# Works Agreements on AI: Typical Contents

Author: Marin Niehues, Niehues Consulting · Published: 2025-10-20 · Updated: 2026-11-16 · AI Law, Risk & Governance
Source: https://www.niehuesconsulting.com/blog/works-agreement-ai/

A works agreement on AI defines what AI systems are used for in the company, which data they process and how the works council stays involved in new applications. For management it is a delivery instrument. As long as purpose and responsibility are not agreed in writing, every rollout remains a one-off that is renegotiated for the next application.

**A framework agreement reduces the introduction of every further AI application from a negotiation process to an annex. The effort arises once and benefits every subsequent initiative.**

## Is a works agreement on AI mandatory?

German law does not require a works agreement, but it does require co-determination. Where a works council exists and the system is capable of monitoring employees' behavior or performance, the rollout needs its consent, and a works agreement is the usual route to it. Without consent or a ruling of the conciliation board, the system may not be introduced. The employer must in any case inform the works council in good time when planning AI.

For AI systems that the employer operates or licenses itself and whose logs it can view, the capability to monitor is almost always present, even if nobody plans an evaluation. What counts is the technical possibility. A labor court drew the lower boundary in 2024: where employees use a publicly available language model through private accounts in the browser, the employer receives no usage data and the co-determination right for technical systems does not apply. Once the company provides a corporate account, the position changes.

This article provides an overview and does not constitute legal advice. How the agreement fits into the full sequence of an introduction is set out in the article [From pilot to scale](/blog/from-pilot-to-scale-ai-operations/).

## What co-determination rights does the works council have on AI?

The works council has four central rights on AI under the German Works Constitution Act (BetrVG), and every works agreement must cover them:

- Section 87(1) no. 6 BetrVG: as soon as a system is objectively capable of monitoring employees' behavior or performance, the works council has a genuine co-determination right over its introduction and use.
- Section 90 BetrVG: the employer must inform the works council when planning working procedures, explicitly including the use of AI, and discuss the effects on employees with it.
- Section 95(2a) BetrVG: guidelines for personnel selection remain subject to consent even when AI drafts or applies them.
- Section 80(3) BetrVG: where the works council has to assess the introduction or use of AI, calling in an expert adviser is deemed necessary.

In addition, order in the establishment under section 87(1) no. 1 applies when the company sets binding usage rules, and vocational training under sections 96 to 98 applies when tasks change to the point where training becomes necessary. A good agreement assigns a building block to each of these rights, so both sides can see that nothing remains open.

The works council agrees the expert adviser's person and scope with the employer, who pays the costs. With several planned AI applications, a standing expert adviser, which the Act expressly provides for, reviews each new annex faster than an assessor starting afresh.

## Why a framework agreement delivers faster than case-by-case arrangements

In many companies the first works agreement on AI emerges under time pressure, and the same clarification repeats itself with the next tool. A framework agreement sets the ground rules once for an entire class of applications, and each individual application then needs only a short annex.

The decision follows clear criteria. If the company plans more than one AI application in the next twelve months, the framework agreement is the shorter route. If it remains a single, narrowly defined tool, an individual agreement suffices. An application that contributes to personnel decisions or restructures entire workflows receives its own, more detailed annex even under a framework.

## The typical contents of a works agreement on AI

Agreements that work in practice follow a similar structure. They are short enough to be read and specific enough to settle disputes in individual cases. These building blocks belong in them:

- Scope and definitions: which systems fall under the agreement, from chat assistants to functions embedded in existing software, and who counts as an employee.
- Purpose and purpose limitation: what the systems are used for and what they explicitly are not used for, for example individual performance appraisal.
- Data categories and storage: which data arise in requests and results, how long they are retained and who has access.
- Transparency: how employees learn where AI is involved and how results come about.
- Human final decision: in which processes a person reviews and takes responsibility for results.
- Training: which training takes place before first use.
- Expert adviser: whether a standing expert adviser is appointed and how he or she reviews new annexes.
- Procedure for new applications: how an annex is created, who reviews it and within what period the works council responds.
- Review and amendment: when the agreement is revised and how complaints are handled.

Two building blocks decide delivery capability. The procedure for new applications determines speed after signing, and purpose limitation determines the workforce's trust. An agreement that leaves both open is circumvented.

**What belongs in the annexes**
Product names and version numbers change faster than any agreement. They belong in annexes and in the internal policy. The works agreement regulates rights and procedures, the policy regulates daily behavior. The article [AI policy for employees](/blog/ai-policy-for-employees/) describes the content and introduction of such a policy.

## Is there a model works agreement on AI?

Model texts and analyses of published works agreements exist, but a template does not replace your own version. The trade-union-affiliated Hans Böckler Foundation analyses company agreements and publishes excerpts, for machine learning applications for instance in an analysis of almost 30 concluded works and service agreements, and trade unions offer practical guidance. These texts show which building blocks are customary and how other parties have worded them.

For your own agreement, a model is useful as an outline. The contents must come from your own applications and data flows. An adopted template that does not fit the company lengthens negotiations, because every clause is examined individually. A draft that names the planned applications concretely shortens them.

## How can negotiations be accelerated?

Negotiations become shorter when management clarifies three points internally beforehand and presents its own draft:

- Which applications are concretely planned for the next twelve months and which data they touch.
- Who in the company decides on new applications and is accountable for their results.
- Which evaluations the company rules out of its own accord.

The third point has the strongest effect. If management declares of its own accord that no individual performance monitoring with AI takes place, the largest point of dispute disappears before it arises. An orderly sequence follows: a joint meeting clarifies the target picture and planned applications, management presents a draft, and the works council and business units respond within an agreed period. With good preparation, you should plan two to four months to signature. How early involvement can be shaped is described in [Works council and AI introduction](/blog/works-council-and-ai-introduction/).

