfect" daily routine. Their descriptions are used to determine which basic assumptions and values are predominant and which need to be established in order to bring about improvements. This results in a roadmap for cultural change.
- Use the KI-ULTRA Evaluation Toolkit to evaluate the predominant organizational culture.
- The values of the respective culture dimension should be embodied by management.
Compared to the technical literature, we have chosen a simple and pragmatic description of the aspects that are relevant to the use of Al (see Figure 4). These cultural dimensions may be harmonized and in some cases may even be mutually dependent or support each other.
Establishing a humanistic organizational culture
A humanistic organizational culture puts the well-being of people at the center of decisions. It is human-centered and based on equal opportunities. Human-centered means that actions are guided by the extent to which decisions impact employees, customers and the social environment. This focus on people makes it easier to involve the relevant stakeholders, particularly employees, in the Al development and introduction process and thereby identify requirements and solutions at an early stage. When there is no such focus, there is the risk that legitimate interests will be ignored during the introduction of Al. This erodes trust in the change process. This can result in people having reservations about Al solutions.
A humanistic organizational culture is also based on equal opportunities, i.e., it gives everybody the same rights. Employees of diverse backgrounds, of any gender and with or without disabilities are not structurally disadvantaged by measures such as the use of Al. A participative organizational culture of this type makes it easier to implement ethical guidelines that promote acceptance of Al applications. A non-discriminatory and inclusive implementation of Al applications provides benefits for the organization, its employees and wider society. In addition, the concrete implementation of Al as well as other processes in the organization (e.g., recruitment processes) have to comply with the legal framework for non-discrimination. Diversity in the organization can also boost innovative capabilities.
Measures and tools
- Promote communication between members of the organization.
- Actively involve affected people in relevant decisions and take their concerns seriously.
- Take into consideration the mental and physical health of employees in the workplace.
- Involve employee representatives in the planning of work procedures and work processes that use Al.
- Define organizational objectives that take into account social values, such as sustainability, social justice, etc.
- Appoint an officer for equal opportunities / diversity.
- Continuously involve employee representatives (e.g., works council/staff council).
- Use anonymized recruitment processes.
- Involve the works council if Al is used to draw up guidelines on employee selection in recruitment, transfers, pay group reassignment or termination.
- Offer diversity training.
- Institutionalize knowledge transfer (e.g., via mentoring programs).
Establishing a culture of failure
A culture of failure is characterized by the fact that the practice of employees taking calculated risks to achieve the organization's goals (e.g., Al development, even if the success of the project is not guaranteed) is generally accepted, or even encouraged. Mistakes are not punished but openly admitted and used as a learning opportunity. A culture of failure is necessary to foster a strong culture of innovation; the key to developing innovative solutions is openly acknowledging problems and tackling them in a constructive manner.
Measures and tools
- Introduce formats for dealing with mistakes in a constructive manner, e.g., "fuck-up nights".
- Involve the works council in the creation of rules on how to handle mistakes.
- Organize meetings where employees can discuss failed projects and the causes of failure.
- If required, offer further training so that identified mistakes are not repeated in the future.
Establishing a culture of trust, cooperation, and innovation
These three dimensions of an organizational culture are mutually dependent and can very rarely be established independently of one another. The culture of trust and the culture of cooperation/collaboration form the foundation for an effective culture of innovation.
A culture of trust is characterized by the basic assumption that employees are generally motivated to act in the best interests of the organization and its members. Important information is shared rather than held back. Employees have the freedom to organize their own work to attain the goals of the organization.
Sharing information is a prerequisite for successful collaboration in teams and thus for a culture of cooperation. The basic assumption behind a culture of cooperation is that strong team cohesion leads to better work results than if everyone works for themselves. Such a culture of cooperation does not need to be implemented in an absolute form. Many organizations successfully use the concept of "coopetition." These organizations provide performance incentives both for cooperation as well as for competition within the organization (e.g., through a cost center structure).
Common concepts behind a culture of innovation include aspects of the other culture dimensions mentioned here (e.g., trust, collaboration, change). It emphasizes the aspect of promoting innovation within the company. Innovation describes the process by which organizations turn ideas into new products, services or processes in order to gain an advantage in their market. Thinking and acting innovatively is not something that can be prescribed, but general conditions that are conducive to innovation can be established. A culture of innovation promotes the search for new ideas and their transformation into something that really exists and works. The basic assumption behind it is that innovation is important to the success of the organization. It is also based on values and standards that allow creative solutions to be found and implemented in everyday work.
Together, these three culture dimensions enable the potential of Al to be harnessed: Al applications can automate routine tasks and free up employees so that they can use their capacities for more demanding activities. A culture of trust allows for autonomy, providing freedom for creative ideas. A culture of cooperation helps to implement these creative ideas as innovation is primarily the result of teamwork. A culture of cooperation is also an important catalyst for organizational learning: Not only do individual employees learn new skills, but whole team structures adapt to new requirements. This type of organizational learning is key to an organization's innovative capacity. A culture of innovation helps to identify new use cases in a creative process, as well as confront the necessary risks and implement an agile development process with the goal of creating a functioning solution. This also involves rearranging existing processes.
Measures and tools
- Implement flexible and agile working models.
- Transfer decision-making authority to employees or self-organized teams, e.g., when selecting tasks.
- Involve the works council in the design of working hours and the creation of ethical guidelines.
- Managers have the role of supporters so that employees can reach their potential.
- Use of team targets instead of, or in addition to, complementing individual targets.
- Create incentives through works council agreements on the wage structure.
- Transfer decision-making authority to employees or self-organized teams, e.g., when selecting tasks.
- Manager as a supporter and mediator who promotes team cohesion.
- Set up a process for evaluating ideas for innovations and select the most promising ones for an innovation project.
- Provide funding and personnel capacity to enable the implementation of new ideas.
- Network with customers, project partners or even competitors so that you can work together to predict developments and trends, and actively help to shape them.
Establishing a culture of change
The basic assumption behind a culture of change is that continuous change is a positive thing, which benefits the organization more than it harms it, or which may even be essential to its survival. This assumption does not exist in every organization. After all, every company has its own (seemingly) proven recipes for success and processes, and change always costs time and money.
It is often the case that innovations in the company can only be implemented successfully if staff actively support them, e.g., by familiarizing themselves with new software and new processes and engaging in further training.
Measures and tools
- Hold regular meetings where employees can discuss innovations and offer praise or criticism without facing any negative consequences.
- Introduce a company suggestion scheme, taking into account the co-determination right of the works council.
- Enable further training for individual employees so that they can prepare for change in the workplace independently and actively help to shape this change.
Establishing a data culture
The basic assumption behind a data culture is that data is extremely valuable to an organization and its success depends on how effectively this data is used. Data drives decision-making processes. Employees feel responsible for the quality and availability of their data and ensure that it is freely accessible within the organization, provided that this complies with data protection law. They also undertake continuing professional development to be better equipped to work with data. The organization supports this professional development.
Measures and tools
- Offer further training measures on the following topics: statistical bases for data-driven decision-making, the potential and limitations of Al, dealing with results from Al applications and new tools, such as software for data analysis.
- Provide data protection guidelines and processes, including role and rights concepts for data access.
- Provide infrastructure and software for data management and analysis.
Empowerment-oriented leadership
In addition to the cultural dimensions mentioned above, leadership behavior — usually exercised by line managers or project managers — also influences whether employees are able to harness the potential freed up by Al, or whether the effects of introducing Al will fizzle out in this area. Particularly if the organizational culture is based on trust and collaboration, the leadership style practiced in the organization should be in harmony with this. We therefore recommend practicing an empowerment-oriented leadership style in the areas relevant to the use of Al.
Empowerment means enabling employees to act and make decisions independently. There are two main reasons why this is advantageous:
- Al introduction projects are generally characterized by a high level of uncertainty and complexity, even if the company prioritizes the use cases with low complexity and high expected earnings. This complexity relates not only to the Al application itself, but also to the work processes and business processes associated with its use. To ensure that projects can be implemented efficiently and successfully, skilled workers with specialist knowledge from different domains must work together. Rather than being restricted by rigid and detailed instructions, they should be able to act independently using their expertise for the benefit of the project.
- Al applications offer the potential to reorganize work in such a way that people and machines can each do what they do best. (See: "Incorporating human and machine capabilities” above). It is therefore a good idea for any routine tasks or calculations to be delegated to the Al application if they are repeated or can be standardized and are too complex for human intelligence. Meanwhile, employees are given the freedom to do what they do best: making decisions based on experience or taking account of ethical principles, being creative, interacting with people on a personal level and advising or inspiring them. This calls for an appropriate level of autonomy.
Transformational leadership is one example of empowerment-oriented leadership. In this leadership style, managers are characterized by four behaviors:
- Idealized influence: The leader acts as a role model with their ethical behavior, and thereby gains respect, trust and admiration.
- Inspiring motivation: The leader motivates employees with a meaningful and positive vision of the future that people can work toward.
- Intellectual stimulation: The leader encourages employees to find their own solutions to problems.
- Individualized consideration: The leader acts as a coach or mentor who supports employees in their personal development. Employees' individual needs (e.g., relating to flexible working hours) are identified and taken into account.
Employees should be able to decide for themselves how to complete their tasks at work. Instead of writing detailed work instructions, the goals that individuals or teams must achieve are defined. These goals can also be agreed mutually within the team. The goals help employees to understand the work results expected of them. Goals should be SMART ("specific, measurable, achievable, relevant and time-bound").
The manager should actively ensure that goals are always up to date and should remove any obstacles if necessary (e.g., by setting up contacts, providing resources). If a manager adopts a leadership style based on achieving goals but only intervenes when there are problems, this is inadequate; this leadership style is called "passive management by objectives" and has a negative impact on the health of employees.
In all of these endeavors, the leader should pay particular attention to ensuring that the scope for each person to make decisions remains within that person's desired parameters so that nobody becomes stressed through overwork.
Measures and tools
- Diagnosis: You can measure the level of empowerment-oriented leadership free of charge with the KI-ULTRA Evaluation Toolkit.
- Diagnosis: You can measure employee autonomy with the Job Diagnostic Survey (available from the German Federal Institute for Occupational Safety and Health (BAuA)).
- Intervention: Select managers who adopt an empowerment-oriented leadership style.
- Intervention: Based on the diagnostic data, managers should be given preventative training so that they know how to take an empowerment-oriented approach. If available: Include empowerment-oriented leadership in the training catalog or list of values.
Action area: Skills development and learning
Employees are indispensable even in digitalized work systems: They initiate, train and control technical systems as well as map datasets or support technical functions that technology cannot execute on its own. The use of innovative technologies requires employees to be trained in new tasks and may also necessitate the recruitment of new employees. It also opens up new career and development paths as the basis for personal growth and corporate flexibility. The need for qualification not only depends on a specific use case, but reflects the overarching Al strategy of the organization. Organizational learning complements the process of organizational development.
Strategic skills management
To ensure Al introduction projects are a success, employees must develop the relevant skills in a systematic way and these skills need to be made available at the right time. Compared to other digitalization projects, the use of Al can cause working methods to change fundamentally and may require special expertise to be developed (e.g., for handling probabilistic system outputs).
In this context, it can be helpful to implement a systematic skills management process based on the company's Al strategy. One example of a systematic approach to strategy-based skills management is the Fraunhofer "Kompetenz-Kompass" (Skills Compass) [9]. This was developed on the basis of DIN PAS 1093:2009-07 and comprises five steps:
- Link to strategy
- Skills model
- Skills measurement
- Skills development
- Skills review
Step 1 "Link to strategy" and step 5 "Skills review" should generally ensure that suitable tools are used to systematically take into account the impact of strategic business decisions on skills requirements and that the effectiveness of activities aimed at developing skills is reviewed.
With regard to the use of Al, it is usually best if the specialist departments involved work with the Al experts or the IT department to identify the work requirements that arise for employees, then use this as a basis for drawing up skills requirements. The Fraunhofer Skills Compass can be applied to steps 2 "skills model" and 3 "skills measurement": In step 2, you identify and describe the skills required for the relevant department so that the new tasks or the tasks changed by Al can be completed. You also define the skill level required to complete the task (basic, advanced or expert level). In step 3, this target skills profile is compared with the actual skills profile (i.e., the skills currently available), in each case at the relevant skill level (Figure 5). This enables you to detect at an early stage which skills need to be developed in step 4 and to what extent.
Guiding questions
- How will the skills requirements for employees change if the Al strategy is implemented as planned?
- Will the use of Al change work procedures and processes in such a way that the works council will need to be involved in the planning stage?
- When will the skills be needed and how long is it likely to take for these to be made available in the organization (through further training or new hires)?
- Do you know how the existing (actual) and required (target) skills can be measured and how the entire skills management process can be evaluated?
Measures and tools
- Fraunhofer Skills Compass "Vorgehensweise zur Kompetenzbedarfsermittlung in Organisationen" ("Procedure for identifying skills requirements in organizations") [9] with guiding questions and tools for strategic skills management.
Helping employees learn how to use AI
The use of Al should improve the agility and resilience of value creation processes. Agility and resilience are the result of using the dedicated abilities of workers. The ability to work in a motivated and self-organized manner can only develop in conjunction with work processes where these abilities are required. Employees' abilities to self-motivate (i.e., goal setting), to efficiently pursue goals and to self-evaluate their work are crucial here. Self-motivated learning and working should not be taught in abstract terms.
Work activities have a motivational potential, which comes to the fore by introducing work design measures that enhance skills. This promotes intrinsic motivation, reinforces attitudes and leads to the development of appropriate coping strategies. Self-motivation requires specific work content that encourages motivation and learning.
Promoting the capacity to learn is a prerequisite for the development of employees in digitalized work processes. However, this is not an ad-hoc solution. Instead, it has to be a component of HR development and work design.
Guiding questions
- Which learning requirements can be implemented in Al-supported work processes, depending on job requirements and individual performance prerequisites?
- How can a systematic change in activities contribute to maintaining the capacity to learn? How does this affect staff deployment and career and development paths? Do the changes in activities involve transfers that require co-determination?
- How can the capacity to learn be promoted, e.g., through hands-on learning in the use of digital components or through retraining?
- How can operational training measures be developed and carried out with the involvement of the works council?
- How can participative ways of working help to maintain the capacity and willingness to learn, e.g., by allocating organizational tasks to affected employees in quality circles?
- Can tandem learning be established where employees contribute their digital expertise and specialists reciprocate with their technological know-how, i.e., by sharing experiences?
Open Lab Days: http://s.fhg.de/old
Promoting the organizational capacity to learn
Growing the organizational capacity to learn is recommended so that organizations and individuals can quickly adapt to newly identified needs and reallocate capacities. Organizational learning is a response to the challenges of organizational change, such as volatility, uncertainty, complexity and ambiguity. It involves establishing scope for action and refraining from making one-sided optimizations of operational efficiency (without an appropriate reference system for innovation and change).
Organizational learning refers to permanent, purposefully introduced behavioral changes in the organization, which are supported by changed processes or structures. A basic requirement for an organization's capacity to learn is continuous communication about changes.
Guiding questions
- How can changes in an organization's environment be predicted by using Al, for example?
- How can open communication on organizational developments be institutionalized and supported through technical measures?
Action area: Data strategy and technology transition
Accuracy of AI
The type of Al method used, the available data and the task itself can determine the accuracy of the results. The quality of the results in turn influences a project's return on investment. Setting reasonable expectations of the success of an Al project is only possible if you understand which factors will affect its accuracy. For example, image recognition applications are generally more reliable than forecasting applications. Just because an Al solution worked well for one task does not mean that it will also be suitable for other tasks. As a rule, Al models cannot be easily transferred to other areas. This means that Al projects must always be guided by specific processes and the available datasets.
Establishing the data strategy
Data forms the foundation of Al projects. While the data required for certain use cases may be available or can be collected during a project, there are many use cases that need a significantly long data history. This raises the question of how the organization will handle the process of collecting data in the medium and long term.
In the same way that processes evolve, datasets that you need for operating an Al model may also change at short notice. Whenever more up-to-date data becomes available, the effectiveness and accuracy of the Al application are affected.
With a growing volume of processed data and associated application software, the question of a unified strategy for data management becomes increasingly important. Growing systems normally result in a significant amount of technical and organizational work on data integration and maintenance. In the medium and long term, this uses up too many resources. There is therefore a need to decide on data consistency and planned structures.
Data storage can either be in the form of large data warehouses (i.e., large volumes of preprocessed data stored for their application purposes) or data lakes (i.e., large volumes of data consistently stored in their raw form) or in the form of many individual databases. In addition to the technical storage of data and data access, there are also organizational challenges: implementing the appropriate role and rights concepts, avoiding redundancies, complying with data protection requirements and implementing suitable documentation processes (especially for the processing of personal data).
