Why works councils need to understand AI

Artificial intelligence rarely enters a workplace under one large banner. It arrives as a new button in office software, a ranking function in recruitment, an automatic translation on a building site or an optimisation feature in a rota. A supplier may call it analytics or automation. Employees can still find that it changes how work is assigned, how performance becomes visible and which skills count.

That makes AI literacy a practical requirement for employee representatives. In Germany, a works council, or Betriebsrat, has statutory information, consultation and co-determination rights. Those rights are useful only when the council can identify the system in front of it. If management and the vendor understand the purpose, data and account structure while the council is still learning the vocabulary, the negotiation starts unevenly.

No works council needs to build a language model. It does need enough knowledge to separate a writing assistant from a scoring tool, a forecast from a decision and a genuine human review from a quick click that confirms whatever the software recommends. These distinctions affect employees and determine which questions, safeguards and legal provisions belong in the discussion.

The difficult part begins after the product demonstration

AI looks tidy in a sales presentation. A meeting assistant produces a neat summary. A scheduling system removes idle time. Recruitment software finds the strongest candidates. Real workplaces contain incomplete records, unusual career histories, machine breakdowns, informal teamwork and exceptions that never reached the process manual.

A council that knows only the demonstration will debate features. A council that understands the workplace asks what happens around the feature. Was a safety instruction summarised from an approved source? Did a worker receive a lower score because a faulty machine reduced output? Does the system retain prompts or editing times? Can an update change the model, the data source or the administrative reports without another local implementation project?

Employee representatives bring information that a software supplier cannot provide. They know where the formal process differs from the work actually done. Technical literacy lets them turn that experience into testable questions. General enthusiasm produces weak scrutiny. General suspicion produces weak alternatives. Both leave the vendor's description largely intact.

Basic literacy means knowing what to ask

A first assessment can begin without code or advanced mathematics. For every proposed system, the works council should be able to obtain clear answers to questions such as these:

• What workplace problem is the system meant to solve, and how will success be measured?
• Which data does it receive, where does that data come from and does it relate to identifiable people?
• What does the system produce: text, a prediction, a score, a ranking, an alert, a recommendation or a decision?
• Who sees the output, and how does it affect hiring, shifts, tasks, appraisals, promotion or dismissal?
• Which prompts, clicks, corrections, response times or other actions are logged and available for later analysis?
• Which errors, biases and fabricated results are known or reasonably foreseeable?
• Who can disregard, correct or contest an output, and is that intervention recorded?
• What happens when the supplier changes the model, connects another data source or adds a reporting function?

Simple questions often expose the real design. Management may say that a person always makes the final decision. The next question is whether that person has the time, evidence and authority to disagree. Ten seconds to approve an automatically generated ranking gives the process a human participant. It does not provide meaningful human judgment.

The review cycle for workplace AI

The five review fields form a cycle: new findings about effects or controls often change the assessment of purpose, data and output.

German law provides tools, not a completed assessment

The German Works Constitution Act, the Betriebsverfassungsgesetz, addresses AI through several provisions. Section 80 requires the employer to give the works council complete information in good time, and to provide documents needed for the council's work. Where the council has to assess the introduction or use of AI, the Act treats the involvement of an expert under Section 80(3) as necessary. An expert can examine contracts or data flows. The council still has to commission the right examination and understand its consequences.^[1]

Section 90 matters before a plan is fixed. When an employer prepares technical installations, work processes, workflows or workplaces, information has to reach the council while objections and proposals can still make a difference. Employee monitoring has its own provision: Section 87(1)(6) covers technical facilities intended to observe behavior or performance. AI appears expressly in Section 95 on selection guidelines. If a system changes jobs or the capabilities they require, the training provisions in Sections 96 and 97 come into play.^[2][3][4][5]

No single provision applies to every AI feature. A tool that drafts an internal note differs from a camera system, an employee score or software that allocates work using personal characteristics. Calling all of them “AI” does not settle the legal analysis. Calling them ordinary software does not settle it either. Purpose, data, technical design and actual use matter.

