Co-Determination Applies as a Rule
If your employees are going to use the tool, § 87 (1) 6 BetrVG will as a rule apply and the works council holds an enforceable co-determination right, at any rate once the system, or the platform underneath it, logs usage per person and the employer can access those logs. Many project teams start by arguing the opposite (we are not monitoring anyone, we are drafting claim letters) and lose, because the provision does not ask about intent.
The test is objective suitability
The provision covers technical systems that are “bestimmt” to monitor the performance or behaviour of employees. The Bundesarbeitsgericht has read that word as an objective test for decades: it is enough that the system is suitable for producing performance or behaviour data about identifiable people. Nobody has to want it. If your assistant writes a user ID, a timestamp and a query into a trace so you can debug a bad answer, you have built such a system. The platform underneath it is one anyway, because a Microsoft 365 tenant attributes activity to individual users before you deploy anything on top of it. The assistant is usually the smallest monitoring-capable component in the stack.
A pilot is not a co-determination-free zone
“It is only a trial” buys no exemption, because a test run with real employees on real data is introduction and use. If the works council hears about the tool from an employee and not from you, the project loses trust it cannot rebuild, and from then on you are negotiating your credibility instead of the scope. Conclude an explicit pilot agreement instead, with one narrow use case, a fixed user group and an expiry date. The works council signs it more readily, because nobody is being asked to approve a permanent state, and the expiry date forces a verdict.
Four Provisions of the BetrVG
Four provisions apply, and they attach at different moments, which you can use.
§ 90 (1) 3 has named Künstliche Intelligenz explicitly since the Betriebsrätemodernisierungsgesetz of 2021. It attaches during planning, it obliges you to inform and consult in good time, and it is not a veto, which is why teams skip it. That is one reason a short negotiation turns into a long one. § 87 (1) 6 is the enforceable one: no agreement means an Einigungsstelle under § 76, and a works council can have a court prohibit the use of a system introduced without co-determination. That switch-off risk lands after you have paid for the build. § 80 (3) 2, also from 2021, says that where the works council has to assess the introduction or use of AI, bringing in an expert counts as necessary to that extent, and you bear the cost. Finally, § 95 (2a) extends co-determination over selection guidelines explicitly to cases where AI is involved.
Many treat that expert clause as a threat, yet it works in your favour. The expert has to write a recommendation, and an expert who cannot follow your data flows will recommend the most restrictive version they can defend, because they can defend no other. Vagueness therefore costs you room. A data-flow document a stranger can understand in an hour is the cheapest way to strengthen your negotiating position.
The Architecture Decides the Scope
A regional insurer builds a retrieval assistant over policy wordings and claim files for its motor claims handlers. The team logs the query, the retrieved chunks, the answer, the user ID and the latency. That is not sloppiness, because you need this data to measure quality and to run the system in production at all. Unmodified, it also yields a report on who asks the most questions and who takes longest per case. The works council sees that in the first meeting.
Teams then usually promise not to look. Make looking structurally hard instead and split the telemetry into a quality store that carries no user ID and serves evaluation and regression tests, and an audit store that is tied to persons, deleted after a fixed period, and reachable only through a defined procedure. Name the cost in the meeting: pseudonymised evaluation logs take away your ability to walk over to the handler and ask what they meant by that query, and debugging gets slower. I would pay that price, because it turns a long argument into a paragraph.
Data minimisation alone does not solve the problem, however. If the system counts as high-risk, the AI Act conversely requires logging and retention. The GDPR wants minimisation (that is a topic of its own), the AI Act wants a record, and § 87 wants no evaluation of performance. These demands do not cancel out. So you keep the logs and make them useless for evaluation through access rights and the agreement instead of deleting them.
The classification, by contrast, you can avoid by design. An assistant that drafts a letter and leaves the decision with the handler differs from one that scores handlers or allocates their work. Once per-person telemetry flows into something that evaluates people, you arguably fall under Annex III of the EU AI Act, conformity assessment included, and you hand the works council its strongest argument at the same time. You cannot avoid co-determination, but you can often avoid the high-risk classification. Teams conflate the two and end up negotiating the wrong one.
What a Betriebsvereinbarung for AI Tools Covers
Before drafting anything, check whether the agreement already exists. Your Microsoft 365, ticketing or ERP agreement may already regulate this system or already forbid what you are planning, because an evaluation ban written years ago for a phone system can cover your new system too. An annex to an existing agreement is often faster than a new one, and your works council chair can answer the question in five minutes.
Beyond that, these agreements in practice mostly regulate the same points:
- Scope — which system, which data sources, which user groups, which purposes. A narrow scope beats a broad one every time.
- Purpose limitation and the evaluation ban as the core of the agreement: the system's data may not be used to assess individuals, with one named exception path for concrete suspicion, under four eyes, with the works council involved.
- Data categories, retention and deletion periods, and role-based access to the logs.
- No automated decision about a person — a human decides, and stays accountable for the decision.
- Transparency and training for the people expected to use it.
- Inspection rights for the works council, including a test account.
- A term and a review date — and what happens to the tool when it lapses.
The change clause
Here an AI agreement differs from the one your company once wrote for time recording. That system did not change, and yours does: the vendor swaps the underlying model, and behaviour shifts with no deployment on your side, no ticket and no announcement. An agreement that fixes “System X, version Y” is either breached continuously or renegotiated every quarter, and neither side wants either.
Describe the system by its data flows and purposes and not by its version. Then define what counts as a material change (a new data source, a new user group, a new purpose, a new decision the system makes) and what does not: model version, interface, retrieval tuning. The first category gets full co-determination, the second a lightweight notification duty. A works council can accept that split, because it assures them they will hear about the first category instead of discovering it in a release note.
What I Would Not Do
I would not start with a framework agreement covering “AI” in general. Settling the question once and covering everything after is tempting. But whoever negotiates an abstract framework with no concrete system negotiates against imagined worst cases, and the framework comes out either so tight that it blocks reasonable uses or so vague that every rollout reopens it. Conclude one narrow agreement for one real tool, then a second, and generalise from those two. The case differs when you are switching on Copilot for the whole company on Monday; then you lack that time and should get help.
Nor would I submit the vendor’s “works-council-ready” datasheet as your documentation. It describes their product, while the expert is assessing your operation, your data sources and your log configuration. And I would not promise anything the system cannot demonstrate. Whoever opens the tracing tool and finds a UPN in a span refutes the sentence “No personal data in the logs”. Promise only what you can show.