A successful data strategy also takes into account legal aspects. When designing the data strategy, it is important to know which data may be used for which purposes as well as to what extent the data to be used for an Al system needs special protection. It must always be ensured that the data can be used in compliance with the law. In terms of personal data, this is only possible if the relevant legal bases exist for the data use and if the data subjects are fully informed of the processing purposes. If data was collected in the past for different purposes (e.g., collecting the master data of an employee for the purpose of handling the employment relationship), the processing purposes can only be changed subsequently if certain requirements are met (purpose limitation principle). For this reason, the data protection officer should be involved in the development of the data strategy as soon as possible.
Aside from data protection considerations, it is possible, in principle, to use data without restrictions. This applies to both self-generated data and data procured from third parties as there is no legal recognition of data ownership, and copyright (or copyright-like law) on raw data, such as measurement data or statistical data, only applies in very few cases. However, this may need to be assessed if data comprises texts or images (see "Minimizing legal risks" above).
If data from publicly available sources or data providers is used, all applicable data licensing regulations must be followed, particularly those that may preclude the use of the respective datasets in Al systems or require additional license fees to do so. To avoid "under-licensing," it is important to document (in the context of a comprehensive data strategy) whether data that was not collected in-house is subject to such restrictions. Appropriate documentation processes also strengthen data management and provide a clearer overview of the data available in the respective organization.
It should also be noted that the data available in an organization may be subject to confidentiality agreements with third parties or may even be classified as trade secrets that must be actively protected pursuant to the German Trade Secrets Protection Act (GeschGehG). Missing documentation on the confidentiality status of the respective data may mean that such data is processed during the use of Al systems and consequently made available to third parties because an Al system cannot readily differentiate between “ordinary" and sensitive data. Disclosing such data may constitute a violation of confidentiality agreements and thus result in contractual penalties. It may also lead to the loss of the organization's own trade secrets.
As the results generated using Al systems are generally not protected by copyright law but can infringe the rights of third parties (see "Minimizing legal risks" above), it is important in every use case to establish how the results obtained using Al systems will be used. As Al results fundamentally lack copyright protection, if they will be available to third parties, it is important to establish whether transferal to third parties will occur under a licensing agreement and what rights and obligations the parties are subject to.
Guiding questions
- Which datasets are absolutely necessary for the Al application to be useful and profitable?
- In which operational units and to what extent is data collected for medium- and long-term purposes?
- How is data quality ensured when collecting data?
- What are the requirements for data storage and data processing?
- Is the usability of data restricted, for example, due to data licensing regulations? Is the data subject to confidentiality agreements with third parties or does it need to be actively protected because it constitutes trade secrets?
- Is there documentation on the data available in the organization?
- Which employees and organizational units can access which data, both now and in the future, using which means of access?
- Which type of data storage fits the organization's requirements?
- Are there different requirements for different locations or divisions, based on the structure and size of the organization?
- For which processing purposes was personal data collected in the past? Which processing purposes were data subjects informed about? Are these processing purposes compatible with the proposed new use of Al?
Measures and tools
- Document the data available in the organization together with possible restrictions based on missing rights or confidentiality requirements.
- Create a role and rights concept that is as broad as possible.
- Conclude a works council agreement on the introduction and application of the programs required for Al.
- Make a conscious decision on the type and scope of data warehousing and how they are integrated into the organizational infrastructure
- Define target processes for general data processing application cases (e.g., using the Unified Modeling Language)
- Define how results generated using Al systems may be used.
- Use (data) licenses for transferring results generated using Al systems.
Technology transition
In order to be able to work with Al applications in a sustainable manner, it is important to regularly maintain and update the technical infrastructure. It may even be necessary to introduce new technologies to the organization (technology transition). These may include data collection systems (sensors) and data storage systems (e.g., dedicated databases, data warehouses or data lakes), as well as new software for data analysis (e.g., statistical programs).
One typical technological change is the transition from traditional on-premises IT to the cloud paradigm, which involves a complete overhaul of working methods, responsibilities and technologies. While these technical changes are necessary to remain operational in the long term, they do initially incur costs and tie up personnel resources. In particular, employees must be trained in the new technologies, and processes and roles may need to be redefined.
The transition from one technology to another should be planned in a centralized manner where possible. Important stakeholders — particularly future users and employee representatives — should be involved in this planning process at an early stage. The objective should be to create an integrated infrastructure, rather than establishing incompatible isolated solutions. This is the only way that data from across the organization can be combined and used efficiently. Having said that, such changes involve costs and can be tiring for employees over time. Some technological advances can be skipped or introduced at a later date.
Guiding questions
- Is the technology transition planned centrally so that incompatible isolated solutions are avoided?
- How will we decide which technological changes are important/indispensable and which can be postponed or avoided?
- Have all processes, roles and responsibilities been defined for the new technology and have the people affected been informed of the changes? Is compliance with data protection law still guaranteed even after the transition (e.g., will personal data that is no longer required be deleted or anonymized)?
- Have the necessary maintenance processes been included in the planning and are the tools required for this available?
- Are the relevant stakeholders, especially users, included in the planning at an early stage?
- What co-determination rights does the works council have? When should the works council be involved?
- Is it clear how a planned technology transition is expected to contribute to achieving the organizational goals?
- Have the expected benefits of a technology transition project been documented and communicated openly?
- Have resources (working hours, budget) been allocated for training employees on how to work with the new technology?
- Is there a clear timeline for the implementation?
- What difficulties should be expected as a result of process dependencies, downtime and data migration?
- What is the expected lifetime of the new system?
- Have you defined how the technology transition process will be monitored and who will ensure any necessary adjustments are made if unexpected problems arise?
Measures and tools
- Carry out a SWOT analysis to weigh up the benefits and risks associated with a technology transition.
Literature
- Proposal of the European Commission for an EU Regulation laying down harmonized rules on artificial intelligence (AI Act) (2021/0106).
- Daugherty, P. R.; Wilson, H. J.: Human + Machine. Reimagining Work in the Age of AI. Harvard Business Review Press, 2018.
- Huchler, N.; Adolph, L.; André, E.; Bauer, W.; Bender, N.; Müller, N.; Neuburger, R.; Peissner, M.; Steil, J.; Stowasser, S.; Suchy, O. (ed.): Kriterien für die menschengerechte Gestaltung der Mensch-Maschine-Interaktion bei Lernenden Systemen (Criteria for the human-centered design of human-machine interaction in learning systems). Munich: 2020. https://www.plattform-lernende-systeme.de/files/Downloads/Publikationen/AG2_Whitepaper2_220620.pdf.
- Huchler, N.: Komplementäre Arbeitsgestaltung. Grundrisse eines Konzepts zur Humanisierung der Arbeit mit KI (Complementary work design. Outline of a concept for the humanization of work with AI). Zeitschrift für Arbeitswissenschaft 76 (2022) pp. 158–175.
- Pokorni, B.; Braun, M.; Knecht, C.: Menschzentrierte KI-Anwendungen in der Produktion. Praxiserfahrungen und Leitfaden zu betrieblichen Einführungsstrategien (Human-centered Al applications in production. Practical experiences and guideline for operational introduction strategies). Project report in the Advanced Al Center for Learning Systems. Stuttgart: Fraunhofer IAO: 2021.
- Floridi, L.; Cowls, J.: A Unified Framework of Five Principles for AI in Society. Harvard Data Science Review. https://doi.org/10.1162/99608f92.8cd550d1, published June 23, 2019.
- Krafft, T. D.; Zweig, K. A.: Transparenz und Nachvollziehbarkeit algorithmenbasierter Entscheidungsprozesse (Transparency and traceability of algorithm-based decision-making processes). https://www.vzbv.de/sites/default/files/downloads/2019/05/02/19-01-22_zweig_krafft_transparenz_adm-neu.pdf
- Schein, E. H.: What is culture. In Frost, P.; Martin, J., Moore, L.; Lundberg, C.; Louis, M. (ed.): Reframing organizational culture. Washington: Sage, 1991, pp. 243–253.
- Schnalzer, K., Karapidis, A., Dworschak, B., Mozer, P. (2022). Vorgehensweise zur Kompetenzbedarfsermittlung in Organisationen (Procedure for identifying skills needs in organizations), Stuttgart. https://publica.fraunhofer.de/handle/publica/419707
Imprint
Fraunhofer Institute for Industrial Engineering IAO
Nobelstraße 12
70569 Stuttgart
Germany
www.iao.fraunhofer.de
Authors
Martin Braun, Jan-Paul Leuteritz, Maike Link, Helmut Zaiser, Damian Kutzias
Editors
Oliver Riedel, Katharina Hölzle, Wilhelm Bauer, Matthias Peissner
Contact
Press and Public Relations Fraunhofer IAO
Nobelstraße 12
70569 Stuttgart
Germany
presse@iao.fraunhofer.de
Fraunhofer Publica
http://dx.doi.org/10.24406/publica-4419
© Fraunhofer IAO, 2025
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License: https://creativecommons.org/licenses/by-sa/4.0/legalcode.en
The legal expertise was provided by the law firms BHO Legal and Oppenländer.
Über diesen Leitfaden
Dieser auf strategischen Wandel ausgerichtete Leitfaden unterstützt Sie, Anwendungspotenziale der Künstlichen Intelligenz zu erkennen und diese unternehmensspezifisch nutzbar zu machen. Darüber hinaus werden Herausforderungen bei der ganzheitlichen Einführung von KI-Anwendungen aufgezeigt. Ein Schwerpunkt ist die menschenzentrierte Gestaltung der KI-gestützten Arbeit. Nur wenn sich Mensch und Künstliche Intelligenz wechselseitig bestärken, lassen sich bestehende Geschäftsprozesse optimieren oder neuartige Geschäftsmodelle erschließen. Der vorliegende Leitfaden fasst vielfältige Arbeits- und Diskussionsergebnisse im Projekt KI-ULTRA zusammen. Allerdings gibt er nur eine grundsätzliche Orientierung für relevante Themen, Zusammenhänge und betriebliche Vorgehensweisen. Es gilt, diese Impulse im jeweiligen Unternehmenskontext zu reflektieren; dies erfolgt erfahrungsgemäß am besten in einem Projektteam unter Beteiligung von Fachexperten, Unternehmensführung und Betriebsrat. Zugleich sind die Mitarbeitenden mit ihren Bedürfnissen und ihrem Erfahrungswissen in die Gestaltungs- und Anwendungsprozesse einzubeziehen, um den Veränderungsprozess zu einem Erfolg zu bringen. Im Gespräch mit diesen Experten und Stakeholdern können Sie fundiert beurteilen, inwiefern die Nutzung von Künstlicher Intelligenz das betriebliche Geschäftsmodell und die Unternehmensstrategie beeinflusst. Dies gilt im Besonderen für rechtliche Fragestellungen, die vor jedem Einsatz von KI im Einzelfall zu prüfen sind und in diesem Leitfaden überblicksartig skizziert werden. Dieser Leitfaden stellt jedoch keine Schritt-für-Schritt-Anleitung dar, die in der vorgegebenen Reihenfolge abzuarbeiten wäre. Er ist vielmehr in sechs Handlungsfelder untergliedert, die Sie jeweils separat umsetzen können:- Betriebliche KI-Strategie
- Menschenzentrierte Gestaltung von Veränderungen
- Strategien der Arbeitsgestaltung
- Soziale Rahmenbedingungen
- Kompetenzentwicklung und Lernen
- Datenstrategie und Technologietransition
Adressat*innen
Der Leitfaden richtet sich an privatwirtschaftliche Unternehmen, Nichtregierungsorganisationen (NGOs) und öffentliche Einrichtungen, die beabsichtigen, Anwendungen der Künstlichen Intelligenz zu implementieren und zu nutzen. Hauptadressaten sind Entscheidende auf der Ebene des betrieblichen Managements, der Projektleitung und der Interessensvertretung der Mitarbeitenden. Darüber hinaus werden sämtliche Akteur*innen angesprochen, die sich mit der Entwicklung und der Anwendung von Künstlicher Intelligenz im Arbeitskontext beschäftigen. Icons markieren im Text, welche Handlungshilfen für welche betriebliche Akteursgruppe von besonderem Interesse sind. Die Erfahrung zeigt: Gute, nachhaltige Lösungen entstehen im Dialog. So können vielfältige Perspektiven reflektiert werden. Suchen Sie das Gespräch mit Fachleuten oder schließen Sie sich überbetrieblichen Arbeitskreisen an.Leseanleitung
Dieser Leitfaden ist in sechs Handlungsfelder gegliedert. Jedes Handlungsfeld beinhaltet mehrere Handlungshilfen, die jeweils für ein bestimmtes Thema möglichst konkrete und eindeutige Handlungsanleitungen beinhalten, um Ihre Entscheidungs- oder Handlungsfähigkeit zu fördern. In einigen Fällen aber kann der Leitfaden Ihnen aber nur allgemeine Hinweise geben bzw. Fragen stellen, mit denen Sie sich beschäftigen sollten, um passende Antworten für Ihre Organisation zu finden. In anderen Fällen sollten Sie externe Berater*innen hinzuziehen, um Methoden erfolgreich anzuwenden. In den Handlungshilfen gibt es folgende Elemente: Titel Zum Beispiel: Strategische Ziele definieren Icons
Beschreibung
Die Beschreibung ist ein Text, der das Problem und den Lösungsansatz kurz darstellt.
Leitfragen
Die Leitfragen geben eine zusätzliche Hilfestellung und ergänzen die Beschreibung. Versuchen Sie, sich selbst die Leitfragen zu beantworten. Die Antworten können je nach Organisation und Branche unterschiedlich ausfallen.
Maßnahmen und Werkzeuge
Wo es möglich ist, verweisen wir auf eigene oder von Dritten entwickelte Methoden und Instrumente, die Sie einsetzen können, um die Leitfragen zu beantworten. Diese Methoden sind frei verfügbar.
Wann lohnt es, sich mit KI zu beschäftigen?
Was meinen wir mit »Künstlicher Intelligenz«?
Es gibt keine eindeutige und allumfassende Definition von »Künstlicher Intelligenz«. In diesem Leitfaden meinen wir mit diesem Begriff Technologien, die bestimmte Verhaltensweisen imitieren können, welche durch herkömmliche Software-Algorithmen nicht beherrschbar wären. Wir beziehen uns dabei sowohl auf standardisierte, funktional eng umrissene KI-Anwendungen (z. B. Verfahren zur Text- oder Spracherkennung) als auch auf Lösungen, die anwendungsspezifisch entwickelt oder angepasst werden müssen. Die Europäische Kommission geht davon aus, dass der Einsatz Künstlicher Intelligenz zur Verbesserung von Prognosen, zur Optimierung von Abläufen und der Ressourcenzuweisung sowie zur Personalisierung der Diensteerbringung beiträgt. Dies kann für die Gesellschaft und die Umwelt von Nutzen sein, sowie Unternehmen und Wirtschaft Wettbewerbsvorteile verschaffen. Auch in rechtlicher Hinsicht ist eine Definition von »KI« nicht ohne weiteres möglich. Der am 14.06.2023 im EU-Parlament diskutierte Entwurf einer KI-Verordnung (KI-VO-E) auf EU-Ebene spricht bei der Definition von KI-Systemen in Art. 3 Abs. 1 allgemein von »ein[em] maschinengestützte[n] System, das so konzipiert ist, dass es mit unterschiedlichem Grad an Autonomie operieren kann und das für explizite oder implizite Ziele Ergebnisse wie Vorhersagen, Empfehlungen oder Entscheidungen hervorbringen kann, die das physische oder virtuelle Umfeld beeinflussen.« Da davon auszugehen ist, dass das für die Zukunft zu erwartende Inkrafttreten der KI-Verordnung der EU einen erheblichen Einfluss auf die Nutzung von KI haben wird, muss der Gesetzgebungsprozess im Auge behalten werden. Wahrscheinlich ist, dass bei der Regulierung von KI-Systemen ähnlich wie bei der Datenschutzgrundverordnung (DSGVO) ein risikobasierter Ansatz gewählt wird, bei dem die Anforderungen an die Erstellung und Nutzung von KI-Systemen graduell höher sein werden, je mehr Risiken der Einsatz der entsprechenden KI darstellt. Künstliche Intelligenz (KI) ist definitionsgemäß eine in Rechenmaschinen ablaufende Abfolg mathematischer Modelle; diese zeichnen mithilfe spezieller Schnittstellen menschliches und maschinelles Verhalten mittels umfangreicher Daten auf, um es mathematisch zu verarbeiten und zu imitieren. Die Datenmuster werden im Trainingsprozess erfahrenen KI-Instruktoren vorgelegt, die festlegen, wie die Künstliche Intelligenz auf einen Input angemessen reagiert. Der antrainierte KI-Automat reagiert mit seinem heuristischen Methodenset fortan auf von ihm selbst identifizierte Datenmuster. Damit ist KI in der Lage, menschliche Aufgaben mit einem Grad der Angemessenheit zu bewältigen, der dem des Menschen zumindest ähnlich ist.Potenziale des KI-Einsatzes
Die hier dokumentierten Handlungshilfen sollen die Vorstellung klären, warum und wie Sie KI in Ihrer Organisation einsetzen, sei es in einem Pilotprojekt oder im größeren Rahmen. Dafür ist ein Grundverständnis der Möglichkeiten und Grenzen von KI-Anwendungen erforderlich. Der Einsatz von KI soll einen Beitrag dazu leisten, dass ein bestimmtes Problem in der Organisation besser gelöst wird. Praxisbeispiele geben Orientierung; man findet sie z. B. auf der KI-ULTRA Website oder im Wissenspool der Zukunftszentren. Wir empfehlen einen menschenzentrierten Gestaltungsansatz für die KI, um deren Wirksamkeit im betrieblichen Einsatz zu verbessern. Das menschliche Gehirn und die Künstliche Intelligenz schließen sich nicht aus, sondern ergänzen einander. Kein KI-System kann mit der Leichtigkeit und Effizienz mithalten, mit der Menschen lernen, verstehen und handeln. Menschen haben kognitive Fähigkeiten, die auch kreative Inspiration und zwischenmenschliche Empathie einbeziehen. Menschen unterscheiden etwa zwischen belebten und unbelebten Gegenständen. Sie greifen neue Begriffe schon nach ein- oder zweimaligem Hören auf und lernen durch Versuch und Irrtum. Umgekehrt beherrscht Künstliche Intelligenz vieles, was der Mensch nicht oder nur schwerlich kann: Muster in riesigen Datenmengen erkennen; komplexe Fertigungsprozesse steuern; Satellitenbilder analysieren, damit Landwirte ihre Ernteerträge steigern können; Finanzbetrug aufdecken; Verbraucherpräferenzen ermitteln. Aber vor allem ist KI in der Lage, arbeitsteilige Einzelaktivitäten von Menschen und Maschinen produktiv zu koordinieren. Im Vergleich zu anderen Digitalisierungslösungen kann man KI-Anwendungen üblicherweise nicht »von der Stange kaufen«. Sie müssen oft betriebsspezifisch angepasst werden. Daher ist der Einsatz von KI (im Vergleich zu anderen Technologien) in vielen Fällen zeit- und kostenaufwendiger, bis die Technologie ihre Stärken ausspielen kann. Auch sind die Unsicherheiten ihres Einsatzes besonders groß: Ob ein Projekt erfolgreich sein wird, ist oft erst spät feststellbar. Welcher Ansatz zum Ziel führt, ist durch wiederholtes Ausprobieren zu erschließen. KI kann damit u. a. dazu beitragen, dass- neue Dienstleistungen oder Produkte angeboten werden können, oder gegenüber der Konkurrenz ein Qualitätsvorsprung erreicht wird (z. B. gezielte Produktvorschläge für Bestandskunden; prädiktive Maschinenwartung vereinfachen);
- Prozesse effizienter gestaltet werden, indem Teilschritte automatisiert werden;
- Mitarbeitende von monotonen und unterfordernden Aufgaben entlastet werden;
- unerfahrene Mitarbeitende bei einer Tätigkeit gezielt unterstützt werden, sodass vor dem Hintergrund des (regionalen) Fachkräftemangels eine größere Anzahl von Bewerber*innen für eine Stelle infrage kommt.