A ChatGPT case from Hamburg shows why architecture matters

In January 2024, the Hamburg Labour Court considered a company policy that allowed employees to use ChatGPT and similar services. Access was voluntary, through a browser and private accounts. The employer could not inspect the chat histories. In that specific interim proceeding, the court rejected the claimed co-determination right under Section 87(1)(6), among other claims. The account and access arrangements were central to the reasoning.^[11]

The decision is sometimes shortened to “no co-determination for ChatGPT”. That reading loses the useful lesson. A company account with central login, usage records, administrative dashboards or employee-level reporting presents a different set of facts. The same branded product can have a different workplace effect depending on its configuration. A council that does not ask about logs, identity, interfaces, storage and administrator access may treat two different systems as though they were identical.

Technical literacy here is modest and concrete. Ask who owns the accounts, where the prompts go, which records remain, whether the supplier trains on the material and who can retrieve usage data. A statement that the supplier is GDPR-compliant does not answer those questions.

A building site makes the range of workplace AI visible

Move from an office to a construction site and AI changes its appearance. At the morning briefing, an app translates the instructions. During the shift, a camera checks whether protective equipment is missing. Software forecasts material needs and builds the crew schedule from availability and qualifications. Another model watches for signs that equipment may fail. By the end of the day, a voice assistant has turned spoken notes into the site report.

The label is the same; the consequences are not. A mistranslated trade term can undermine the safety benefit of a multilingual briefing. The camera may catch a missing helmet, yet its feed also creates the possibility of constant observation. A faster crew schedule might ignore physical strain, hard-won experience or the value of keeping people together. Even the site report needs checking, because an invented detail must not enter the official record.

The useful question is therefore narrower than “Should construction use AI?” For each application, the council needs to know which task changes, which data appears and who carries the cost of a wrong result. That is the basis for negotiating access, training, review and limits.

Consultation can improve the implementation

Technology projects sometimes describe employee consultation as delay. OECD workplace surveys offer a different perspective. In its 2022 survey, workers were more likely to report positive effects on performance and working conditions when workers or their representatives had been consulted about new workplace technologies. Employers that consulted also reported positive effects more often. The report describes an association and warns against treating it as simple proof of causation. Even with that caution, the mechanism is credible. Employees know the workarounds, dependencies and exceptions missing from a process map.^[9]

An informed works council can make a pilot more useful to the employer as well as safer for the workforce. Time saved is one measure. Error rates, rework, stress, safety, distribution of demanding tasks and the ability to challenge a recommendation are measures too. A trial that records only speed can reward a system that moves hidden work onto employees.

Training deserves the same scrutiny. The OECD found that trained AI users were more likely to report improvements in working conditions, although training was also associated with greater concern about job stability. Teaching people where to click is insufficient. They need to recognize unreliable output, know which data must stay out of the tool and have somewhere to report a failure. Managers need rules for cases in which an AI output must not become evidence for a personnel decision.

The EU AI Act turns literacy into an organizational matter

AI literacy is a duty under Article 4 of the EU AI Act. Providers and deployers must, to the best of their ability, take measures that give staff and others operating systems on their behalf an appropriate level of knowledge. According to the Commission, the requirement has applied since 2 February 2025. Its guidance reaches an everyday use case: employees who ask ChatGPT to write advertising copy or translate text should know about risks such as hallucinated content.^[6][8]

This does not mean giving everyone the same course. The system, the operator's role and the people affected determine what knowledge is appropriate. Reviewing draft translations calls for different preparation than supervising an employment system. The Act treats certain recruitment and personnel-management systems as high-risk, including tools that allocate tasks from personal characteristics or monitor performance and behavior. Under Article 26, worker representatives and affected employees are also to be informed before a high-risk system is used at work.

