AI works agreements: when the works council has a say

The actual use determines the AI works agreement

The use case determines the participation rights. An AI works agreement sets conditions agreed between an employer and its works council for using a system at work. Whether co-determination is required depends on its functions and the workflow. Technical monitoring of performance or behaviour is one important trigger, but not the only one.

Our recommendation for smaller businesses: start with a one-page description of the intended use rather than a generic contract template. The works council needs concrete details of the employee data generated and who can access it. This article covers German workplaces within the scope of the Works Constitution Act and does not replace an assessment of an individual case.

The rules on works agreements differ from a unilateral instruction. An internal AI usage policy tells staff what they may do; a works agreement establishes jointly settled conditions. A new software logo does not remove that distinction.

Step 1: inform the council while decisions remain open

The works council must still be able to influence planning.

Section 90 of the Act explicitly includes AI in work methods and workflows. It requires timely information, the necessary documents and consultation about effects on employees. This does not automatically amount to a consent right, but saying “we are not monitoring anyone” does not make it disappear.

Prepare one specific task for the first discussion, such as drafting copy from public product information. Include a fictional input and output. Describe which work disappears, which new review step is added and who owns it. Show the administration screens as well as the friendly chat interface.

Section 80 provides information rights tied to the council’s duties. Where assessing AI is necessary for those duties, bringing in an expert is deemed necessary; the provision still requires a more detailed agreement with the employer. Make room for unresolved technical questions.

Step 2: check what the system makes observable

Technical capability matters more than promised restraint.

The Federal Labour Court’s monitoring decision focuses on objective capability. An employer need not intend to monitor staff. Live technical monitoring can qualify without storing information. The case concerned headsets; it neither authorizes nor prohibits AI in general.

Apply that question to your configuration: can managers identify individuals, processing times, inputs or performance? Which reports can be enabled or exported? A technical error log is not automatically a performance report. Identification, informational value and available functions matter.

Where the relevant technical employee monitoring exists, section 87(1)(6) provides co-determination rights unless statutory or collective-bargaining rules govern the matter. For introduction and use subject to co-determination, obtain agreement with the competent works council before deployment, or a conciliation committee decision replacing that agreement. Promising not to look at available individual reports does not replace that process.

Three tools lead to different reviews

These examples distinguish functions rather than brands.

The fictional configurations below help with initial classification. They make no claims about the current features of a particular supplier. Other participation rights must also be considered for the actual decision.

• Text assistant: it processes public product information; the verified configuration offers the employer no identifiable usage reports or live views. Calling it AI does not by itself establish technical employee monitoring. Information and consultation about the workflow still need consideration.
• Service assistant: it links response times, error rates and conversation ratings to individual employees, letting the team leader compare them. This clearly calls for assessment and arrangements under section 87(1)(6), even if the stated goal is simply better customer answers.
• Recruitment assistant: it develops consistent criteria for shortlisting applicants. Selection guidelines under section 95 raise a separate issue. The absence of employee usage statistics does not settle that question.

Step 3: keep personnel rules on a separate track

Personnel decisions need their own review.

Personnel questionnaires and general appraisal principles require works council consent under section 94. Section 95 covers selection guidelines, including those developed with AI. Its threshold of more than 500 employees concerns the council’s right to demand that guidelines be established, not a general threshold for consent to existing guidelines.

For a small business, we would exclude applicant scoring, rankings and performance points from the first text-assistant trial. Otherwise “improving wording” can quietly become “comparing people”. Record a later change of purpose as a new decision even when the software stays the same.

Also check the exact scope of existing IT agreements. The Federal Labour Court addresses extensions of technical systems as possible participation issues too. Record the version, enabled features and connected services. An agreement titled “IT” does not prove that new AI analytics are already covered.

Use this discussion card to expose unanswered questions

The card gives the discussion a concrete starting point.

Copy these lines for a limited text assistant. They are suggestions for preparation, not a works agreement ready for signature. Replace empty fields with evidence; explicitly leave anything unknown unresolved.

• Purpose: drafts based on public product information. Excluded: staff appraisal, applicant selection and individual productivity rankings.
• Scope: [system, version, features, department]. Owner: [role]. Changes: present additional features before use.
• Data: fictional examples and public text. Employee identification in accounts, logs and reports: [specific fields and associations].
• Access: [who can see which content, reports or live views]. Retention and deletion: [deadline, process, evidence].
• Participation: record information and consultation; list potential co-determination issues separately; identify relevant provisions in existing agreements.
• Acceptance: demonstrate settings, check agreed boundaries and record training. Incident: [reporting route and responsible role].

Preparation costs €135 in this example

The initial cost is staff time.

Assume one responsible person spends 45 minutes preparing, three people each spend 30 minutes discussing the draft, and one person spends another 45 minutes documenting it. The calculation is 45 + 3 × 30 + 45 = 180 person-minutes, or three hours. At an assumed fully loaded internal cost of €45 an hour, that is €135.

This is a checked budgeting assumption, not a measured project duration or a lawyer’s fee. Negotiations, experts, software and technical changes may add costs. If individual-level reports remain unclear, we would first obtain the missing technical evidence rather than build an entire agreement on assumptions.

Use the AI literacy record for subsequent staff instruction. The participation card and the training record serve different purposes. Keep both with the approved workflow and appoint someone to reconcile them when it changes.

Three questions determine the next action

The next action is a documented feature review.

Do we need a works agreement without a works council? Without a competent works council, the other party to the agreement is missing. An internal usage policy can establish working rules but does not remove other obligations. The AI data-protection review remains a separate step.

Does a personal account avoid participation requirements? It creates no blanket exemption. The work-related use, technical access and other participation triggers still matter. Personal accounts therefore do not replace a properly governed purchasing process.

Can we test first and agree terms later? Calling something a trial does not settle participation rights. Describe the limited trial beforehand and resolve the applicable rights. Our AI consulting and implementation services can document the functions and data flows for that discussion. Fill in the card now, starting with the access permission that nobody can yet demonstrate.