Grenzen des KI-Einsatzes
Grenzen für den Einsatz von KI-Systemen ergeben sich hauptsächlich aus den aktuell gegebenen technischen Möglichkeiten, aus gesetzlichen Rahmenbedingungen und dem, was durch Mitarbeitende als akzeptabel eingestuft wird. KI-Systeme stoßen z. B. an Grenzen, wenn emotionale Fähigkeiten oder komplexes Abwägen imitiert werden sollen, etwa bei der Personenauswahl oder der Lösung sozialer Konflikte:- KI-Systeme haben keine Emotionen. Damit können sie auch nicht empathisch handeln.
- KI-Systeme verfügen über keine moralische Urteilsfähigkeit. Die Fähigkeit des komplexen Abwägens moralischer Gründe kann nicht maschinell simuliert werden. Nur Menschen können moralisch abwägen.
- KI-Systeme verfügen über kein Problembewusstsein: Sie snd zwar in der Lage, gewisse Denkprozesse zu imitieren. Dennoch liegen dem maschinellen Prozess kein verständiges Erfassen, keine Intuition und keine Einsicht zugrunde.
- Sie kommen zur Überzeugung, dass der KI-Einsatz keinen wesentlichen Beitrag zum Erreichen der Ziele Ihrer Organisation leisten kann. Damit besteht kein weiterer Handlungsbedarf in diesem Themenfeld.
- Sie streben einen ganzheitlichen Gestaltungsansatz an, der beim KI-Einsatz die menschlichen Fähigkeiten und Leistungsbeiträge berücksichtigt. So werden mögliche unerwünschte Nebenwirkungen des KI-Einsatzes kompensiert und die Fähigkeiten der Mitarbeitenden durch den KI-Einsatz gefördert.
Rechtliche Vorgaben für den KI-Einsatz
Neben den technisch und organisatorisch bedingten Grenzen des KI-Einsatzes ist stets zu berücksichtigen, dass sich der konkrete Einsatz von KI-Systemen in vielerlei Hinsicht innerhalb der geltenden (und aktuell teilweise einem starken Wandel unterworfenen) gesetzlichen Bestimmungen bewegen muss, wobei hier nicht nur die bereits erwähnte KI-Verordnung zu beachten ist. Besonderes Augenmerk ist bei der strategischen Planung des KI-Einsatzes insbesondere auf folgende Bereiche zu richten: KI-Einsatz generell:- Rechteklärung bei der Beschaffung von Trainingsdaten
- Nutzung der KI durch den Anwender / Eingabekontrolle (z. B. Datenschutzrecht, Geschäftsgeheimnisse)
- Arbeitsrecht
- Datenschutzrecht
- Allg. Persönlichkeitsrecht
- Schutzfähigkeit von mittels KI produzierten Daten
- Gesetzliche Schutzrechte an den Daten
- Patente / Gebrauchsmuster
- Design (im gesetzlichen Sinne)
- Urheberrechte / Leistungsschutzrechte
- Vertraglicher Schutz der Daten
Handlungsfeld: Betriebliche KI-Strategie
Strategische Ziele definieren
Wenn Ihr Betrieb über die KI-Einführung nachdenkt, sollten Sie vorab klären, welche Erwartung Sie mit der Einführung verbinden, welche ersten Projekte sich dafür eignen und ob Sie langfristig über die entsprechende Expertise in Ihrer Belegschaft verfügen. Falls Sie zum ersten Mal vor einem KI-Projekt stehen, kommt Ihnen das vermutlich sehr herausfordernd vor. Wenn Sie wissen, worauf Sie sich während dieses Prozesses zu konzentrieren haben, erscheint das Vorhaben in einem anderen Licht. Die Entwicklung einer KI-Strategie ist zweckmäßig, wenn einer der folgenden Punkte auf Ihren Betrieb zutrifft:- Sie planen die umfassende Einführung von KI-Anwendungen.
- Für die geplante KI-Nutzung wird über Abteilungsgrenzen hinweg kooperiert.
- Für den erfolgreichen Einsatz von KI, z. B. auch für das Aufsetzen von Folgeprojekten, haben Sie die Unterstützung der relevanten Entscheidungsträger gesichert (z. B. Unternehmensführung, Datenschutzbeauftragte*r, weitere Stakeholder).
- Welche wirtschaftlichen Ziele verfolgen wir mit dem Einsatz von KI (z. B. Marktstellung verbessern, Kosten senken, Kooperationsfähigkeit mit Kund*innen erhalten)?
- Wie trägt die KI-Anwendung bei, diese Ziele zu erreichen, und wie ist die Zielerreichung nachzuvollziehen (z. B. Erschließung neuer Marktpotenziale, Automatisierungseffekte)?
Klären Sie den betrieblichen Nutzen des KI-Einsatzes. Eine der ersten Fragen lautet: »Warum meinen wir, dass sich diese Investition lohnen wird?« Dafür müssen Sie wissen, welche Probleme (z. B. Engpässe im Betriebsablauf) Sie angehen oder wo Sie Effizienz und Innovationskraft verbessern wollen (etwa bei der Digitalisierung von Absatzkanälen). KI-Projekte sollten sich mit Prozessen befassen, die sich konkret auf Aufwände und Ressourcen auswirken und die betriebliche Profitabilität merklich beeinflussen.
Wie bedeutsam die strategische Nutzung von KI für die Organisation ist, wird beispielhaft an 3 Stufen dargestellt:
- Stufe 1: KI-Potenziale sind (aktuell) für den Erfolg der Organisation von geringer Bedeutung. Kern der Strategie ist es daher zunächst, die Digitalisierung voranzutreiben, um Effizienzgewinne zu realisieren und ggf. für die Zukunft bessere Voraussetzungen für den KI-Einsatz zu schaffen.
- Stufe 2: In einzelnen Anwendungsbereichen werden Verbesserungen durch KI angestrebt (z. B. in Fertigung und Montage); alle anderen möglichen Anwendungsfelder (z. B. Customer Relations Management) arbeiten wie bisher weiter.
- Stufe 3: Das zukünftige Geschäftsmodell basiert wesentlich auf dem Einsatz von KI; ein umfassender KI-Einsatz wird als zentraler Erfolgsfaktor betrachtet.
Leitfragen
- Welchen Nutzen erwarten Sie von KI-Anwendungen insgesamt im Unternehmen?
- Welche strategischen Ziele werden durch den Einsatz von KI-Anwendungen erreicht?
- Welche konkreten Aufgaben wollen Sie angehen, um Erfahrungen mit KI zu sammeln. Wo wollen Sie etwa Effizienz oder Innovationskraft stärken?
- Welche Arten von KI-Anwendungen kommen für Ihre Organisation infrage – und in welchem Umfang?
- Welche Kompetenzen zur Einführung und Umsetzung von KI-Anwendungen sollen vorgehalten werden?
- Sollen KI-Anwendungen eingekauft oder (langfristig) selbst entwickelt werden (siehe: Strategischer Ansatz zu »make or buy«)
- Wie erkennen Sie, dass KI evtl. doch nicht zur Lösung der drängenden Probleme beiträgt? Unter welchen Bedingungen werden KI-Projekte abgebrochen?
- Wie kann sichergestellt werden, dass sich der KI-Einsatz positiv auf die Situation der arbeitenden Menschen auswirkt? Fallen Arbeitsplätze weg?
- Besteht weitergehender Qualifikationsbedarf? Verändert sich die Art der Kommunikation?
- Wie wird die KI-Strategie datenschutzkonform umgesetzt?
- Wie wird sichergestellt, dass alle relevanten Stakeholder beteiligt werden?
- Welche zentralen Geschäftsprozesse werden vom geplanten KI-Einsatz berührt?
- Wird eine mitbestimmungspflichtige Änderung der Betriebsorganisation angestrebt?
- Welche Ressourcen werden für die Umsetzung der KI-Strategie bereitgestellt?
- Welche zeitlichen Freiräume werden dafür gewährt?
Maßnahmen und Werkzeuge
- Festlegung unter Einbindung relevanter Stakeholder, welchen Beitrag KI zur Erreichung der Vision bzw. der Ziele der Organisation leisten soll.
- Einbindung eines möglichst ranghohen Entscheidungsträgers, um die Umsetzung der KI-Strategie zu begleiten, und durch Freigabe von Ressourcen (z. B. Projektmitteln) zu ermöglichen.
- Frühe Einbeziehung des Betriebs- bzw. Personalrats, um potenziellen Ängsten entgegenzuwirken und eine frühzeitige Kommunikation für maximale Transparenz zu etablieren.
- Einbindung des Datenschutzbeauftragten und Einplanung einer ggf. erforderlichen Datenschutz-Folgenabschätzung von Anfang an.
- Einbeziehung der Rechtsabteilung, um sicherzustellen, dass Prozesse rechtskonform ausgestaltet werden.
- Regelmäßige Überprüfung und ggf. Anpassung von Vision und Strategie.
- Veranschaulichung und Kommunikation der KI-Strategie anhand einer Strategy Map (vgl. Abbildung 1).
- Bei der Formulierung einer Unternehmensstrategie kann ein externer Coach den Blick ausweiten, die Diskussion reflektieren und kritische Verständnisfragen stellen.
Transformationsbedarfe ermitteln
Um aus der KI-Strategie Veränderungsmaßnahmen abzuleiten, bietet es sich an, zunächst den Ist-Zustand zu ermitteln, d. h. welche günstigen Rahmenbedingungen für das Ausschöpfen von KI-Potenzialen vorliegen und welche nicht. Als orientierendes Bewertungsinstrument steht das KI-ULTRA Evaluation Toolkit3 zur Verfügung. Zweck der Bewertung ist es, diejenigen strategischen Handlungsfelder zu identifizieren, in denen das größte Verbesserungspotenzial steckt. Jedes Ergebnisfeld des Evaluation Toolkit verweist auf relevante Handlungshilfen in diesem Dokument, die zur Verbesserung der Rahmenbedingungen für den KI-Einsatz herangezogen werden können. Dieser Leitfaden bildet somit ein Reifegradmodell für den KI-Einsatz; selbstverständlich können auch andere Reifegradmodelle verwendet werden. Reifegradbewertungen beziehen sich vor allem auf folgende Kernthemen:- Organisation und Arbeitsmethodik
- Infrastruktur
- Datenbestände und -verfügbarkeit
- Kultur und Zusammenarbeit
- Ethische, soziale und rechtliche Aspekte (engl.: Ethical, Legal and Social Implications – ELSI).
Leitfragen
- Ist Ihr Unternehmen aufgrund der Rahmenbedingungen derzeit überhaupt in der Lage, das Thema KI operativ anzugehen?
- Stehen angemessene Ressourcen und Kompetenzen zur Verfügung?
- Gibt es ein klares Commitment der obersten Unternehmensführung?
- Welche rechtlichen Fragen müssen geklärt werden? Ist die notwendige rechtliche Expertise vorhanden bzw. eingeholt worden?
- Was wäre zu tun, um die nötigen Rahmenbedingungen herzustellen?
- Wen müssen Sie einbeziehen, um die Rahmenbedingungen in den einzelnen Handlungsfeldern zu verbessern?
- Welche Mitbestimmungsrechte des Betriebsrats sind zu beachten? Liegt eine Planung von Arbeitsverfahren oder Arbeitsabläufen vor, über die der Betriebsrat unterrichtet und über die beraten werden muss?
Maßnahmen und Werkzeuge
- KI-ULTRA Evaluation Toolkit
Strategischer Ansatz zu »make or buy«
Sollten Sie Ihre KI-Anwendungen einkaufen oder selbst entwickeln? Diese Frage muss nicht gleich beim ersten KI-Projekt adressiert werden, weil Sie hier vermutlich externe Unterstützung einbeziehen werden. Mittelfristig stellt sich jedoch die Frage des »make or buy«, wenn KI im größeren Umfang eingesetzt werden soll oder wenn die KI-Lösungen ein Alleinstellungsmerkmal der Organisation darstellen sollen. Für diese Entscheidung sind u. g. Leitfragen relevant. Die Gewichtung dieser Aspekte bei der Entscheidungsfindung https://www.ergebnisse.ki-ultra.iao.fraunhofer.de sollte jedoch organisationsspezifisch auf Grundlage der KI-Strategie festgelegt werden. Prognosen über das Kundenverhalten in Branchen wie Versicherungen, Banken oder Werbung.Leitfragen
- Sind geeignete Kompetenzen und Strukturen (z. B. Data Scientists, Data Engineers, Anwendungsdesigner) vorhanden, sodass sich Aufwand und Kosten für die KI-Entwicklung realistisch abschätzen und einer Kostenstelle zuordnen lassen?
- Welche Variante verspricht eine bessere Berücksichtigung gesicherter arbeitswissenschaftlicher Erkenntnisse zur menschengerechten Gestaltung der Arbeit?
- Können eher über den Einkauf von KI oder über die eigene Entwicklung mögliche Bedenken des Betriebsrats zu den Auswirkungen auf die Arbeitnehmer Rechnung getragen werden?
- Entsteht durch den Einkauf der Lösung eine einseitige Abhängigkeit von einem externen Dienstleister z. B. durch Abwanderung von Know-how oder durch Offenlegung von Datenbeständen?
- Welche Folgekosten entstehen beim Einkauf durch externe Dienstleister (z. B. Updates, Wartung)?
- Ist der Dienstleister als zuverlässig bekannt?
- Kann man darauf vertrauen, dass bei Problemen schnell und mit angemessenem Aufwand reagiert wird?
- Gibt es DSGVO-konforme Lösungen für die zu bewältigenden Aufgaben? Welche Vereinbarungen zur Verantwortlichkeit über die Datensätze können mit dem Dienstleister getroffen werden (z. B. zur Erfüllung der Informationspflichten, zur Durchführung von Datenschutz- Folgenabschätzungen, zur technischen Absicherung der Datensätze etc.)?