A council looking for a firm start date will not yet find one. The EU simplification process continues. Following the political agreement reported in May 2026, the Commission's current implementation page points to 2 December 2027 for employment systems and other listed high-risk uses. Before deployment, the organization must check the consolidated law then in force. None of this changes the council's immediate task when a product appears: understand how it works and test the controls promised for it.^[7]

Build a map before the next formal AI project arrives

Many councils begin with the tool that management has just submitted for consultation. A short internal inventory provides a better starting point. Which AI features are already enabled in office platforms? Which services are employees using on their own? Which systems process HR records, customer conversations, safety information or performance data? Which suppliers can change a model or add analytics without a conventional local software upgrade?

No elaborate register is required. A short table can record what each tool does, whose data it uses and whether it influences a decision. Put an informal translation aid beside an applicant filter and the difference in scrutiny becomes obvious. The same is true of a chatbot with no personal data and a platform that reports activity by user. Sorting the tools this way preserves council time for health, rights and consequential decisions.

Employees do not always wait for a formal AI project. If there is no sanctioned tool, someone may open a private account and paste in commercial secrets or personal information. That data has now left the organization's controlled environment. A blanket ban may hide the behavior without ending it. Providing an approved service and rules that work on the job is often the safer response.

A useful works agreement can survive a product update

The map is only a starting point because AI services change more often than traditional workplace equipment. A supplier may replace the underlying model, connect a new source or introduce employee-level reports. A works agreement tied only to a product name and version can become outdated while the interface still looks familiar.^[10]

The agreement should state the permitted purpose and excluded uses. It should cover data categories, retention, access, logs, human review, complaints and the handling of material changes. For employment decisions, it should say what role an AI output may play and which judgment remains with a person. A pilot needs measures agreed in advance and a real decision point at which the parties can continue, revise or stop it.

Employee feedback belongs in that design. The workforce will notice when a translation repeatedly mishandles a trade term, when reporting creates extra work or when one group receives poorer recommendations. A simple reporting route gives both employer and council evidence that will never appear in the vendor brochure.

Outside expertise works best with an informed client

Some questions need outside help. When a council cannot tell where information goes after a user clicks Submit, a specialist can trace the route. The same review may examine a data protection assessment, a vendor contract or the legal position. German law explicitly provides for experts when AI must be assessed, making that support easier to obtain. The council still has to decide which workplace concern comes first.

A useful expert brief begins with something that can be tested. Take one log: the reviewer traces whether it can be linked to an individual worker and whether it travels outside the European Economic Area. For a personnel ranking, the reviewer compares its formal status with its role in the actual decision. The same assignment can identify changes the vendor may make without renewed approval. This gives the council evidence for negotiation.

Training need not turn every council member into a technical specialist. Some members can go deeper; everyone else still needs command of the basic questions. Section 37(6) of the Works Constitution Act provides paid release when knowledge is required for council work. Whether a specific AI course qualifies depends on conditions in the establishment and the tasks approaching the council. A current or planned deployment supplies a concrete connection that can be assessed and documented.

AI literacy gives the council something useful to negotiate with

A works council is neither the sales team for workplace technology nor a committee formed to stop it. Its job is to help shape good work and to set limits where monitoring, discrimination, loss of professional judgment or unclear responsibility threaten employees. A product name provides too little information for that task.

A capable council can enter before the design is fixed. It can tell whether a pilot measures the right outcomes, distinguish a drafting tool from an evaluation system and test claims of human control against the actual workflow. It can also propose workable alternatives: a smaller data set, narrower access, a training plan, a limited trial, an appeal route or removal of a problematic function.

Basic AI knowledge has therefore become part of institutional capacity. Representatives asked to shape tomorrow's work need to understand how today's software prepares decisions, records behavior and distributes tasks. They do not need every technical answer in advance. They need enough command of the subject to recognize when an answer is missing.

When management, a works council and employees need a shared starting point for an AI program, a keynote can establish the vocabulary before the negotiation begins. I would be pleased to develop a talk that connects technical basics with workplace examples, employee participation and the limits of automated judgment.