The provisions
- § 87(1) no. 6 BetrVG: enforceable co-determination on the introduction and use of technical devices designed to monitor employees' behaviour or performance.
- § 90(1) no. 3 BetrVG: information and consultation on the planning of working procedures and workflows “including the use of artificial intelligence”, early enough for suggestions and concerns to still influence the planning.
- § 80(3) sentence 2 BetrVG: if the works council has to assess the introduction or use of AI, calling in an expert is deemed necessary to that extent. A more detailed agreement with the employer under sentence 1 is still required.
- § 95(2a) BetrVG: selection guidelines for hiring, transfers or dismissals remain subject to co-determination, even if AI is used in drawing them up.
Why intent does not help
A device is designed for monitoring if it itself collects and records individualised or individualisable behavioural or performance data, regardless of whether the employer intends to analyse it. A de minimis threshold is irrelevant.
The case concerned an Excel spreadsheet for recording attendance times. The court had already established in its decision of 13 December 2016, 1 ABR 7/15, that what counts is the objective suitability for monitoring, not the employer's intent. “We never look at it” is therefore no defence; nor is “it's only a pilot”, and “participation is voluntary” even less so.
With AI tools, this is triggered sooner than you might think: user ID, project, timestamp, executed command, usage, acceptance rate of suggestions, response time or quality ratings make a tool objectively suitable for monitoring, even if it is introduced as a security or cost measure.
What metrics a pilot may collect
| Defensible | Activation per team, aggregated task completion, aggregated run time and reliabilityIn addition, professionally assessed results, prevented security incidents and voluntary interviews whose raw answers managers do not see. |
|---|---|
| To avoid | Anything counted per personActive hours, inputs per person, acceptance rate per person, response speed, individual quality or risk ratings, rankings, breakdowns for managers, derived working time. |
| Minimum group | Analysis only above a group size at which individuals cannot be inferredPseudonymous IDs remain personal data as long as you or the provider can restore the link, via project, timestamp, device or usage patterns. According to the Federal Labour Court, pure group data only triggers co-determination if the monitoring pressure carries through to the individual group members (1 ABR 7/15, para. 27). |
What belongs in the works agreement
- Permitted purposes and an express ban on use for performance appraisal and disciplinary measures
- The fields collected, and those technically switched off
- Aggregation and minimum group size
- Who has access in a security incident and in support, and in what role
- Retention and deletion, including deletion at the end of the pilot
- A documented procedure for suspicions and investigations instead of access without cause
- Employees' rights to information, access, rectification and objection
- Involvement of the data protection officer, works council and internal audit
- Change control for model, sub-processors, data paths and logging
- Consequences of a breach, including switching off the function concerned
The legal basis, and why not consent
For employee data, § 26 BDSG (Federal Data Protection Act) comes into consideration; its wording still applies unchanged as of 23 September 2026, and no separate employee data protection act is in force. A works agreement can serve as a legal basis under § 26(4) BDSG in conjunction with Art. 88 GDPR if it meets the requirements of Art. 88(2) GDPR. Relying solely on § 26(1) has been uncertain since the CJEU judgment of 30 March 2023: a national employment rule is only a “more specific rule” if it satisfies Art. 88(2); otherwise it is not applied (C-34/21).
Consent is no way out. § 26(2) BDSG requires the dependency inherent in the employment relationship to be taken into account when assessing whether consent is voluntary; a pilot called “voluntary” does not cure unnecessary or disproportionate data collection.
And the upper limit: Art. 22 GDPR restricts decisions based solely on automated processing that have significant effects. According to the CJEU's SCHUFA judgment of 7 December 2023 (C-634/21), the automated generation of a value can itself be such a decision if the downstream party relies on it to a decisive extent. A human who nominally reviews but regularly follows the value is not a reliable safeguard.
What usage data does Biwak generate?
Biwak collects usage data per account; each person signs in with their own access. If several people work as a team from one quota, which is possible from the Rope Team plan upwards, the team owner sees how many credits each person has used in the current billing month, but not their tasks, files or conversations. For every model request, the account records when it ran, with which model and to what extent, and whether it came from the app or the browser; quota and billing are calculated from this. The data is stored in a database in Frankfurt am Main. The app itself sends no telemetry and no usage statistics; conversations from the app only reach Biwak in an error report or a message that the user sends themselves. If someone uses the browser workspace, their conversations are also stored in the database; Biwak only looks at them for a named reason, and every access is logged. Details are in the privacy policy.
On the computer, Biwak keeps a local log of which tasks the user gave, which files the agent read and changed, and the checkpoints before every change. Biwak does not collect keystrokes, screen content or presence, and there are no rankings by personal performance; the overview is on the security page. Biwak does not record working hours as such, but the times of the tasks would show when someone worked with Biwak.
Usage data with timestamps per account is data that can be attributed to a person. Whether the use falls under § 87(1) no. 6 BetrVG therefore depends on who can view and analyse it: only the individual in their own account, or the employer too, for example as administrator of an organisation in the Biwak account. Before the roll-out, clarify which details administrators see and discuss this with the works council. Two things apply in any case: § 90 BetrVG requires timely information, and anyone who collects or analyses the local logs creates the data collection themselves.
A model works agreement for the use of Biwak is in progress; on request to kontakt@biwak.ai you will receive the current draft.
Overlap with the EU AI Act
AI systems that help decide on hiring, promotion or dismissal, allocate tasks based on individual behaviour or evaluate performance and behaviour are high-risk under Annex III no. 4 of the EU AI Act; a writing or analysis assistant is not high-risk merely because employees use it. For high-risk systems, employers must inform workers' representatives and employees before deployment from 2 December 2027 (Art. 26(7)). Classification and deadlines are covered in the guide to the EU AI Act.
Frequently asked questions
We don't have a works council. Does that settle the question?
No. Without a works council, co-determination and the duty to inform under the BetrVG no longer apply, but the obligations under data protection and employment law remain: legal basis, proportionality, informing employees, deletion concept. In the public sector, the relevant staff representation law applies.
Is it enough to inform the works council once we have already decided?
No. § 90 BetrVG requires information early enough for suggestions and concerns to feed into the planning. Informing the works council after the contract has been signed does not fulfil the purpose of the provision and creates exactly the conflict you wanted to avoid.
The provider says there is no monitoring function. Is that enough?
Ask about fields, not functions. What matters is which data is collected and stored and who can see it, not whether there is an analysis screen. An event log with user ID and timestamp is suitable for monitoring even if nobody has built a report.
What is the smoothest way to introduce such a tool?
In this order: define purpose and fields; inform and consult early; document the threshold assessment for the data protection impact assessment; limit metrics to team level and switch off individual ones; conclude an agreement that expressly excludes performance appraisal; only then install. The less a tool collects about people, the shorter this path becomes, but it cannot replace it.
Sources
- § 87 BetrVG (co-determination rights)
- § 90 BetrVG (rights to information and consultation)
- § 80 BetrVG (general duties, experts)
- § 95 BetrVG (selection guidelines)
- BAG, decision of 23 October 2018, 1 ABN 36/18
- BAG, decision of 13 December 2016, 1 ABR 7/15
- § 26 BDSG (employee data)
- CJEU, judgment of 30 March 2023, C-34/21
- CJEU, judgment of 7 December 2023, C-634/21
This text is not legal advice. It is the groundwork we had to do for ourselves, with the legal references, so that your lawyer does not have to start from scratch. Where a question depends on your circumstances, the text says so.