- Wo werden die Daten verarbeitet?
Systematische Suche nach KI-Anwendungsfällen
Das Vorgehensmodell zur Projektdurchführung beginnt mit einem definierten Anwendungsfall für KI. Eine zentrale Frage für viele Organisationen ist, wie Sie überhaupt zweckmäßige Anwendungsfälle identifizieren können, um »Low Hanging Fruits« zu ernten. Geeignete Kandidaten sind:- Tätigkeiten, die zeit- und arbeitsintensiv sind (wie das Lesen umfangreicher Dokumentationen, um Handlungsbedarfe zu ermitteln).
- Arbeitsabläufe, die eine intensive Bilderkennung erfordern.
- Prozesse, die sich durch eine Sprachanalyse ergänzen lassen (etwa im Kundendienst).
- Ausgehend von den Problemstellungen der Fachabteilungen bzw. der Fachleute (z. B. Mitarbeitende, Qualitätssicherung und Arbeitsplanung)
- Ausgehend von den Potenzialen der KI-Anwendung (d. h. Personen mit Technologieexpertise, Personen mit KI-Expertise)
Ansatz 1: KI-Technologie für Problemstellung suchen
Das problemorientierte Vorgehen (»Market Pull«) stellt früh sicher, dass es für die KI-Anwendung einen konkreten Bedarf gibt, was das weitere Vorgehen vereinfachen kann. Lösungen finden sich oft in bereits identifizierten Problemstellungen eines Geschäftsprozesses. So kann anlassbezogen eine interne Fachabteilung oder ein Dienstleister kontaktiert werden, um den Einsatz von KI zu untersuchen. Es ist auch zu prüfen, ob im Unternehmen bereits geeignete KI-Lösungen vorliegen, die sich anpassen lassen.Leitfragen
- Wurden bei vergleichbaren Aufgabenstellungen bereits KI-Technologien eingesetzt? Wenn ja, welche?
- (Sind eventuell vergleichbare Anwendungen aus anderen Betrieben bekannt?)
- Was macht KI besser oder anders als bisherige Technologien? Wo liegt der Unterschied?
- Verbessert der Einsatz von KI Arbeitsverfahren und Abläufe zugunsten der Mitarbeitenden?
- Welche Optimierungsziele werden durch den Einsatz von KI-Technologien erreicht?
- Welche Geschäftsprozessdaten stehen in angemessener Qualität zur Verfügung? Welche Daten sind zusätzlich zu erheben?
- Unter welchen Voraussetzungen kann die gewählte Technologie für das bekannte Problem rechtskonform eingesetzt werden?
- Sind insbesondere geeignete Trainingsdaten unter Beachtung aller Schutzrechte (siehe: Rechtliche Risiken minimieren) verfügbar?
Ansatz 2: Problemstellung für KI-Technologie suchen
Alternativ können Sie prüfen, ob und welche Problemstellungen der Geschäftsprozesse sich durch KI-Technologien optimieren lassen (datengetriebene Herangehensweise, »Technology Push«). Lassen sich aus der Verfügbarkeit von (großen) Datenbeständen neue Geschäftspotenziale erschließen? Kann man etwa die Kundenbedürfnisse besser ergründen und zielgruppenspezifische Produkte herstellen? Kann man den zeitlichen Verlauf des Kaufverhaltens besser vorhersagen, etwa im Weihnachtsgeschäft? Kann man betriebsintern das Verschleißverhalten von Maschinen besser vorhersagen und Instandhaltungen effizienter planen? Bei dieser Vorgehensweise sollten Personen mit KI-Expertise kontinuierlich die Nutzenpotenziale von KI-Technologien identifizieren und bewerten. Im Rahmen konkreter Projekte formulieren sie Anforderungen an aussichtsreiche Technologien. Diese sollten spezifisch, aber nicht lösungseinschränkend sein. Beispielsweise kann gefragt werden, wie eine Qualitätskontrolle durch Bilderkennung automatisiert werden kann. Bedenken Sie allerdings: KI wird immer nur einzelne Phasen eines Prozesses verbessern; daher ist es schwierig, den Beitrag zum Gesamtergebnis zu berechnen.Leitfragen
- Welchen Wettbewerbsvorteil bringt es der Organisation, wenn wiederkehrbare Muster in einem Prozess erkennbar werden, wenn sich ein Prozessverhalten prognostizieren ließe, wenn Entscheidungen durch statistische Verfahren unterstützt werden?
- Gibt es wiederkehrende, zeit- bzw. aufwandsintensive, aber unabdingbare Aufgaben?
- Gibt es Aufgaben, bei denen die Analyse großer Datenmengen Fehlentscheidungen verringert?
- Können Sie für konkrete Anwendungen aufzeigen, wozu KI fähig ist – und wozu nicht?
- Gibt es Bereiche, in denen sich Arbeitsverfahren und Abläufe zugunsten der Mitarbeitenden verbessern lassen?
Maßnahmen und Werkzeuge
Unternehmen können ihr Vorschlagswesen nutzen, um Problemstellungen und KI-Technologie zusammenzuführen. Bewährt haben sich auch themenspezifische Workshops. Wenn Data Sciencists und Datenexperten im Unternehmen tätig sind, ist die Gründung eines »KI-Hubs« zu erwägen. Dieser kann Folgendes leisten:- Austausch von Expert*innen aus unterschiedlichen Unternehmensteilen fördern, um z. B. bei Problemen in einem Projekt Tipps und Hilfe zu bekommen oder aktuelle Fachinformation zu teilen (z. B. neue Technologien, Erfahrungsaustausch).
- Mitglieder des Hubs können gezielt Fachabteilungen ansprechen, um im Gespräch über die täglichen Arbeitsroutinen Entlastungs- oder Unterstützungspotenziale zu identifizieren.
- KI-Schulungen für die Mitarbeitenden anbieten, um Grundlagenwissen und Kernkompetenzen im Umgang mit KI zu vermitteln.
- Arbeitsmittel zur Verfügung stellen (z. B. digitale Austauschplattform, internes Forum)
- Ausreichende Kapazitäten bereitstellen
- Interne Kommunikation: Die Mitarbeitenden sollten den Hub, seine Mitglieder und Aufgaben kennen.
Handlungsfeld: Menschenzentrierte Gestaltung von Veränderungen
Change-Management, um KI-Potenziale zu nutzen
Die Einführung von KI-Anwendungen kann Veränderungen in der betrieblichen Organisation erfordern. Im Rahmen eines Change-Managements werden diese Veränderungen gezielt herbeigeführt und betroffene Mitarbeitende frühzeitig eingebunden, um die Organisation an die neuartigen Bedarfe anzupassen. Bei einer grundlegenden Änderung muss mit dem Betriebsrat ein Interessenausgleich angestrebt und evtl. gar ein Sozialplan geschlossen werden. Versetzungen von Mitarbeitenden müssen vom Arbeitsvertrag gedeckt sein und sind nur mit Zustimmung des Betriebsrats zulässig. Wenn Sie zunächst einen KI-Piloten planen, um erste Erfahrungen zu sammeln, dann bietet Ihnen der Leitfaden zur Durchführung von KI-Projekten eine Grundlage, um im Verlauf des Einführungsprojekts die wichtigsten Veränderungsbedarfe zu berücksichtigen. Ebenso können Sie in Bezug auf das Change-Management, im Hinblick auf eine spezifische KI-Anwendung, den Leitfaden Einführung von KI-Systemen in Unternehmen der Plattform Lernende Systeme5 anwenden. Die folgende Handlungshilfe richtet sich an Organisationen, die die Potenziale von KI umfassend ausschöpfen wollen und daher KI als Querschnittstechnologie einsetzen wollen. Damit werden unterschiedliche Unternehmensbereiche (z. B. in Personalwesen, Fertigung und Kundenkontakt) zusammengeführt. Ausgangspunkt des Change-Managements ist die KI-Strategie (siehe: Strategische Ziele definieren). Sie legt dar, welchen Beitrag KI zum Erfolg der Organisation leistet und wie dieser Beitrag entsteht. Die für den Einsatz von KI relevanten Zielsetzungen ergeben sich aus den folgenden strategischen Handlungsfeldern (Arbeitsgestaltung, Kompetenzmanagement und Lernen, Datenstrategie) und den entsprechenden Handlungshilfen. Weitere Zielsetzungen können sein:- Stärkung der Rolle des Menschen als Kompetenzträger (im Sinne einer komplementären Funktionsteilung von Mensch und KI)
- Weiterentwicklung der Organisationsstruktur
- Schaffen neuer Stellen oder Abteilungen, um KI-Aktivitäten zu bündeln
- Bereitstellen von Infrastruktur
Leitfragen
Beantworten Sie im Rahmen des Change-Managements folgende Fragen:- Wurde ein Entwicklungsziel definiert, das die Potenziale des Zusammenwirkens von Mensch und KI würdigt?
- Ergeben sich aus der geplanten Nutzung von KI Änderungen des Geschäftsmodells, was Änderungen in der Organisation erfordert?
- Liegt eine mitbestimmungspflichtige Betriebsänderung vor?
- Trägt die Kommunikation bei, qualifizierte Entscheidungen in angemessenem Zeitrahmen herbeizuführen, um die Vorteile von KI zu nutzen?Trägt die Kommunikation bei, qualifizierte Entscheidungen in angemessenem Zeitrahmen herbeizuführen, um die Vorteile von KI zu nutzen?
- Passt unsere Arbeitsweise zur geplanten KI-Einführung? Wie lassen sich KI-Entwicklung und -Einführung unter Beachtung der Mitbestimmungsrechte der Vertretung der Mitarbeitenden beschleunigen?
- Müssen wir neue Rollen oder Geschäftsprozesse definieren?Müssen wir neue Rollen oder Geschäftsprozesse definieren?
KI-Lösungen menschenzentriert entwickeln
Menschenzentrierte Gestaltung von Systemen stellt die tatsächlichen Benutzer*innen sowie weitere Betroffene in den Fokus der Betrachtung und bezieht diese aktiv in den Gestaltungsprozess ein. Müssen wir neue Rollen oder Geschäftsprozesse definieren? Für den Umgang mit personenbezogenen Daten ist z. B. eine für den Datenschutz verantwortliche Person zu bestimmen. Sollen Datenbereiche über Bereichsgrenzen nutzbar gemacht werden, sollten Prozesse definiert werden, die den datenschutzkonformen Zugriff regeln. Bei der Evaluation sollten neben rein funktionalen Zielkriterien auch die relevanten menschbezogenen Auswirkungen Die Zahl der Entscheidungen, die in Unternehmen getroffen werden müssen, nimmt zu. Wesentliche Entscheidungen betreffen die Problemlösung, die Analyse, das strategische Denken oder die Fähigkeit zur Priorisierung. Entscheidungen verlangen von Menschen die Fähigkeit, mögliche Wirkungen dieser Entscheidung abzuschätzen bzw. Alternativen abzuwägen. Dieser Aspekt kann durch KI-Systeme wirksam unterstützt werden. Werden entscheidungsrelevante Daten auf Plattformen geteilt, lassen sich mehr Menschen an Entscheidungsprozessen beteiligen, wodurch deren Akzeptanz zunimmt. Folgende Entscheidungsszenarien sind realisierbar:- Einige KI-Anwendungen sorgen dafür, dass das abschließende Fachurteil von Entscheidern ein stärkeres Gewicht bekommt. Die Nutzung bildgebender Verfahren in der Medizin ist ein Beispiel.
- In anderen Fällen bündelt KI die Entscheidungsfindung, ohne die Entscheidungsträgerschaft zu ändern. Ein Beispiel ist die Personalauswahl.
- Einige KI-Anwendungen zentralisieren den Entscheidungsprozess. Die Validierung des Kreditkarten- Einsatzes ist ein typisches Beispiel.
Leitfragen
- Welche Fähigkeiten der Mitarbeitenden sind unabdingbar für Kundenorientierung und Unternehmenserfolg? Wie werden sie durch KI zweckmäßig ergänzt?
- Welche Einstellungen haben die Mitarbeitenden zu KI und auf welchen Erfahrungen und Annahmen beruhen diese Einstellungen?
- Wie lässt sich ein Grundverständnis von KI in der Breite der Organisation schaffen?
- Welche Formate und Medien bieten sich an, um intern über das Thema KI zu informieren?
- Welche Erfahrungen zur Einbeziehung der Mitarbeitenden und ihrer betrieblichen Interessensvertretung liegen im Unternehmen vor?
- Welche Bedeutung haben sie in Abhängigkeit des Organisationskonzepts? Gibt es Vereinbarungen mit dem Betriebsrat zur Einführung und Anwendung von Programmen?
- Beteiligung von Mitarbeitenden erfordert betriebliche Regeln. Nur so können sich Verbindlichkeit, Vertrauen und Verantwortung im täglichen Handeln entwickeln. Inwiefern ist dies in der Organisationskultur berücksichtigt?
- Ist davon auszugehen, dass betroffene Personen und Mitarbeitende in die Verarbeitung ihrer Daten einwilligen (soweit dies erforderlich ist)?
- Sofern personenbezogene Daten verarbeitet werden sollen: Ist sichergestellt, dass eine rechtliche Verarbeitungsgrundlage besteht?
- Kann die Verarbeitung der Daten auf eine wirksame, freiwillige Einwilligung der jeweiligen betroffenen Personen gestützt werden?
- Wie können die betroffenen Personen vorab verständlich darüber informiert werden, was mit ihren Daten zu welchem Zweck geschieht? (siehe: Personenbezogene und sonstige sensible Daten schützen; siehe: Datenzugriff im Leitfaden zur Durchführung von KI-Projekten)
Maßnahmen und Werkzeuge
- Informationen über den geplanten KI-Einsatz transparent gestalten und öffentlich zugänglich machen
- Mitbestimmung des Betriebsrats ermöglichen
- Planung des Einsatzes von KI rechtzeitig mit dem Betriebsrat beraten
- Austauschformate für alle betroffenen Mitarbeitenden schaffen
Handlungsfeld: Strategien der Arbeitsgestaltung
Auch angesichts des Einsatzes innovativer Technik wird Arbeit auch zukünftig maßgeblich durch die menschlichen Leistungsbeiträge geprägt werden. Ein menschenzentrierter Gestaltungsansatz schafft günstige Voraussetzungen für ein motiviertes und qualifiziertes Handeln und Lernen. Bei der menschenzentrierten Arbeitsgestaltung geht es darum, menschliche Leistungspotenziale und technische Funktionspotenziale in Einklang zu bringen. Wie dies operativ umgesetzt wird, entscheidet sich immer im konkreten Projekt. Einige Grundsätze sollten Sie aber auf strategischer Gestaltungsebene reflektieren; nur so können KI-Anwendungen zweckmäßig ausgewählt und wirkungsvoll eingesetzt werden.Fähigkeiten von Mensch und Maschine einbeziehen
Wie eine angemessene Funktionsteilung von Mensch und Maschine aussieht, entscheidet sich im Einzelfall auf operativer Projektebene. Dennoch sollten Sie entsprechende Voraussetzungen auf strategischer Ebene schaffen. In der Einführung wurden bereits Potenziale und Grenzen von KI-Anwendungen dargestellt (siehe: Potenziale des KI-Einsatzes). Darüber hinaus sollten Sie die Stärken von Menschen gegenüber KI-Systemen berücksichtigen, wie sie z. B. im Modell der fehlenden Mitte dargestellt sind (vgl. Abbildung 2). KI ermöglicht grundsätzlich, menschliche Arbeit zu substituieren (d. h. Automatisierung), zu assistieren und zu augmentieren (d. h. der Verstärkung menschlicher Fähigkeiten). Eine zentrale Aufgabe der Funktionsteilung ist es, Einzelaufgaben der Technik oder den Menschen zuzuordnen. Dabei soll der Mensch einen angemessenen Handlungs- und Entscheidungsspielraum erhalten. Zusammen mit dem Anforderungsniveau der Tätigkeit legt dieser Spielraum fest, welche Lern- und Anpassungsleistungen der Mensch erbringt, um nicht planbare Anteile von Arbeit flexibel zu bewältigen. Eine an den Stärken des Menschen ausgerichtete Funktionsteilung schafft günstige Voraussetzungen für eine agile Arbeitsweise. Insbesondere sollen KI-Anwendungen nicht dazu eingesetzt werden, um Entscheidungen zu treffen, die sich unmittelbar auf den Menschen auswirken; solche Entscheidungen sind vielmehr Gegenstand zwischenmenschlicher Aushandlungsprozesse. Beispiele sind die Schicht- oder Urlaubsplanung. In vielen Bereichen sind (vollständig) automatisierte Entscheidungen durch KI-Lösungen auch aus rechtlicher Sicht nur in engen Grenzen zulässig. Außerdem sollen KI-Anwendungen nicht dazu eingesetzt werden, menschliches Verhalten zu kontrollieren und zu lenken; dieser Ansatz wird auch als Re-Taylorisierung bezeichnet. Hier geben IT-Systeme den Menschen ideale Verhaltensweisen vor (»one best way«) und kontrollieren deren Einhaltung. Bisher gibt es keine Belege dafür, dass sich durch diesen Optimierungsansatz tatsächlich das Ergebnis verbessert. Die Erfahrung deutet jedoch auf vielfältige Probleme hin, wenn menschliches Leistungsverhalten vor allem als eine technische Herausforderung betrachtet wird. Stattdessen soll eine Arbeitsgestaltung die menschlichen Leistungsvoraussetzungen fördern, etwa durch Verständnis und Sinnerleben. Bei KI-Systemen ist die Mensch-Maschine-Interaktion so zu gestalten, dass nicht ausgerechnet jene Arbeitsinhalte und Kommunikationsanteile substituiert werden, die motivierend und lernförderlich wirken. Dies kann beispielsweise der Fall sein, wenn interessante Programmiertätigkeiten an einem Roboter verlagert werden, das routinemäßige Einlegen von Werkstücken in eine Stanzmaschine jedoch bei den Mitarbeitenden verbleibt. Hierzu sind Möglichkeiten zu schaffen, um die Interaktion zwischen Mensch und Maschine transparent zu gestalten. Diese Aufgabe wird im Modell der »fehlenden Mitte« ausgeführt.
