Works Council and AI Adoption: What § 87 BetrVG Requires Before Go-Live
The question comes up in almost every project, usually shortly before go-live: does the works council have to approve the AI assistant? In most cases the answer is yes, and the reason is not the artificial intelligence. The reason is that almost every one of these systems incidentally generates data about what employees do and when. That is exactly what the works council's co-determination right is about, and it has been since 1972.
This article classifies the usual use cases in mid-sized companies, names the rights of the works council with the relevant section, describes what belongs in a works agreement and shows how the procedure stays short. It is written from the perspective of someone who builds and introduces such systems, not from that of a lawyer; where interpretation in an individual case is required, we say so.
In brief: Under § 87 para. 1 no. 6 BetrVG the works council co-determines when a technical device is introduced that is suitable for monitoring the behaviour or performance of employees. Under settled case law suitability is enough; intent is irrelevant. Phone assistants, back-office automation and company access to chat services almost always generate such data. The works council must also be informed at the planning stage under § 90 BetrVG, may call in an expert for AI under § 80 para. 3 BetrVG, and for high-risk systems Art. 26 para. 7 of the AI Act requires informing employee representatives before putting the system into service. Anyone who bypasses this risks a shutdown by urgent order. Anyone who plans it early is done in weeks.
Does the works council have to approve before the AI goes live?
Yes, as soon as the system is suitable for monitoring the behaviour or performance of employees. § 87 para. 1 no. 6 BetrVG speaks of technical devices intended to monitor behaviour or performance; the Federal Labour Court has interpreted that broadly for decades: what matters is not whether the employer wants to monitor, but whether the device can objectively do so. A system that stores transactions with a timestamp and the name of the person handling them can.
For the use cases that actually occur in mid-sized companies, a clear mapping follows:
| Use case | Does the system generate data about employees? | Co-determination under § 87 para. 1 no. 6 |
|---|---|---|
| Website chatbot that answers customer questions and hands over to a shared mailbox | No | No, as long as nothing is assigned to individuals |
| Website chatbot with handover to individual employees and measurement of response time | Yes | Yes |
| AI phone assistant that answers calls, books appointments and forwards to individuals | Yes, forwarding logs with timestamps | Yes |
| Back-office automation (incoming invoices, inbox) with approval by clerks | Yes, approvals with person and time | Yes |
| Internal knowledge assistant answering staff questions | Yes, if queries are logged per user | Yes, unless the logs are not personal |
| Company access to ChatGPT or another chat service with employer admin rights | Yes, usage logs | Yes |
| Private use of ChatGPT through employees' own accounts | No, no employer access | No (Hamburg Labour Court, 16 January 2024, 24 BVGa 1/24) |
| AI-supported pre-selection of job applications | Yes, additionally high-risk under the AI Act | Yes, plus § 95 para. 2a BetrVG (selection guidelines) and Art. 26 para. 7 AI Act |
The table shows what matters: not the label AI, but the question whether somewhere a person is stored with a timestamp. That is also the lever for design. An incoming-invoice agent that counts documents and assigns approvals to a role rather than a person is a different system from one that evaluates processing time per clerk. Both do the same work.
The decision of the Hamburg Labour Court of 16 January 2024 is so far the best-known case on the question. A group works council wanted to have the use of ChatGPT prohibited. The court declined: employees used their own accounts, the employer had no access to the usage data, so the possibility of monitoring was missing. The employer's usage guideline moreover governed work conduct, meaning the way work is to be done, and not the orderly conduct subject to co-determination under § 87 para. 1 no. 1 BetrVG. The decision is a first-instance ruling in urgent proceedings, not a line set by the highest court. But it shows the logic courts follow: who has access to which data.
What rights does the works council have besides § 87?