A pilot phase with a commitment has also proved its worth. The agreement initially applies to a limited group of users, and both sides agree in advance on the criteria under which regular operation follows. The negotiation on regular operation then rests on observations instead of apprehensions.

**Case-by-case arrangement and framework agreement**
Criterion | Case-by-case arrangement | Framework agreement | With Niehues Consulting
Time to the next application | New negotiation for each system | Short annex under a fixed procedure | Procedure and annex template built from the start
Clarity for employees | Inconsistent across units | One set of rules for all applications | Rules made understandable and anchored in enablement
Overview for management | Knowledge scattered across individual cases | Register of approved applications | Register integrated into the AI portfolio with metrics
Adaptation to change | Laborious, as each arrangement applies separately | Amendment through annexes | Regular review planned into operations

## What happens if employer and works council cannot agree?

If no agreement is reached, the conciliation board decides, and its ruling replaces the agreement between employer and works council. It consists of an equal number of members from each side and an impartial chair, appointed by the labor court in case of dispute. Until the ruling, the system may not be introduced, and the employer covers the costs of the proceedings. A permanent block is therefore ruled out; a delay of several months is the realistic risk. Companies that declare exclusions early and present a concrete draft rarely need the conciliation board.

## Operations after signing: responsibilities and review

The agreement takes effect only in operations. Someone must maintain the register of approved applications, a fixed rhythm for review must exist and a reporting channel for irregularities must be open. Many companies set up a small body of works council and business unit representatives for this, supplemented by data protection and IT. The principles for its work are described in [Building AI governance](/blog/ai-governance-framework/).

Three metrics show whether the agreement delivers:

- Lead time of an annex from draft to approval.
- Share of the applications in use that are listed in the register.
- Share of users with completed training.

The second value reveals shadow use. If it is low, employees use tools that nobody has reviewed, and the agreement runs past practice. The response is a faster procedure for annexes and better offerings, so that the approved tools are the first choice in daily work.

## Niehues Consulting as a partner for the rollout

[Niehues Consulting](/) supports companies in taking AI from prototype to regular operation and treats the works agreement as part of the rollout. In the Niehues Delivery Model with its phases Understand, Build, Introduce, the draft is created in parallel with the prototype. In the Understand phase, the review question is who decides, who delivers and who uses, which produces the overview of applications and data flows. In the Build phase, the model asks whether the prototype proves its value in the real workflow with real data. This shows what is actually evaluated and provides the basis for purpose limitation and data categories. In the Introduce phase, the question is who is accountable for operations from day one, meaning who maintains the register and prepares new annexes.

The legal drafting is done by the parties with their legal advisers. Niehues Consulting provides the preparatory work required for it: the overview of applications and data and the draft of the procedure for new applications, together with the metrics for operations. The entry point is the free initial consultation of 30 minutes via [contact](/#kontakt).

**Procedure first, applications second**
A works agreement on AI accelerates the rollout when it is built as a framework with annexes. Management clarifies applications and exclusions in advance, negotiates early with the works council and maintains a register of approved applications. Proceeding in this way replaces recurring negotiations with a procedure and saves time on every further application.

**Sources**
- German Federal Legislature: Betriebsverfassungsgesetz (BetrVG), statute, as amended 2021
- German Federal Legislature: Betriebsrätemodernisierungsgesetz, amending statute, 2021
- Arbeitsgericht Hamburg: decision 24 BVGa 1/24 on the use of ChatGPT, court ruling, 2024
- Thieltges, Andree (Hans-Böckler-Stiftung, I.M.U.): Machine-Learning-Anwendungen in der betrieblichen Praxis, Mitbestimmungspraxis No. 33, 2020

## Frequently asked questions

### Is a works agreement required for the use of AI?

No, the works agreement itself is not mandatory, but the works council's consent is, as soon as the system is capable of monitoring behavior or performance. The works agreement is the usual and legally secure route to that consent. Without a works council there is no co-determination, while data protection law continues to apply.

### Which sections of the BetrVG apply to AI?

The decisive provisions are section 87(1) no. 6 (co-determination on technical systems), section 90 (information during planning), section 95(2a) (selection guidelines) and section 80(3) (expert adviser). The last three mention AI explicitly. Section 87(1) no. 1 on usage rules and sections 96 to 98 on training also apply.

### Can the works council bring in an expert adviser on AI?

Yes, under section 80(3) BetrVG an expert adviser is deemed necessary when the works council has to assess the introduction or use of AI. The details are agreed with the employer, who pays the costs. Where several applications are planned, a standing expert adviser is possible.

### Is ChatGPT subject to co-determination?

ChatGPT is subject to co-determination as soon as the company provides a corporate account and usage data reach the employer. For use through private accounts in the browser, a labor court has denied co-determination on technical systems. Binding rules on use remain co-determined as a matter of order in the establishment.

### What happens if employer and works council cannot agree?

The conciliation board then decides, and its ruling replaces the missing agreement. Until then the system may not be introduced. The realistic risk is a delay of several months.

### Is there a model works agreement on AI?

Yes, analyses and model texts are published by the Hans Böckler Foundation and by trade unions, among others. They are useful as an outline for your own agreement. The contents must come from your own applications and data flows, otherwise the template lengthens the negotiation.

### What belongs in a works agreement on AI?

A works agreement on AI covers scope, purpose limitation, data categories, transparency, human final decision, training and a procedure for new applications. Technical details belong in annexes so that the agreement remains stable when products change.

### Can a works agreement also cover later AI applications?

Yes, if it is built as a framework with annexes. The framework regulates principles and procedures, and each new application receives an annex with purpose, data and responsibility. The agreement itself therefore does not have to be renegotiated for every new tool.