Leitfragen
- Welche Aufgaben können zuverlässig vom Menschen übernommen werden? Wie wirkt sich eine Automatisierung durch KI auf die menschliche Erfahrungsmöglichkeit, auf die Tätigkeitsvielfalt und die Motivation der arbeitenden Menschen aus?
- Welche rechtlichen Vorschriften und ethischen Aspekte sind beim KI-Einsatz betroffen?
- Liegt ein Risikomodell mit geeigneten Kritikalitätsstufen vor, um mögliche Risiken der KI-Anwendung zu erkennen und zu minimieren?
- Welche Mitbestimmungsrechte der Interessenvertretung der Mitarbeitenden sind zu beachten?
- Werden automatisierte Entscheidungen getroffen, die keiner menschlichen Kontrolle mehr unterliegen?
- Welche Funktionen können KI-Anwendungen grundsätzlich übernehmen? Inwiefern können sie in das soziale Gefüge des Unternehmens eingreifen?
- Welche Auswirkung hat dies auf die KI-Strategie?
- Ist die KI tatsächlich hinreichend flexibel, um sich Erfordernissen des Betriebs und Bedürfnissen des Menschen situativ anzupassen?
- Oder reichen Methoden einer »klassischen« Automatisierung aus?
- Sofern die KI die menschliche Arbeit substituiert: Ist ein Sozialplan zu schließen?
- Sind Versetzungen nötig?
Maßnahmen und Werkzeuge
- Formen der Funktionsteilung von Mensch und KI definieren und mit allen Beteiligten abstimmen.
- Die strategischen Rahmenbedingungen für alle Entwickler verfügbar machen.
- Maßnahmen einer menschenzentrierten Arbeitsgestaltung ableiten. Hierzu kann auf die einschlägige Literatur zurückgegriffen werden ([2], [3], [4], [5]).
- In kleineren Betrieben kann an diese Stelle ein Workshop mit Mitarbeitenden treten, in denen festgestellt wird, inwieweit die Arbeitsumstände den Gestaltungszielen entsprechen und welche Veränderungen gemeinsam angestrebt werden.
Sicherheitsrisiken minimieren
Technische Systeme sind grundsätzlich sicher und zuverlässig einzusetzen; einschlägige Sicherheitsstrategien sind u. a. in Gesetzen, Verordnungen und Normen geregelt. Bei robotischen und risikobehafteten KI-Systemen steht die Unfallvermeidung im Vordergrund. Unfälle sind zum einen auf technische Parameter zurückzuführen. Zum anderen ist menschliches Fehlverhalten zu vermeiden, das sich bei Monotonie oder psychischer Sättigung häuft. Bei der Implementierung von KI-Anwendungen sind die gesetzlichen Anforderungen des Arbeitsschutzes zu berücksichtigen. Insbesondere kann eine umfassende Gefährdungsbeurteilung erforderlich und der Betriebsrat wegen Fragen des Gesundheitsschutzes zu beteiligen sein. Gleichermaßen ist die zu erwartende risikobasierte Regulierung von KI-Systemen durch eine KI-Verordnung der EU zu berücksichtigen. Ein grundlegendes Kriterium menschenzentrierter Gestaltung betrifft die Schädigungslosigkeit bei der Interaktion von Mensch und Maschine. Die Sicherheitsrisiken einer Maschine sind auf ein allgemein akzeptiertes Maß zu reduzieren. Einen wesentlichen Anteil an diesem Schutzkonzept haben Sicherheitssteuerungen. In KI-Systemen lassen sich Algorithmen und Entscheidungsabläufe nicht mehr detailliert durchschauen und begründet nachvollziehen (d. h. »Black Box«-Verhalten). Folglich schätzen Menschen die Prozesse nicht immer zutreffend ein und greifen zuweilen unangemessen in diese ein. Daher sind immer auch übergeordnete Sicherheitsfunktionen und Notfall-Strategien vorzusehen. Grundsätzlich ist nicht hinreichend bekannt, in welchem Umfang KI-Algorithmen sichere Entscheidungen treffen können. Daher soll das Sicherheitsniveau von KI-Systemen anhand starker Indizien vertrauenswürdig belegt werden. Dazu zählen Kriterien wie Erklärbarkeit und Nachvollziehbarkeit von Entscheidungen, Übertragbarkeit auf unterschiedliche Situationen, Überwachung während der Laufzeit und transparente Formen der Mensch-Maschine-Interaktion. Grundlage einer sicheren Gestaltung ist üblicherweise eine Gefährdungsbeurteilung (wie sie auf operativer Ebene durchgeführt wird). Auf strategischer Ebene sind strukturelle Voraussetzungen für den Arbeitsschutz zu schaffen, durch Einbindung von qualifiziertem Personal, durch Bereitstellung von Prüfmitteln oder durch Unterweisung von Beschäftigten. Die Missachtung einschlägiger Vorschriften kann im Einzelfall dazu führen, dass KI-Anwendungen stillgelegt werden müssen. Die Datenethikkommission der Bundesregierung (DEK) hat eine sog. »Kritikalitätspyramide« entwickelt, mit der sich KI-basierte Entscheidungssysteme bewerten und die Wahrscheinlichkeit und Schwere von Schädigungen berücksichtigen lassen. Sie definiert Anforderungen und Kontrollverfahren und bietet Orientierung bei der Einschätzung der Rahmenbedingungen von KI-Anwendungen. In Zukunft wird insbesondere auch die »Positivliste« sogenannter »Hochrisiko-KI-Systeme« der zu erwartenden KI-Verordnung der EU zu berücksichtigen sein.Leitfragen
- Welche Sicherheitsrisiken sind beim KI-Einsatz grundsätzlich zu berücksichtigen?
- Werden einschlägige Gesetze hinsichtlich Produkt- und Betriebssicherheit des KI-Systems erfüllt?
- Wie ist der betriebliche Arbeitsschutz beim Einsatz von KI-Systemen zu organisieren?
- Muss die Interessenvertretung der Mitarbeitenden wegen Fragen des Gesundheitsschutzes beteiligt werden?
Maßnahmen und Werkzeuge
- Identifizierung der anwendbaren Gesetze und Vorschriften. Insbesondere Prüfung, ob die einzusetzenden KI-Systeme den gesetzlichen Anforderungen an die Produktsicherheit genügen
- Prüfung, ob der Einsatz von KI-Systemen im konkreten Anwendungsbereich von bestehenden Versicherungen abgedeckt ist oder ob zusätzlicher Versicherungsschutz erforderlich ist
- Verfahren zur Gefährdungsbeurteilung (Informationen vermitteln z. B. die Leitfäden der BAuA).
Personenbezogene und sonstige sensible Daten schützen
Das Kernprinzip aller KI-Systeme ist es, Daten massenhaft zu erfassen und diese zu analysieren. Die Erfassung, Speicherung und Auswertung dieser Daten werden immer effizienter. Im Arbeitskontext lassen sich Datenanalysen u. a. zur Fehler- vermeidung nutzen – aber auch zur Leistungskontrolle und Überwachung. Werden personenbezogene Daten erhoben und analysiert, so ist der Betriebsrat zu beteiligen. Bei der Entwicklung und beim Einsatz von KI sind insbesondere die Gesetze zum Schutz der Persönlichkeitsrechte der Beschäftigten einzuhalten, von der DSGVO bis zum Arbeitsrecht, insbesondere die Mitbestimmungsrechte der Vertretungen der Mitarbeitenden. Personenbezogene Daten Dritter, z. B. von Kund*innen, sind ebenfalls nur unter Berücksichtigung des geltenden Datenschutzrechts zu verarbeiten. Für die Ermittlung der rechtlichen Rahmenbedingungen ist es erforderlich, zwischen den verschiedenen Datenarten und den betroffenen Personen zu differenzieren, deren Daten im konkreten Einzelfall beim Training der KI und im anschließenden Praxisbetrieb verarbeitet werden sollen. Soweit personenbezogene Daten verarbeitet werden, muss dafür eine Rechtsgrundlage vorliegen. Personenbezogen sind alle Informationen, die sich auf eine identifizierte oder identifizierbare natürliche Person beziehen. Umfasst sind etwa der Name, die Anschrift und die Kontaktdaten der Mitarbeitenden oder sonstiger Personen, aber auch Informationen, aus denen sich erst durch Kombination mit weiteren Daten eine Person identifizieren lässt, etwa die IP-Adresse oder eine Beschäftigtennummer. Eine Verarbeitung dieser Daten ist nur zulässig, soweit entweder eine ausdrückliche gesetzliche Rechtsgrundlage besteht oder die Einwilligungen der Betroffenen eingeholt worden sind. Letzteres ist im Beschäftigungsverhältnis häufig kritisch, da wegen des Unter- / Überordnungsverhältnisses zwischen Arbeitgeber und Arbeitnehmer strenge Anforderungen an die Wirksamkeit einer freiwilligen Einwilligung gestellt werden. Für besonders sensible Datensätze mit Personenbezug gelten noch strengere Vorgaben. Im Beschäftigtenkontext fallen hierunter etwa Daten zur ethnischen Herkunft, zur religiösen Überzeugung, zur Gewerkschaftszugehörigkeit, zur Gesundheit sowie genetische und biometrische Daten. Eine Verarbeitung dieser Daten ist untersagt, soweit nicht einer der eng begrenzten Ausnahmefälle der DSGVO eingreift. Daneben gilt es, die allgemeinen Datenschutzvorgaben möglichst frühzeitig in der Projektplanung zu berücksichtigen. Insbesondere müssen Betroffene über die Verarbeitungsvorgänge informiert werden. Soweit Daten in der Vergangenheit für bestimmte Zwecke erhoben wurden (etwa bei der Einstellung eines Mitarbeitenden), muss wegen des Grundsatzes der Zweckbindung geprüft werden, ob diese Daten überhaupt für andere Zwecke (etwa das Training von KI-Lösungen oder die Leistungskontrolle) genutzt werden dürfen. Bei der Verwendung neuer Technologien, wie KI-Lösungen, muss der Verantwortliche nach Datenschutzgrundverordnung vor der Verarbeitung eine Datenschutz-Folgenabschätzung vornehmen. Dies gilt vor allem im Beschäftigtenkontext, da die Rechte der Beschäftigten im Verhältnis zu jenen des Arbeitgebers als besonders schützenswert eingeordnet werden. Sollen KI-Lösungen auch zur Vorbereitung von Entscheidungen genutzt werden, die rechtliche Wirkung entfalten (etwa bei der Personalgewinnung oder -entwicklung), gelten besonders strenge Anforderungen; eine Nutzung von KI-Lösungen im Beschäftigtenkontext ist hier meist unzulässig. Ein besonderes Augenmerk ist zudem auf KI-Systeme zu legen, die sich selbstständig Daten über das Internet beschaffen, z. B. durch das »Crawlen« (»Text- und Data Mining«) durch Datenbanken und Internetseiten sowie durch die Nutzung von Suchmaschinen. Inwieweit dies datenschutzrechtlich möglich ist und ggf. noch weitere Risiken birgt, ist im konkreten Fall zu prüfen und zu dokumentieren. Soweit nur mit Daten gearbeitet wird, die sich keinem individuellen Mitarbeitenden oder sonstigen Personen zuordnen lassen (etwa rein betriebsbezogene oder aggregierte Datensätze), ist keine datenschutzrechtliche Erlaubnis erforderlich. Daher kann es sinnvoll sein, den Einsatz von KI-Lösungen von vorneherein auf aggregierte Daten ohne Personenbezug zu beschränken oder Daten zu anonymisieren, bevor sie in einem KI-System genutzt werden. Soweit nur mit Daten gearbeitet wird, die sich keinem individuellen Mitarbeitenden oder sonstigen Personen zuordnen lassen (etwa rein betriebsbezogene oder aggregierte Datensätze), ist keine datenschutzrechtliche Erlaubnis erforderlich. Darüber hinaus ist die Verarbeitung personenbezogener Daten nicht erforderlich und damit unzulässig, soweit der Verarbeitungszweck auch mit anonymen Datensätzen erreicht werden kann. Folglich ist vorrangig zu prüfen, ob eine Verwendung anonymer Daten ausreicht. Erfolgt ein Rückgriff auf personenbezogene Daten, ist es regelmäßig weniger eingriffsintensiv, diese Daten vor der Nutzung in einem KI-System zu anonymisieren. Auch eine Anonymisierung von Daten kann aus rechtlicher Sicht eine Verarbeitung darstellen, die einer Rechtsgrundlage bedarf. Findet eine Trennung von aggregierten und individualisierten Daten nicht statt, sind individuelle Leistungskontrollen möglich. Daten, die dahingehend rekonstruierbar sind, dass sie wieder personenbeziehbar sind, sind ebenfalls nicht anonymisiert, sondern pseudonymisiert, sodass die Verarbeitung auch weiterhin dem Datenschutzrecht unterfällt. KI, die eine Leistungskontrolle ermöglicht, bedarf einer Vereinbarung mit dem Betriebsrat. Gerade die Methoden der prädiktiven Analytik (d. h. Vorhersage zukünftiger Ereignisse durch Analyse von historischen Daten) bergen Gefahren von illegaler Diskriminierung und Verhaltensbeeinflussung. Verstöße gegen das Allgemeine Gleichbehandlungsgesetz müssen vollständig vermieden werden. Kommt bei der Aufstellung von Auswahlrichtlinien KI zum Einsatz, hat der Betriebsrat zwingend mitzubestimmen.Leitfragen
- Welche Datensätze sollen für die Entwicklung bzw. das Training der KI genutzt werden? Welche Datensätze werden im anschließenden Praxisbetrieb verarbeitet?
- Welche rechtlichen Rahmenbedingungen müssen dabei beachtet werden?
- Ist es für die Zwecke der jeweiligen Anwendung ausreichend, anonyme Daten zu verwenden?
- Besteht eine ausreichende Rechtsgrundlage für die Verarbeitung personenbezogener Daten?
- Wer erhält Zugriff auf die Daten?
- Wie sind betriebliche Datenbestände gegen Missbrauch oder Betriebsspionage geschützt?
- Müssen die Risiken für personenbezogene Daten vorab in einer Datenschutz-Folgenabschätzung geprüft und dokumentiert werden? Wer führt diese durch?
- Wie werden Mitbestimmungsrechte hinsichtlich des Schutzes personenbezogener Daten wahrgenommen?
- Wie werden die Kriterien der Datensparsamkeit und Zweckbindung der Datennutzung umgesetzt?
- Wie werden Mitarbeitende und sonstige Betroffene über Belange des Persönlichkeits- und Datenschutzes informiert?
- Ist die Nutzung von KI-Lösungen kompatibel mit den Verarbeitungszwecken, die an Mitarbeitende kommuniziert wurden?
- Werden durch die KI automatisiert Entscheidungen vorbereitet oder durchgeführt, die rechtliche Wirkung gegenüber Mitarbeitenden oder sonstigen Personen entfalten (etwa im Bereich prädiktive Analytik)?
Ethische Verträglichkeit gewährleisten
Zu den humanen Kriterien gehören auch gesetzliche und ethische Anforderungen (ELSI). ELSI ist ein Akronym für »Ethical, Legal and Social Implications«. Technik besitzt keine moralische Urteilsfähigkeit. Um die menschliche Entscheidungs- und Handlungsautonomie in moralischer Hinsicht zu erhalten, sind ethische Grundsätze bei der Systemgestaltung einzuhalten. Eine Diskussionsgrundlage bieten die Gestaltungsanforderungen von Floridi und Cowls [6]:- Wohltätigkeit: Das Prinzip betont die Bedeutung des Wohlergehens der Menschen, der Wahrung ihrer Würde und des Erhalts der Ökosphäre.