Five more, and three of them were created or extended specifically for AI in 2021 with the Works Council Modernisation Act:
| Provision | Right | What it means in the project |
|---|---|---|
| § 90 para. 1 no. 3 BetrVG | Information and consultation when planning work procedures and workflows, expressly including the use of AI | The works council belongs in the planning, not in the acceptance. Anyone who only arrives with the finished system has already breached this duty |
| § 80 para. 3 sentences 2 and 3 BetrVG | Calling in an expert; for AI the necessity is deemed given | By closer agreement with the employer, the works council can bring in an external adviser at the employer's expense without having to justify the need. Plan for it instead of disputing it |
| § 95 para. 2a BetrVG | Selection guidelines for hiring, transfer and dismissal remain subject to co-determination even when AI is used in drawing them up | Whether every AI pre-sorting of applications falls under it is disputed; anyone introducing it should assume so |
| § 87 para. 1 nos. 1 and 7 BetrVG | Order in the establishment and conduct of employees; health protection | Rules on private use of AI services and questions of workload caused by new systems can fall under these |
| Art. 26 para. 7 AI Act | For high-risk systems in the workplace, employee representatives and affected employees must be informed before the system is put into service | Applies in addition to the BetrVG, regardless of whether a works council exists |
Then there is data protection law. A works agreement can be the legal basis for processing employee data under Art. 88 GDPR and § 26 para. 4 BDSG; in the employment relationship it thus often replaces consent, which in practice would rarely be freely given anyway. That is one reason to see the agreement not as a hurdle but as the document that legally carries the rollout.
What belongs in the works agreement?
Nine points, and a well-planned project has most of them documented anyway:
- Purpose and scope. What the system is used for and what it is not used for. The negative list matters more than the positive one.
- The specific systems. Provider, function, where the data is processed. In a framework agreement as an annex per system.
- Data about employees. Which data arises, where it is stored, who sees it. The honest answer is usually: timestamp, handler, transaction.
- Ban on performance and behaviour monitoring. Evaluation per person is excluded; evaluations are only aggregated or anonymised. This sentence is the core of every agreement and the reason it comes about at all.
- Access rights and deletion periods. Who may inspect logs, in which cases, and when they are deleted.
- Training. Who is trained, when, with what content. This also satisfies Art. 4 of the AI Act, and the training documentation serves both purposes.
- Rights of the works council in operation. Information about changes, inspection of evaluations, involvement in extensions.
- Procedure for changes. What counts as a change that needs a new annex and what counts as maintenance.
- Review. A date, usually after twelve months, at which both sides check whether the agreement still fits the system.
Bitkom published a guide on AI and co-determination in February 2026 that contains model structures. It does not replace a negotiation, but it shortens one, because both sides start from the same outline.
What happens if the works council is bypassed?
The system stands still, and at the request of the works council, not by decision of the employer. The works council has a claim for cessation that it can enforce through urgent proceedings before the labour court; the court can prohibit use until an agreement or a ruling of the conciliation committee under § 87 para. 2 BetrVG exists. For a project whose benefit lies in ongoing relief, that is more expensive than any negotiation beforehand.
Two consequences last longer. First: data collected in breach of the co-determination right stands on shaky ground in later dismissal proceedings. A ban on using it does not follow automatically from the breach alone, as the Federal Labour Court held on 29 June 2023 (2 AZR 296/22), but the evidence becomes contestable, and the employer carries that risk, not the employee. Second: the next rollout takes longer, because the works council now examines every system as if it were the first.
How does the procedure stay short?
Through three decisions that all come before the negotiation.
Involve early instead of informing late. § 90 BetrVG requires information at the planning stage; anyone who invites the works council to the process-recording workshop fulfils the duty and at the same time wins over the person who later co-signs the agreement. The process-recording workshop is the natural place for that; the exceptions the agreement will later deal with are discussed there anyway.
Count transactions, not people. The system is built so that evaluations per person are technically not provided for: approvals are assigned to roles, processing times are aggregated only per transaction type, logs are anonymised after a short period. What is technically impossible does not have to be prohibited, and the agreement becomes half as long.
Framework instead of individual case. A framework works agreement on AI settles the principles once and adds each new system as an annex. The second chatbot then needs a two-page annex instead of a three-month negotiation.
The cost of the expert under § 80 para. 3 BetrVG is money well spent: a works council that understands the system signs faster than one that fears it. The technical documentation it needs is produced anyway for the AI inventory under the AI Act and should be part of what the provider delivers.
What applies without a works council?
Then co-determination falls away, but not the rest. A works council can be elected in establishments with at least five permanent employees entitled to vote; many mid-sized companies have none, and there nobody co-determines under § 87 BetrVG. What remains is data protection law with § 26 BDSG and Art. 6 GDPR for employee data, the information duty under Art. 13 GDPR, the literacy duty under Art. 4 of the AI Act and, for high-risk systems, informing the affected employees under Art. 26 para. 7 of the AI Act.