- Schädigungslosigkeit: Dieses Prinzip warnt vor den negativen Folgen der übermäßigen Nutzung oder des Missbrauchs von KI-Systemen. Von besonderer Bedeutung ist es, eine Verletzung der Privatsphäre zu vermeiden und die menschliche Autonomie zu respektieren.
- Autonomie (bzw. Entscheidungskompetenz): Bei der Nutzung von KI-Systemen überlässt der Mensch einen Teil seiner Entscheidungskompetenz einem technischen Artefakt. Es ist zu verhindern, dass KI-Systeme die Entscheidungskompetenz an sich ziehen und mithin die menschliche Autonomie untergraben.
- Gerechtigkeit: Das Prinzip der Gerechtigkeit soll einen gleichberechtigten Zugang zu den Vorteilen von KI-Systemen schaffen, um eine Diskriminierung von gesellschaftlichen Gruppen zu vermeiden. Zudem umfasst es die unverzerrte Nutzung von Datensätzen, die für KI-Systeme verwendet werden.
- Erklärbarkeit: Dieses Kriterium bezieht sich auf die Transparenz und Verständlichkeit von KI-basierten Entscheidungsprozessen im erkenntnistheoretischen Sinn sowie auf eine Rechenschaftspflicht hinsichtlich getroffener Entscheidungen.
- Entscheidungskompetenz: Entscheidungen mit ethischer Relevanz werden vom Menschen und nicht von intelligenten Maschinen getroffen, wodurch Handlungsträgerschaft und Verantwortlichkeit eindeutig geklärt sind.
- Transparenz: Die Funktionsweise ist klar, konsistent und nachvollziehbar; intransparente Entscheidungsfindungen mittels Maschinenlernen werden vermieden.
- Erklärbarkeit: Fähigkeit, die Funktionsweise in verständlicher Sprache zu erklären.
- Überprüfbarkeit: Ermöglicht Dritten, den Dateninput zu bewerten und sicherzustellen, dass Ergebnisse vertrauenswürdig sind.
- Personenbezogene Daten: Der Umfang und die Art der Verwendung personenbezogener Daten sind geklärt; alle Beteiligten haben eingewilligt.
Leitfragen
- Ist für alle Beteiligten klar, welche Auswirkungen die ethische Gestaltung von KI-Anwendungen haben bzw. welche Funktionen sich ggf. nicht realisieren lassen?
- Ist geklärt, wer in welchem Umfang für die Nutzung eines KI-Systems verantwortlich ist?
- Sind mögliche Schäden durch den Abschluss entsprechender Versicherungen abgedeckt?
- Wie sind die Mitsprache- und Mitgestaltungsmöglichkeiten der Mitarbeitenden und deren Interessensvertretungen u. a. auch für Experimentier- und Testphasen geregelt?
Maßnahmen und Werkzeuge
- Festlegen einer ethischen Richtlinie für die KI-Anwendung (in Abstimmung mit Mitarbeitenden bzw. der Beschäftigtenvertretung).
- Erstellung einer Gefährdungsbeurteilung (z. B. nach [7]) für den Einsatz von KI-Systemen in bestimmten Anwendungsfällen.
- Absicherung der Nutzung von KI-Systemen durch Zertifizierungen und Abschluss von Versicherungen.
Referenzdokumente
- Ethische Leitlinien, wie Gutachten der Datenethikkommission für die Bundesregierung zum Einsatz von Algorithmen und Künstlicher Intelligenz,
- Europäische Ethik-Leitlinien für vertrauenswürdige KI,
- »Hambacher Erklärung zur künstlichen Intelligenz«.
Rechtliche Risiken minimieren
Aufgrund des dargestellten Kernprinzips von KI-Systemen, Daten massenhaft zu erfassen und diese zu analysieren, muss der Rechtmäßigkeit der Beschaffung dieser (Trainings-) Daten ein besonderes Augenmerk gewidmet werden.Vorsicht: Trainingsdaten können gesetzlich geschützt sein
Unabhängig davon, ob Trainingsdaten als personenbezogene oder -beziehbare Daten anzusehen sind und dadurch die strengen Anforderungen des Datenschutzrechts auf die Verarbeitung solcher Daten anwendbar sind, können die Trainingsdaten selbst unter mehreren rechtlichen Gesichtspunkten geschützt sein, die ggfs. ein Einverständnis eines Rechteinhabers in die Sammlung und Verarbeitung solcher Daten erforderlich machen. Darüber hinaus können aber auch die Datenbanken, denen die Trainingsdaten entstammen, selbst gegen das Auslesen der (ansonsten nicht selbstständig geschützten) Dateninhalte geschützt sein. Ein Schutz der Trainingsdaten kann sich ergeben, wenn es sich bei den Trainingsdaten um urheberrechtlich geschützte Werke oder designrechtlich geschützte, eingetragene oder nichteingetragene Designs handelt. Als relevante Beispiele lassen sich z. B. folgende Werkarten identifizieren:- Literarische Werke: Romane, Gedichte, Theaterstücke, Essays, Artikel usw.
- Musikalische Werke: Kompositionen, Songs, Liedtexte, einzelne Klangsequenzen
- Bildende Kunst: Gemälde, Skulpturen, Zeichnungen, Fotografien, Grafiken usw.
- Architektonische Werke: Gebäude, Bauwerke, Skizzen, Pläne
- Filmwerke: Filme, Animationen usw.
- Computerprogramme: Softwarecode, Applikationen, Algorithmen
- Äußere Gestaltungen von Bedarfsgegenständen aller Art
- Stoffmuster und Mustergestaltungen aller Art
Datenbanken auslesen?
Neben dem bzw. unabhängig vom (inhaltlichen) Schutz der Trainingsdaten kann zugleich die Datenbank, der Trainingsdaten entstammen, selbst urheberrechtlichen Schutz genießen. So können z. B. Gedichtelisten oder medizinische Lexika sowie Websites mit umfangreich geordnetem Inhalt und einer Suchfunktion sowie Datenbanken allgemein, bei denen die Beschaffung, Überprüfung oder Darstellung eine nach Art oder Umfang wesentliche Investition erfordert, unter urheberrechtlichen bzw. urheberrechtsähnlichen Schutz stehen. Besteht ein urheberrechtlicher Schutz der Datenbank als solcher, kommt es auf die Schutzfähigkeit des Datenbankinhalts (also der einzelnen Datensätze, die selbst nicht gesetzlich schutzfähig sein müssen) nicht mehr an. Sind die Trainingsdaten oder die Datenbanken, aus denen die Trainingsdaten stammen, urheberrechtlich geschützt, richtet sich die Zulässigkeit der Entnahme von Daten und ggfs. der Datenbankstruktur nach den urheberrechtlichen Ausnahmebestimmungen für das sogenannte Text- und Data Mining. Dieses, auch kommerziellen Unternehmen zustehende Recht, schränkt die Rechte des Urhebers zugunsten des KI-Nutzers deutlich ein. Zu beachten ist jedoch, dass Urheber ihre veröffentlichten Werke von der Zulässigkeit des Text- und Data Mining ausnehmen können, wenn sie sich ihre Rechte „in maschinenlesbarer Form” – also auch nur eine niederschwellige Information in einem Website-Impressum oder in AGB – vorbehalten haben.Trainingsdaten kaufen?
Zwischenzeitlich existieren auf dem Markt viele Angebote von Trainingsdatensätzen. Kommt die Beschaffung solcher Trainingsdatensätze in Betracht, ist ein Augenmerk auf die Vertragsgestaltung und -verhandlung zu setzen. Die Eigenschaften der zu beschaffenden Datensätze sollten im Vertrag so konkret wie möglich beschrieben sein. Außerdem sollte der Anbieter eine verbindliche Aussage zur rechtlichen Schutzfähigkeit und insbesondere zu Rechtesituation treffen können und sich auch vertraglich zu einer bestimmten Beschaffenheit verpflichten. Hierfür sind insbesondere Haftungsfreistellungen erforderlich, die allein den Anbieter der Trainingsdatensätze für Schadenersatzansprüche Dritter haften lassen, wenn die Nutzung seiner Trainingsdaten Rechte solcher Dritter verletzt.Leitfragen
- Ist den für die Beschaffung von Trainingsdaten zuständigen Mitarbeitenden bewusst, dass die Rechtesituation bezüglich aller zu beschaffender Trainingsdaten eindeutig geklärt werden muss, bevor die Trainingsdaten beschafft werden?
- Lassen sich die zu beschaffenden Trainingsdaten in eine der gesetzlichen Schutzkategorien einteilen: Sind es personenbezogene oder -beziehbare Daten, können die Trainingsdaten urheberrechtlichen Schutz genießen, weil es sich um Bilder, Filme, Texte etc. handelt?
- Welche Personen (z. B. eigene Rechtsabteilung, Beauftragung externen Rechtsrats) sind einzubeziehen, wenn geprüft wird, welche Schutzrechte für die zu beschaffenden Daten gelten und wie sie verwendet werden dürfen?
Handlungsfeld: Soziale Rahmenbedingungen
Die folgenden Empfehlungen sind bei der Nutzung von KI unabdingbar, um angemessene Rahmenbedingungen für einen KI-Einsatz zu schaffen. Ohne diese Rahmenbedingungen kann der KI-Einsatz scheitern. Zumindest bleiben die erwünschten Potenziale ungenutzt. Den Vorteilen der hier empfohlenen Ausprägungen der Organisationskultur stehen stets Aufwände und Kosten gegenüber. Daher ist gemeinschaftlich anhand der KI-Strategie zu entscheiden, welche Kulturaspekte zu fördern sind.KI-förderliche Organisationskultur
Was ist Organisationskultur?
Die Organisationskultur prägt die Wahrnehmung, das Denken, die Empfindungen und letztlich das Handeln der Mitglieder einer Organisation. Die Organisationskultur lässt sich auf drei Ebenen betrachten [8]:- »Artefakte« sind die sichtbaren Auswirkungen der Organisationskultur. Diese umfassen z. B. Dresscode und Umgangsformen, wie Arbeit organisiert wird, welche Formen des Austauschs es gibt, worüber man sich austauscht, und welche Rollen und Prozesse es für verschiedene Aufgaben in der Organisation gibt.
- »Werte und Normen« bilden die implizite Basis für die Artefakte. Das Artefakt »höhenverstellbare Bürotische« spiegelt z. B. die Norm wieder, dass die Gesundheit der Mitarbeitenden zu schützen ist.
- »Grundannahmen« bilden wiederum die Grundlage der Werte und Normen. Sie beschreiben implizit, wie die Welt und die Menschen sind. Eine Grundannahme kann z. B. sein: »Menschen sind grundsätzlich motiviert, Ziele zu erreichen.« Daraus ergibt sich die Norm, Menschen selbstbestimmt Ihre Arbeit einteilen zu lassen, was sich wiederum im Artefakt einer flexiblen Arbeitszeitregelung darstellt.
Generelle Maßnahmen des Kulturwandels
Die Mitglieder der Organisation sollten immer gemeinschaftlich beschließen, bestimmte Werte zu verankern und die entsprechende Kultur zu leben. Besonders wichtig ist, dass die Unterstützung durch die Unternehmensführung sichtbar wird, indem das Verhalten zu den beschlossenen Werten passt. Das schließt ein, dass dann die nötigen Strukturen geschaffen und Ressourcen freigegeben werden.- Festlegung auf gemeinsame Zielwerte, z. B. in Workshops oder durch Befragungen
- Moderierter Transformationsprozess
- Bereitstellung von Ressourcen durch die Führungsebene, um Artefakte (z. B. Fortbildungsprogramme) zu etablieren
- Vorleben der Werte durch die Führungsebene (z. B. beim Treffen strategischer Entscheidungen)
- Regelmäßige Evaluation der Fortschritte
Generelle Maßnahmen und Werkzeuge
- »My perfect day«: Im Rahmen eines moderierten Workshops beschreiben z. B. 5 bis 8 Mitarbeitende einen »perfekten« Tagesablauf. Aus den Beschreibungen wird ermittelt, welche Grundannahmen und Werte vorherrschen und welche man etablieren müsste, damit es besser wäre. Daraus ergibt sich die Roadmap für den Kulturwandel.
- Verwendung des KI-ULTRA Evaluation Toolkit zur Beurteilung der vorherrschenden Organisationskultur.
- Die Werte der jeweiligen Kulturdimension werden durch die Unternehmensführung vorgelebt.
Im Vergleich zur Fachliteratur wählen wir hier eine einfache und pragmatische Darstellung der für den KI-Einsatz relevanten Aspekte (vgl. Abbildung 3). Diese Kulturdimensionen können miteinander in Einklang gebracht werden und sich sogar teilweise gegenseitig bedingen oder fördern.
Etablieren einer humanistischen Organisationskultur
Eine humanistische Organisationskultur stellt das Wohl der Menschen in den Mittelpunkt von Entscheidungen. Sie ist menschenzentriert und chancengleichheitsorientiert. Menschenzentriert bedeutet, dass handlungsleitend ist, inwiefern sich Entscheidungen auf Mitarbeitende, die Kund*innen und das gesellschaftliche Umfeld auswirken. Dieser Fokus auf die Menschen erleichtert es, im KI-Entwicklungs- und Einführungsprozess die relevanten Stakeholder, insbesondere Mitarbeitende, einzubeziehen und dadurch Bedarfe und Lösungen frühzeitig zu entdecken. Fehlt dieser Fokus, besteht das Risiko, dass berechtigte Interessen bei der KI-Einführung unberücksichtigt bleiben. Darunter leidet das Zutrauen in den Change-Prozess. In der Folge können Vorbehalte gegen die KI-Lösungen entstehen. Zudem ist eine humanistische Organisationskultur auf Chancengleichheit ausgerichtet, d. h., sie räumt allen Menschen dieselben Rechte ein. Mitarbeitende unterschiedlicher Herkunft, aller Geschlechter und mit oder ohne Behinderungen werden durch Maßnahmen wie den Einsatz von KI nicht strukturell benachteiligt. Eine solche, partizipative Organisationskultur erleichtert die Umsetzung ethischer Richtlinien, welche die Akzeptanz der KI-Anwendung fördern. Eine diskriminierungsfreie und inklusive Umsetzung der KI-Anwendung hat Vorteile für die Organisation, die Mitarbeitenden und die Gesellschaft. Darüber hinaus sind bezüglich der Diskriminierungsfreiheit sowohl in Bezug auf die konkrete Umsetzung der KI als auch in Bezug auf andere Abläufe in der Organisation (z. B. Einstellungsverfahren) rechtliche Rahmenbedingungen einzuhalten. Diversität in der Organisation kann die Innovationsfähigkeit stärken.Maßnahmen und Werkzeuge
- Förderung der Kommunikation zwischen den Mitgliedern der Organisation
- Bei relevanten Entscheidungen werden Betroffene aktiv einbezogen und Bedenken ernst genommen.
- Am Arbeitsplatz wird die psychische und körperliche Gesundheit der Mitarbeitenden berücksichtigt.
- Die Mitarbeitendenvertretung wird bei der Planung von Arbeitsverfahren und Arbeitsabläufen unter Einsatz von KI beteiligt.
- Definition von Zielen der Organisation, die gesellschaftliche Werte wie Nachhaltigkeit, soziale Gerechtigkeit o. Ä. berücksichtigen.
- Etablieren einer beauftragten Person für Chancengleichheit / Diversity
- Kontinuierliche Einbindung der Mitarbeitendenvertretung (z. B. Betriebs- / Personalrat).
- Anonymisierte Einstellungsverfahren
- Beteiligung des Betriebsrats, wenn KI zur Aufstellung von Richtlinien zur personellen Auswahl bei Einstellungen, Versetzungen, Umgruppierungen oder Kündigungen eingesetzt wird.