Even without a works council, it makes sense to inform the workforce before the start and putting the rule "transactions, not people" in writing. Not because a law requires it, but because an assistant the team distrusts is not used, and an unused assistant does not pay off.
Frequently asked questions
Do I have to involve the works council before the AI phone assistant goes live? As a rule, yes. A phone assistant logs which calls it forwards to whom and how long handling takes. That makes it a technical device suitable for monitoring behaviour or performance, and § 87 para. 1 no. 6 BetrVG gives the works council a co-determination right over its introduction and use. Whether monitoring is intended is irrelevant under the settled case law of the Federal Labour Court; objective suitability is enough.
Does co-determination apply even if only customers talk to the chatbot? Only if the system generates or evaluates data about employees. A website chatbot that answers customer questions and routes handovers to a shared mailbox does not touch the workforce. As soon as it assigns handovers to individual employees, measures response times or attributes conversations to individual persons for quality control, the threshold is crossed.
Is private use of ChatGPT by employees subject to co-determination? According to the decision of the Hamburg Labour Court of 16 January 2024 (24 BVGa 1/24), not as long as employees use their own accounts and the employer has no access to the usage data. Then the possibility of monitoring is missing, and a usage policy governs work conduct rather than orderly conduct. With a company account whose logs the employer can inspect, the picture changes.
What happens if we introduce the AI without the works council? The works council can demand that use be stopped, if necessary through urgent proceedings before the labour court, and the system has to be switched off until an agreement or a ruling of the conciliation committee exists. Data collected in breach of the co-determination right is contestable in later dismissal proceedings. On top comes the loss of trust that slows down every subsequent rollout.
What belongs in a works agreement on AI? Purpose and scope, the specific systems, the data processed about employees, an explicit ban on performance and behaviour monitoring per person, access rights and deletion periods, training under Article 4 of the AI Act, information and inspection rights of the works council, the procedure for changes and a date for review. A framework agreement that adds new systems as annexes has proven itself.
How long does co-determination take, and how does it get shorter? Months without preparation, weeks with it. Three things shorten the procedure: involving the works council at the planning stage under § 90 BetrVG instead of arriving with the finished system, building the system so that it counts transactions rather than people, and working with a framework agreement so that not every system needs its own negotiation.
Conclusion
Co-determination is not an obstacle to AI adoption but its condition, as soon as the system generates data about employees. Anyone who treats it as part of process recording rather than as a hurdle before go-live has the works council at the table during planning, a system that counts transactions rather than people, and an agreement that at the same time provides the data protection basis.
Which duties the AI Act sets per use case is in the article on the EU AI Act for AI agents; what the managing director answers for during the introduction is in the guide Introducing AI in a mid-sized company.
This is not legal advice. Whether a specific system is subject to co-determination and how a works agreement should be drafted in an individual case is for an employment lawyer to determine.
Sources
- § 87 BetrVG - Co-determination rights - Gesetze im Internet
- § 90 BetrVG - Information and consultation rights - Gesetze im Internet
- § 80 BetrVG - General duties - Gesetze im Internet
- § 95 BetrVG - Selection guidelines - Gesetze im Internet
- Artificial intelligence and co-determination, guide (February 2026) - Bitkom e.V.
- First ruling on works council rights when AI is used (Hamburg Labour Court, 24 BVGa 1/24) - Bird & Bird
- Article 26 AI Act: Obligations of deployers of high-risk AI systems - artificialintelligenceact.eu
- § 26 BDSG - Data processing for employment-related purposes - Gesetze im Internet
AI process automation with a legal framework
One bounded process first, with a data processing agreement, EU hosting and the technical documentation the works council and the data protection officer want to see.
More Articles
AI Automation for SMEs: A Guide to Boosting Efficiency in 2025
36 percent of German firms cannot fully fill open positions, according to the DIHK. Which processes can be automated first, how to recognise them, and why most projects fail.
WhatsApp Business API & AI: How Companies Generate More Revenue
WhatsApp reaches customers more reliably than email. But without documented consent under Section 7 UWG, every promotional message is a legal risk. What has to be settled first.