- Anbieten von Diversity-Trainings
- Wissenstransfer institutionalisieren (z. B. über Mentoring-Programme)
Etablieren einer Fehlerkultur
Eine Fehlerkultur zeichnet sich dadurch aus, dass es allgemein akzeptiert oder gewünscht ist, dass Mitarbeitende wohlüberlegte Risiken eingehen, um Ziele der Organisation zu erreichen (z. B. KI-Entwicklung, auch wenn der Erfolg des Projekts nicht garantiert ist). Irrtümer werden nicht bestraft, sondern offen zugegeben und als Lernmöglichkeit genutzt. Eine Fehlerkultur ist erforderlich, damit eine Innovationskultur zum Tragen kommen kann; nur wenn Probleme offen eingestanden und konstruktiv bearbeitet werden, können innovative Lösungen entstehen.Maßnahmen und Werkzeuge
- Einführung von Formaten zum konstruktiven Umgang mit Fehlern, z. B. sog. »fuck-up-nights«
- Beteiligung des Betriebsrats zur Aufstellung von Regeln zum Umgang mit Fehlern
- Besprechungen, in denen sich Mitarbeitende über gescheiterte Projekte und die Ursachen des Scheiterns austauschen können
- Bei Bedarf Weiterbildungen anbieten, falls dadurch die Wiederholung von festgestellten Fehlern vermieden werden kann
Etablieren einer Vertrauens-, Kooperations-, und Innovationskultur
Diese drei Dimensionen der Organisationskultur bedingen sich gegenseitig und können kaum voneinander unabhängig etabliert werden. Die Vertrauenskultur und die Kooperations- bzw. Zusammenarbeitskultur sind Grundlagen für eine wirksame Innovationskultur. Eine Vertrauenskultur zeichnet sich durch die Grundannahme aus, dass Mitarbeitende generell dazu motiviert sind, im Sinne der Organisation und deren Mitglieder zu handeln. Wesentliche Informationen werden geteilt, statt sie zurückzuhalten. Mitarbeitende dürfen sich weitgehend selbst organisieren, um die Ziele der Organisation zu erreichen. Das Teilen von Informationen ist eine Voraussetzung für die erfolgreiche Zusammenarbeit in Teams und damit für eine Kooperationskultur. Diese zeichnet sich durch die Grundannahme aus, dass eine starke Kohäsion im Team zu besseren Arbeitsergebnissen führt, als wenn jeder für sich arbeitet. Diese Zusammenarbeitskultur muss nicht in absoluter Form umgesetzt werden. In vielen Organisationen wird erfolgreich das Konzept der »Coopetition« eingesetzt. Hier gibt es sowohl Anreize zur Zusammenarbeit als auch zum Wettbewerb und damit als Leistungsanreiz innerhalb der Organisation (z. B. durch die Organisation in Cost Centern). Gängige Konzepte einer Innovationskultur umfassen Teile der anderen hier genannten Kulturdimensionen (z. B. Vertrauen, Zusammenarbeit, Wandel). Sie betont den Aspekt, dass Innovation im Unternehmen gefördert wird. Innovation beschreibt hier den Prozess, durch den Organisationen Ideen in neuartige Produkte, Dienstleistungen oder Prozesse verwandeln, um dadurch einen Vorsprung im eigenen Markt zu erzielen. Innovatives Denken und Handeln kann nicht angeordnet werden, aber man kann förderliche Rahmenbedingungen herstellen. Eine Innovationskultur fördert das Auffinden neuer Ideen und deren Umsetzung in etwas, das wirklich existiert und funktioniert. Sie basiert auf der Grundannahme, dass Innovationen wichtig für den Erfolg der Organisation sind und auf Werten und Normen, die es erlauben, kreative Lösungen zu finden und diese alltagstauglich zu machen. Gemeinsam machen diese drei Kulturdimensionen die Potenziale von KI nutzbar: KI-Anwendungen können Routinetätigkeiten automatisieren und damit Kapazitäten der Mitarbeitenden freisetzen, die für anspruchsvolle Tätigkeiten eingesetzt werden können. Eine Vertrauenskultur sichert Autonomie zu und schafft dabei Freiraum für kreative Ideen. Die Kooperationskultur hilft dabei, diese kreativen Ideen umzusetzen, denn Innovationen entstehen v. a. in Teams. Eine Zusammenarbeitskultur ist auch ein wichtiger Katalysator für das organisationale Lernen: Die Mitarbeitenden lernen nicht nur für sich neue Fertigkeiten, sondern ganze Teamstrukturen passen sich an neue Anforderungen an. Dieses organisationale Lernen ist ebenfalls bedeutsam für die Innovationsfähigkeit von Organisationen. Eine Innovationskultur hilft dabei, in einem Kreativprozess neue Anwendungsfälle zu identifizieren, die nötigen Risiken einzugehen und einen agilen Entwicklungsprozess hin zur funktionierenden Lösung durchzuführen. Das schließt auch die Neuordnung existierender Prozesse ein.Maßnahmen und Werkzeuge
- Flexible und agile Arbeitsmodelle
- Übertragung von Entscheidungskompetenzen auf Mitarbeitende oder selbstorganisierte Teams, z. B. bei der Aufgabenauswahl
- Beteiligung des Betriebsrats bei der Arbeitszeitgestaltung oder der Aufstellung von Ethikregeln
- Rolle der Führungskraft als Unterstützer, damit Mitarbeitende Ihre Potenziale entfalten können
- Einsatz von Teamzielen statt oder in Ergänzung zu individuellen Zielvorgaben
- Schaffung von Anreizen durch Betriebsvereinbarungen zur Lohngestaltung
- Übertragung von Entscheidungskompetenzen auf Mitarbeitende oder selbstorganisierte Teams, z. B. bei der Aufgabenauswahl
- Führungskraft als Unterstützer*in und Mediator*in, um den Teamzusammenhalt zu stärken
- Aufsetzen eines Prozesses, um Ideen für innovative Neuerungen zu bewerten und die vielversprechendsten für ein Innovationsprojekt auszuwählen
- Bereitstellung von Finanzierung und Personalkapazität für die Umsetzung neuer Ideen
- Vernetzung mit Kund*innen, Projektpartnern oder auch Konkurrent*innen, um gemeinsam Trends und Entwicklungen vorherzusehen und aktiv mitzugestalten
Etablieren einer Kultur des Wandels
Eine Kultur des Wandels basiert auf der Grundannahme, dass kontinuierliche Veränderung etwas Positives ist und der Organisation mehr nützt als schadet, wenn es nicht sogar für diese überlebensnotwendig ist. Diese Annahme ist nicht in allen Organisationen gegeben. Schließlich gibt es auch in jedem Unternehmen (scheinbar) bewährte Erfolgsrezepte und Vorgehensweisen, und Wandel kostet immer Zeit und Geld. Neuerungen im Unternehmen können oftmals nur erfolgreich umgesetzt werden, wenn die Belegschaft diese aktiv unterstützt, sich z. B. in neue Software einarbeitet, sich mit neuen Prozessen vertraut macht und sich weiterbildet.Maßnahmen und Werkzeuge
- Regelmäßige Termine für den Austausch zu Neuerungen, an denen Beschäftigte Lob und Kritik offen und ohne negative Konsequenzen äußern können.
- Einführung eines betrieblichen Vorschlagswesens unter Beachtung des Mitbestimmungsrechts des Betriebsrats.
- Individuelle Weiterbildung ermöglichen, damit sich Beschäftigte eigenständig auf den Wandel an Ihrem Arbeitsplatz vorbereiten und diesen aktiv mitgestalten können.
Etablieren einer Datenkultur
Eine Datenkultur fußt auf der Grundannahme, dass Daten für die Organisation sehr wertvoll sind und dass der Erfolg der Organisation davon abhängt, wie gut diese Daten genutzt werden. Entscheidungen werden auf der Grundlage von Daten getroffen. Die Mitarbeitenden fühlen sich für die Qualität und die Verfügbarkeit ihrer Daten verantwortlich und machen diese Daten innerhalb der Organisation frei zugänglich, soweit dies datenschutzrechtlich zulässig ist. Zudem bilden sich die Mitarbeitenden weiter, um für die Arbeit mit Daten gerüstet zu sein. Diese Weiterbildung wird durch die Organisation gefördert.Maßnahmen und Werkzeuge
- Anbieten von Weiterbildungsmaßnahmen zu statistischen Grundlagen für das datengetriebene Entscheiden, zu den Potenzialen und Grenzen von KI, zum Umgang mit den Ergebnissen von KI-Anwendungen, sowie zu neuen Werkzeugen, z. B. Software zur Datenanalyse
- Bereitstellung von Datenschutzrichtlinien und Datenschutzprozessen, einschließlich Rollen- und Rechtekonzepten für den Datenzugriff
- Bereitstellung von Infrastruktur und Software zur Datenhaltung und -analyse
Empowerment-orientierte Führung
Neben den o. g. Kulturdimensionen hat auch das Führungsverhalten, in der Regel durch Linienvorgesetzte oder Projektleiter*innen ausgeübt, einen Einfluss darauf, ob sich die durch den KI-Einsatz freigelegten Potenziale der Mitarbeitenden entfalten können oder ob der Effekt der KI-Einführung in diesem Bereich verpufft. Insbesondere wenn die Organisationskultur auf Vertrauen und Zusammenarbeit ausgelegt ist, dann sollten die praktizierten Führungsstile damit in Einklang sein. Wir empfehlen daher, dass in den für den KI-Einsatz relevanten Bereichen ein Empowerment- orientierter Führungsstil praktiziert wird. Empowerment meint die Ermächtigung und Selbstbefähigung der Mitarbeitenden, eigenständig zu handeln und zu entscheiden. Dies ist aus zwei wesentlichen Gründen von Vorteil:- KI-Einführungsprojekte zeichnen sich, selbst wenn das Unternehmen gezielt diejenigen Anwendungsfälle mit niedriger Komplexität und hoher Ertragsaussicht priorisiert, meist durch eine hohe Unbestimmtheit und Komplexität aus. Diese Komplexität betrifft nicht nur die KI-Anwendung selbst, sondern auch die mit ihrer Nutzung verbundenen Arbeits- oder Geschäftsprozesse. Damit die Projekte effizient und erfolgreich durchgeführt werden können, müssen Fachkräfte mit Spezialwissen aus unterschiedlichen Domänen zusammenarbeiten. Diese sollten dabei nicht von starren und detaillierten Anweisungen eingeschränkt werden, sondern auf Basis ihrer Kompetenzen selbstständig zum Wohle des Projekts handeln können.
- KI-Anwendungen bieten das Potenzial, Arbeit derart neu zu organisieren, dass Menschen und Maschinen jeweils das tun, was sie am besten können (siehe: Fähigkeiten von Mensch und Maschine einbeziehen). Es bietet sich also an, standardisierbare, wiederkehrende Routinetätigkeiten oder Berechnungen, die für das menschliche Auffassungsvermögen zu komplex sind, an die KI-Anwendung zu delegieren. Auf der anderen Seite bleibt für die Mitarbeitenden ein Freiraum, um das zu tun, was sie am besten können: Entscheidungen auf Erfahrungsbasis oder unter Berücksichtigung ethischer Grundlagen zu treffen, kreativ zu sein, mit Menschen auf persönlicher Ebene zu interagieren und sie zu beraten oder für etwas zu begeistern. Das erfordert eine angemessene Autonomie.
- Vorbildfunktion: Die Führungsperson agiert als Vorbild für ethisches Verhalten und gewinnt dadurch Respekt, Vertrauen oder Bewunderung.
- Inspirierende Motivation: Die Führungsperson motiviert die Mitarbeitenden durch ein bedeutungsvolles, positives Zielbild der Zukunft, auf das man gemeinsam hinarbeitet.
- Intellektuelle Anregung: Die Führungsperson ermutigt die Mitarbeitenden, eigene Lösungen für Probleme zu finden.
- Individualisierte Unterstützung: Die Führungsperson agiert als Coach oder Mentor, um die Mitarbeitenden bei ihrer persönlichen Entwicklung zu unterstützen. Individuelle Bedarfe (z. B. bzgl. Arbeitszeitflexibilität) der Mitarbeitenden werden erkannt und berücksichtigt.
Maßnahmen und Werkzeuge
- Diagnose: Das Ausmaß an Empowerment-orientierter Führung kann kostenfrei mit dem KI-ULTRA Evaluation Toolkit gemessen werden.
- Diagnose: Die Autonomie von Mitarbeitenden kann mit der deutschen Version des Job Diagnostic Survey gemessen werden (erhältlich bei der BAuA – Bundesanstalt für Arbeitsschutz und Arbeitsmedizin).
- Intervention: Auswahl von Führungskräften, die sich durch einen Empowerment-orientierten Führungsstil auszeichnen
- Intervention: Basierend auf Diagnosedaten sollten Führungskräfte präventiv geschult werden, damit sie Empowerment-orientiert vorzugehen wissen.
- Sofern vorhanden: Empowerment-orientierte Führung im Schulungs- oder Wertekanon aufnehmen.
Handlungsfeld: Kompetenzentwicklung und Lernen
Die Mitarbeitenden sind auch im digitalisierten Arbeitssystem unersetzbar: Sie initiieren, trainieren und lenken technische Systeme, sie mappen Datenbestände oder unterstützen technische Funktionen, zu deren Ausführung die Technik nicht in der Lage ist. Der Einsatz innovativer Technologien erfordert es, Mitarbeitende für neue Aufgaben zu qualifizieren und ggf. auch neue Mitarbeitende zu rekrutieren. Hinzu kommen erweiterte Karriere- und Entwicklungspfade als Grundlage persönlichen Wachstums und betrieblicher Flexibilität. Der Qualifikationsbedarf hängt dabei nicht nur von einem konkreten Anwendungsfall ab, sondern bildet die übergeordnete KI-Strategie der Organisation ab. Organisationales Lernen ergänzt den Prozess der Organisationsentwicklung.Strategisches Kompetenzmanagement
Damit KI-Einführungsprojekte gelingen, müssen einschlägige Kompetenzen der Mitarbeitenden systematisch entwickelt und rechtzeitig verfügbar gemacht werden. Im Vergleich zu anderen Digitalisierungsvorhaben können sich durch den KI-Einsatz Arbeitsweisen grundlegend ändern oder spezielle Kompetenzen, z. B. für den Umgang mit probabilistischen Systemausgaben, erforderlich werden. Hier hilft ein systematisches Kompetenzmanagement, das an der KI-Strategie des Unternehmens ausgerichtet ist. Ein Beispiel einer systematischen Vorgehensweise für ein strategieverbundenes Kompetenzmanagement bietet der Fraunhofer-Kompetenz-Kompass [9]. Dieser wurde in Anlehnung an die DIN PAS 1093:2009-07 entwickelt und umfasst fünf Schritte:- Strategieanbindung
- Kompetenzmodell
- Kompetenzmessung
- Kompetenzaufbau
- Kompetenzbilanz
Leitfragen
- Wie ändern sich die Anforderungen an die Kompetenzen der Mitarbeitenden, wenn die KI-Strategie wie geplant umgesetzt wird?
- Verändert der Einsatz von KI die Arbeitsverfahren und Arbeitsabläufe, sodass der Betriebsrat bei der Planung zu beteiligen ist?
- Wann werden die Kompetenzen benötigt und wie lange dauert es voraussichtlich, diese in der Organisation verfügbar zu machen (durch Weiterbildung oder Neueinstellung)?
- Ist bekannt, wie die vorhandenen (IST) und benötigten (SOLL) Kompetenzen gemessen werden können und wie der gesamte Kompetenzmanagementprozess evaluiert werden kann?
Maßnahmen und Werkzeuge
- Fraunhofer Kompetenz-Kompass »Vorgehensweise zur Kompetenzbedarfsermittlung in Organisationen« [9] mit Leitfragen und Instrumenten für ein strategisches Kompetenzmanagement.
Lernfähigkeit der Mitarbeitenden im Umgang mit KI fördern
Der KI-Einsatz soll die Agilität und Resilienz der Wertschöpfungsprozesse erhöhen. Agilität und Resilienz sind Ergebnisse der engagiert eingesetzten Fähigkeiten der arbeitenden Menschen. Die Befähigung zu motiviertem und selbstorganisiertem Arbeiten kann sich nur in Wechselwirkung mit den Arbeitsprozessen ergeben, in denen die Fähigkeiten benötigt werden. Ausschlaggebend sind die Befähigungen zur Selbstmotivation (als Zielsetzen), zum effizienten Zielverfolgen und zur Selbstbewertung beim Arbeiten. Selbstmotiviertes Lernen und Arbeiten sind nicht abstrakt zu vermitteln. Das Motivierungspotenzial der Arbeitstätigkeiten wird durch kompetenzförderliche Arbeitsgestaltungsmaßnahmen geschaffen. Dies fördert intrinsische Motivation, verfestigt Einstellungen und führt zu geeigneten Bewältigungsstrategien. Selbstmotivierung bedarf konkreter, motivations- und lernbegünstigender Arbeitsinhalte. Die Förderung der Lernfähigkeit ist eine Voraussetzung für die Mitarbeiterentwicklung in digitalisierten Arbeitsprozessen. Sie ist jedoch keine Ad-hoc-Lösung, sondern ein Bestandteil der Personalentwicklung und der Arbeitsgestaltung.Leitfragen
- Welche Lernanforderungen lassen sich in den KI-gestützten Arbeitsprozessen realisieren, in Abhängigkeit von Tätigkeitsanforderungen und individuellen Leistungsvoraussetzungen?
- Wie kann durch einen systematischen Tätigkeitswechsel die Lernfähigkeit erhalten werden?
- Welche Auswirkungen hat dies auf den Personaleinsatz und die Karriere- und Entwicklungspfade? Stellen die Tätigkeitswechsel mitbestimmungspflichtige Versetzungen dar?
- Wie lässt sich Lernfähigkeit fördern, etwa durch praxisnahes Lernen im Umgang mit digitalen Komponenten oder durch Umschulung?
- Wie können betriebliche Bildungsmaßnahmen unter Beteiligung des Betriebsrats entwickelt und durchgeführt werden?
- Wie kann durch partizipative Arbeitsweisen, z. B. Vergabe von Organisationsaufgaben an Betroffene zur Bearbeitung in Qualitätszirkeln, die Lernfähigkeit und -bereitschaft erhalten werden?
- Lassen sich Lerntandems bilden aus Beschäftigten, die digitale Kompetenz beisteuern, und aus Fachleuten, die sich im Austausch mit technologischem Know-how revanchieren können im Sinne eines Erfahrungsaustauschs?
Organisationale Lernfähigkeit fördern
Damit sich Organisationen und Individuen zügig auf neu erkannte Bedarfe einstellen und Kapazitäten verlagern können, empfiehlt sich ein hoher Grad an organisationaler Lernfähigkeit. Das organisationale Lernen ist eine Antwort auf die Herausforderungen des betrieblichen Wandels, wie Volatilität, Unsicherheit, Komplexität und Mehrdeutigkeit, indem Freiräume geschaffen werden und von einseitiger betrieblicher Effizienzoptimierung (ohne angemessenes Referenzsystem für Innovation und Wandel) Abstand genommen wird. Organisationales Lernen bedeutet dauerhafte und absichtsvoll herbeigeführte Verhaltensänderungen in der Organisation, die durch veränderte Prozesse oder Strukturen unterstützt werden. Eine Grundlage der organisationalen Lernfähigkeit ist die kontinuierliche Kommunikation über Veränderungen.Leitfragen
- Wie lassen sich Veränderungen im Umfeld der Organisation prognostizieren, etwa durch einen KI-Einsatz?
- Wie lässt sich eine offene Kommunikation über betriebliche Entwicklungen institutionalisieren und technisch unterstützen?
Handlungsfeld: Datenstrategie und Technologieeinsatz
Genauigkeit der KI Die Art der eingesetzten KI-Methode, die verfügbaren Daten und die Aufgabenstellung können die Genauigkeit der Ergebnisse bestimmen. Die Ergebnisgüte wiederum beeinflusst die Investitionsrendite eines Projekts. Nur wenn Sie verstehen, welche Faktoren sich auf die Genauigkeit eines KI-Projekts auswirken, können Sie angemessene Erwartungen an seinen Erfolg formulieren. So sind beispielsweise Bilderkennungsanwendungen in der Regel zuverlässiger als Prognoseanwendungen. Nur weil eine KI für eine Aufgabe gut funktioniert, heißt das noch nicht, dass sie auch für andere Aufgaben geeignet ist. In der Regel lässt sich ein KI-Modell nicht ohne Weiteres auf andere Bereiche übertragen. Das heißt: KI-Projekte müssen sich immer an konkreten Prozessen und den hierbei verfügbaren Datenbeständen orientieren.Datenstrategie etablieren
Daten sind die Grundlage für KI-Projekte. Während für manche Anwendungsfälle die notwendigen Daten vorliegen mögen oder während der Projektlaufzeit erhoben werden können, so existieren viele Anwendungsfälle, die eine langfristige Datenhistorie benötigen. Hier stellt sich die Frage, wie in der Organisation mit der mittel- und langfristigen Erfassung von Daten umgegangen wird. Genauso, wie sich Prozesse verändern, können sich auch Datenbestände kurzfristig ändern, die Sie für den Betrieb eines KI-Modells brauchen. Immer dann, wenn aktuellere Daten auftauchen, hat das Einfluss auf die Effektivität und Genauigkeit einer KI-Anwendung. Mit zunehmender Menge von verarbeiteten Daten und entsprechender Anwendungssoftware gewinnt die Frage nach einer einheitlichen Strategie zur Datenverwaltung an Bedeutung. Wachsende Systeme resultieren üblicherweise in hohen technischen und organisatorischen Aufwänden zur Datenintegration und -pflege, was mittel- und langfristig dazu führt, dass zu hohe Ressourcen aufgewendet werden müssen. Es stellen sich daher die Fragen nach Datendurchgängigkeit sowie geplanten Strukturen. Große Datenlager wie »Data Warehouses« (d. h. Daten werden im großen Stil bereits vorverarbeitet für ihre Anwendungszwecke abgelegt) oder »Data Lakes« (d. h. Daten werden im großen Stil einheitlich in Rohform abgelegt) stehen vielen einzelnen Datenbanken gegenüber. Neben der rein technischen Ablage der Daten sowie dem Zugriff auf die Daten gibt es auch organisatorische Herausforderungen: die Umsetzung angemessener Rollen- und Rechtekonzepte, Vermeidung von Redundanzen, Einhaltung von Datenschutzanforderungen sowie die Dokumentation (insbesondere von der Verarbeitung personenbezogener Daten). Zu einer erfolgreichen Datenstrategie gehört auch die Berücksichtigung rechtlicher Aspekte. Zum einen gehört zur Datenstrategie das Wissen darum, welche Daten für welche Zwecke genutzt werden dürfen und zum anderen auch, inwieweit Daten, die für die Nutzung eines KI-Systems eingesetzt werden sollen, eines besonderen Schutzes bedürfen. Grundsätzlich muss sichergestellt sein, dass die Daten rechtskonform nutzbar sind. Bei personenbezogenen Daten setzt dies u. a. voraus, dass für die Datennutzung entsprechende Rechtsgrundlagen bestehen und Betroffene über die Verarbeitungszwecke umfassend informiert werden. Soweit Daten in der Vergangenheit für andere Zweck erhoben wurden (etwa die Stammdaten eines Mitarbeitenden zur Abwicklung des Beschäftigtenverhältnisses), können die Verarbeitungszwecke nachträglich nur unter bestimmten Voraussetzungen geändert werden (Grundsatz der Zweckbindung). Bei der Entwicklung einer Datenstrategie sollte daher möglichst frühzeitig der Datenschutzbeauftragte eingebunden werden. Abgesehen von datenschutzrechtlichen Aspekten ist die Nutzung von Daten dem Grunde nach frei möglich. Dies gilt sowohl für selbst generierte als auch für von Dritten beschaffte Daten, da es kein klassisches Eigentum an Daten gibt und ein Urheberrecht (oder urheberrechtsähnliches Recht) an reinen Daten wie z. B. Messdaten oder statistischen Angaben nur in wenigen Fällen zum Tragen kommt; anders kann dies zu bewerten sein, wenn Daten aus Texten oder Bildern bestehen (siehe: Rechtliche Risiken minimieren). Wenn Daten aus öffentlich verfügbaren Quellen oder aber von Datenanbietern beschafft werden, sind aber ggf. Regelungen in Datenlizenzen zu beachten, die einer Nutzung der entsprechenden Datensätze in KI-Systemen entgegenstehen könnten oder aber zusätzliche Lizenzgebühren auslösen. Um eine »Unterlizensierung« zu vermeiden, sollte im Sinne einer umfassenden Datenstrategie dokumentiert werden, ob Daten, die nicht selbst erhoben wurden, entsprechenden Einschränkungen unterliegen. Eine entsprechende Dokumentation stärkt zudem das Datenmanagement und den Überblick über die in der jeweiligen Einrichtung vorhandenen Daten. Ebenfalls zu beachten ist, dass in einer Organisation vorhandene Daten Vertraulichkeitsvereinbarungen mit Dritten unterliegen können oder sogar Geschäftsgeheimnisse verkörpern, die nach dem Geschäftsgeheimnisschutzgesetz aktiv zu schützen sind. Eine fehlende Dokumentation zur Vertraulichkeit entsprechender Daten kann dazu führen, dass diese bei der Nutzung von KI-Systemen verarbeitet und im Ergebnis Dritten zugänglich gemacht werden, da das KI-System ohne entsprechende Einstellung ggf. nicht zwischen »gewöhnlichen« und besonders zu schützenden Daten unterscheiden kann. Eine Offenlegung solcher Daten kann sowohl eine Verletzung von Vertraulichkeitsvereinbarungen bedeuten und damit Vertragsstrafen nach sich ziehen, als auch den Verlust von eigenen Geschäftsgeheimnissen bewirken. Da die mithilfe von KI-Systemen generierten Ergebnisse in der Regel nicht urheberrechtlich geschützt sind, jedoch Rechte Dritter verletzen können (siehe: Rechtliche Risiken minimieren), ist je Anwendungsfall festzulegen, wie die mit den KI-Systemen erzielten Ergebnisse genutzt werden. Sollen KI-generierte Ergebnisse Dritten zur Verfügung gestellt werden, ist im Hinblick auf den grundsätzlich fehlenden urheberrechtlichen Schutz KI-generierter Ergebnisse zu prüfen, ob eine Weitergabe an Dritte unter Zugrundelegung einer Lizenzvereinbarung erfolgt, welche die Rechte und Pflichten der Parteien festlegt.Leitfragen
- Welche Datenbestände sind für eine zweckmäßige und profitable KI-Anwendung unabdingbar?
- An welchen betrieblichen Stellen und in welchem Umfang werden Daten für mittel- und langfristige Zwecke gesammelt?
- Wie wird Datenqualität beim Aufnehmen der Daten sichergestellt?
- Welche Anforderungen an Datenhaltung und Datenverarbeitung existieren?
- Bestehen Beschränkungen der Nutzbarkeit von Daten z. B. aufgrund von Regelungen in Datenlizenzen? Unterliegen Daten Vertraulichkeitsvereinbarungen mit Dritten oder sind sie sogar als Geschäftsgeheimnisse aktiv zu schützen?
- Besteht eine Dokumentation zu den in der Organisation vorhandenen Daten?
- Welche Mitarbeitende und Organisationseinheiten greifen heute sowie in Zukunft auf welche Art und Weise auf welche Daten zu?
- Welche Art der Datenhaltung ist angemessen für die Anforderungen der Organisation?
- Gibt es je nach Organisationsstruktur und -größe Unterschiede nach Standort oder Bereich?
- Für welche Verarbeitungszwecke wurden personenbezogene Daten in der Vergangenheit erhoben? Über welche Verarbeitungszwecke wurden Betroffene informiert?
- Sind diese Verarbeitungszwecke kompatibel mit dem geplanten neuen Einsatz der KI?
Maßnahmen und Werkzeuge
- Dokumentation der in der Organisation vorhandenen Daten mitsamt möglichen Einschränkungen aufgrund fehlender Rechte oder aber dem Erfordernis der Vertraulichkeit
- Aufstellung eines möglichst allgemeinen Rollen- und Rechtekonzepts
- Betriebsvereinbarung zur Einführung und Anwendung der für KI erforderlichen Programme
- Bewusste Entscheidung für Art und Menge von Datenlagern sowie deren Integration in die Organisationsinfrastruktur
- Definition von Soll-Prozessen für allgemeine Anwendungsfälle aus der Datenverarbeitung (z. B. mithilfe der Unified Modeling Language)
- Festlegung der Nutzung von mithilfe von KI-Systemen generierten Ergebnissen
- Nutzung von (Daten-)Lizenzen bei der Weitergabe mittels KI-Systemen generierter Ergebnisse.
Technologieeinsatz
Um mit KI-Anwendungen nachhaltig arbeiten zu können, ist es erforderlich, die technische Infrastruktur regelmäßig zu warten, zu aktualisieren oder gar neue Technologien in der Organisation einzuführen (Technologietransition). Dazu können Datenerfassungssysteme (Sensorik) und Datenhaltungssysteme (z. B. dedizierte Datenbanken, Data Warehouses oder Data Lakes) gehören, ebenso wie neue Software zur Analyse von Daten (z. B. Statistikprogramme). Ein typischer Technologiewandel ist der Wechsel von klassischer »On-Premises« IT zum »Cloud«-Paradigma, womit umfangreiche Wechsel von Arbeitsweisen, Verantwortlichkeiten und Technologien einhergehen. Diese technischen Veränderungen sind notwendig, um langfristig arbeitsfähig zu bleiben, verursachen jedoch zunächst Kosten und binden Arbeitskraft. Insbesondere müssen sich die Beschäftigten weiterbilden, um mit den neuen Technologien zu arbeiten, oder Prozesse und Rollen müssen neu definiert werden. Der Übergang von einer Technologie zur anderen sollte möglichst zentral geplant werden und gleichzeitig unter frühzeitiger Einbeziehung der wichtigen Stakeholder, insbesondere der späteren Benutzenden und der Vertretung der Mitarbeitenden erfolgen. Ziel sollte es sein, eine durchgängige Infrastruktur zu schaffen, statt inkompatible Insellösungen zu etablieren. Nur so können Daten organisationsweit zusammengeführt und effizient genutzt werden. Andererseits bringen Veränderungen Kosten mit sich und können Mitarbeitende mit der Zeit ermüden. Manche Technologiesprünge kann man auslassen oder zu einem späteren Zeitpunkt nachholen.Leitfragen
- Wird die Technologietransition zentral geplant, sodass inkompatible Insellösungen vermieden werden?
- Wie entscheiden wir, welche Technologietransitionen wichtig / unumgänglich sind, und welche wir aufschieben oder auslassen können?
- Sind für die Arbeit mit der neuen Technologie alle Prozesse sowie Rollen und Verantwortlichkeiten definiert und sind die betroffenen Personen über Änderungen informiert worden?
- Ist auch nach der Umstellung der datenschutzkonforme Einsatz gewährleistet (z. B. werden nicht mehr erforderliche personenbezogene Daten gelöscht oder anonymisiert)?
- Sind die nötigen Wartungsprozesse mitgeplant worden und sind die dafür nötigen Werkzeuge dafür vorhanden?
- Werden die relevanten Stakeholder, insbesondere Benutzende, frühzeitig in die Planung einbezogen? Welche Mitbestimmungsrechte des Betriebsrats gibt es?
- Wann ist der Betriebsrat zu beteiligen?
- Ist bekannt, wie eine geplante Technologietransition positiv auf die Erreichung der Unternehmensziele einzahlen soll?
- Wurden die Vorteile, die von einem Technologietransitionsprojekt erwartet werden, dokumentiert und transparent kommuniziert?
- Sind Ressourcen (Arbeitszeit, Budget) für die Weiterbildung der Mitarbeitenden vorgesehen worden, um den Umgang mit der neuen Technologie zu erlernen?
- Gibt es einen klaren Zeitplan für die Umsetzung?
- Welche Schwierigkeiten sind durch Prozessabhängigkeiten, Downtime und Datenmigration zu erwarten?
- Wie lange wird das neue System voraussichtlich nutzbar bleiben?
- Ist definiert, wie der Technologietransitionsprozess überwacht wird und wer für nötige Anpassungen sorgt, wenn unerwartete Probleme auftreten?
Maßnahmen und Werkzeuge
- SWOT-Analyse, um Chancen und Risiken einer Technologietransition abzuwägen.
Literatur
- Vorschlag der Europäischen Kommission für eine EU-Verordnung über die Regulierung von künstlicher Intelligenz (AI Act) (2021/0106).
- Daugherty, P. R.; Wilson, H. J.: Human + Machine. Reimagining Work in the Age of AI. Harvard Business Review Press, 2018.
- Huchler, N.; Adolph, L.; André, E.; Bauer, W.; Bender, N.; Müller, N.; Neuburger, R.; Peissner, M.; Steil, J.; Stowasser, S.; Suchy, O. (Hrsg.): Kriterien für die menschengerechte Gestaltung der Mensch-Maschine-Interaktion bei Lernenden Systemen. München: 2020. https://www.plattform-lernende-systeme.de/files/Downloads/Publikationen/AG2_Whitepaper2_220620.pdf.
- Huchler, N.: Komplementäre Arbeitsgestaltung. Grundrisse eines Konzepts zur Humanisierung der Arbeit mit KI. Zeitschrift für Arbeitswissenschaft 76 (2022) S. 158-175.
- Pokorni, B.; Braun, M.; Knecht, C.: Menschzentrierte KI-Anwendungen in der Produktion. Praxiserfahrungen und Leitfaden zu betrieblichen Einführungsstrategien. Projektbericht im Fortschrittszentrum »Lernende Systeme«. Stuttgart: Fraunhofer IAO, 2021.
- Floridi, L.; Cowls, J.: A Unified Framework of Five Principles for AI in Society. Harvard Data Science Review. https://doi.org/10.1162/99608f92.8cd550d1, veröffentlicht am 23. Juni 2019.
- Krafft, T. D.; Zweig, K. A.: Transparenz und Nachvollziehbarkeit algorithmenbasierter Entscheidungsprozesse. https://www.vzbv.de/sites/default/files/downloads/2019/05/02/19-01-22_zweig_krafft_transparenz_adm-neu.pdf
- Schein, E. H.: What is culture. In Frost, P.; Martin, J., Moore, L.; Lundberg, C.; Louis, M. (Hrsg.): Reframing organizational culture. Washington: Sage, 1991, S. 243–253.
- Schnalzer, K., Karapidis, A., Dworschak, B., Mozer, P. (2022). Vorgehensweise zur Kompetenzbedarfsermittlung in Organisationen, Stuttgart. https://publica.fraunhofer.de/handle/publica/419707
