AI Tools Need the Works Council: What Germany's New Co-Determination Practice Means for Recruiting Software



A TA leader has run the business case, picked the vendor, survived the demo. The next step, as far as they're concerned, is IT sign-off – maybe a glance from legal at the AI Act.
What's almost always missing from that sequence: the works council – Germany's elected employee representation body with statutory co-determination rights. Not as a box to tick at the end, but as a body that can stop the rollout – retroactively, via emergency injunction, with a ban on using any data the system has produced up to that point.
That's no longer a theoretical threat. It's the practical consequence of a line of case law that has shifted noticeably since 2024 – and in 2026 it's converging from three directions at once: a stricter Federal Labor Court (BAG) reading of Section 87(1) No. 6 of the Works Constitution Act (Betriebsverfassungsgesetz, BetrVG), the AI Act's transparency obligations that took effect on 2 August 2026, and the regular works council elections between March and May 2026, which in many companies are putting more technically literate councils in office.
The Line That Moved
The penny usually drops at Section 87(1) No. 6 BetrVG: introducing and using technical equipment designed to monitor employee behavior or performance is subject to mandatory co-determination – not merely a duty to inform. For a long time, the decisive question was how narrowly to read "designed to."

One example of the old, narrower line comes from the Hamburg Labor Court (Case No. 24 BVGa 1/24) in a dispute over a ChatGPT rollout: because the company itself had no way to intervene in, control, or access the tool, the court denied that it was a monitoring device within the meaning of the statute – and classified the AI use as unregulated "work conduct" rather than co-determined "order-related conduct." Anyone relying on that logic in 2026 is arguing against the wind: recent BAG rulings increasingly focus on the objective capability to monitor, not on the employer's intent or on whether the data is actually analysed.
If a system is technically capable of logging usage behaviour, co-determination applies – regardless of whether that capability is ever used.
For recruiting software, that's not a borderline case. CV parsing with ranking, scoring models, interview transcription, even a simple usage log of who opened which profile or overrode which suggestion and when – all of it meets the new, broad reading almost automatically.
A careers-page chatbot that stores conversation histories is covered. An interview assistant that logs talk time or response latency is covered. A scheduling tool that tracks how quickly a recruiter responds to a request is covered. Even a plain matching tool with no visible dashboard is covered, the moment it logs in the background which suggestions were accepted and which were overridden.
The bar today is so low that the genuinely interesting question is no longer "does Section 87 No. 6 apply", but "how quickly can we get a workable agreement in place."
Whose Conduct Is Actually Being Monitored
This is where TA teams most often trip up: Section 87(1) No. 6 BetrVG doesn't protect candidates – it protects the company's own employees, typically the recruiters operating the tool. How they use the software, their response times, how often they override the score: that's the data Section 87 No. 6 is actually about.
Candidates enter the picture through a different door: Section 95 BetrVG gives the works council co-determination over selection guidelines for hiring, and an AI ranking that systematically favours certain profiles counts as such a guideline – even if nobody called it that when it was configured in the tool. Two separate claims, two separate subjects of negotiation, one and the same tool.
Three Other Levers That Rarely Get Considered
Beyond Section 87 No. 6 and Section 95 BetrVG, it's worth looking at three provisions that get cited less often but are just as effective:
Section 90(1) No. 3 BetrVG requires timely notice of new working methods being introduced – before the decision, not after. On its own that isn't a co-determination right, but breaching it makes every later negotiation harder, because the works council can then rightly point to a fait accompli.
Section 80(2) and (3) BetrVG gives the council a right to information about the entire AI inventory in the company, plus the right to bring in outside experts – at the employer's expense. Anyone hoping a tool can stay "small" because nobody on the council understands the technology is increasingly mistaken: the 2026 works council elections have, in many companies, brought in candidates with noticeably stronger IT and data-protection expertise.
Section 94 BetrVG applies as soon as an AI system defines assessment criteria or personnel-questionnaire-style attributes – for instance, when an interview tool automatically fills in competency profiles.
What Happens If You Bypass the Works Council
The consequences aren't symbolic. If the employer breaches an enforceable co-determination right, the works council can obtain an immediate shutdown of the system through preliminary injunction proceedings – often taking just two to six weeks. On top of that, a usage ban looms: results from an AI introduced without proper process – scores, rankings, automated rejections – become unusable under employment law. A rejection based on such a score stands on shaky ground if challenged. Add to that possible GDPR fines, if the underlying
data processing ran without a solid legal basis for lack of an effective works agreement – and, as discussed in detail in an earlier post here, potentially personal liability for management, because bypassed co-determination is exactly the kind of governance gap that courts treat as a failure to provide "adequate information."

The Roadmap to a Works Agreement
With a cooperative approach, an AI works agreement isn't a year-long project. The realistic sequence: first, mapping the AI inventory via a Section 80(2) information request, then assigning each tool to the relevant co-determination provisions, an outside expert if needed, followed by a draft and typically three rounds of negotiation. With consensus, four to eight weeks is realistic. If agreement fails, the conciliation board (Einigungsstelle) decides – that costs two to four additional months, but it does force a result either way.
In terms of content, a standard core has emerged for AI works agreements, confirmed independently by guides like Bitkom's and the HUMAINE toolkit: a clearly defined scope, a closed list of purposes instead of a blanket approval, a positive list of approved tools with a process for adding new ones, clear rules on human oversight (no fully automated personnel decisions), training obligations, a complaints channel, and a term with a review clause – two years is typical, because the tools change faster than the negotiation rhythm.

The 2026 Works Council Elections as an Accelerator
It's no coincidence that 2026, of all years, is becoming the turning point. The regular works council elections, held between 1 March and 31 May 2026, land almost exactly alongside the AI Act's transparency obligations taking effect on 2 August 2026.
Newly elected councils are stepping straight into a term where AI topics aren't a footnote but a core agenda item – and, as several trade articles on the elections independently note, they're bringing in noticeably more candidates with technical and data-protection backgrounds than previous cohorts. A works council that can assess recruiting software on its technical merits negotiates differently than one that has to rely entirely on outside advice.
And in Switzerland?
Anyone rolling out a tool from Switzerland for a team in Germany should know the difference rather than underestimate it. Switzerland has no equivalent to Section 87(1) No. 6 BetrVG. The Swiss Participation Act (Mitwirkungsgesetz) provides for an employee commission once a company has 50 or more staff, if at least a fifth of the workforce requests one – but its rights are considerably leaner: a right to information on the economic situation, a right to consultation on major decisions such as mass layoffs, and genuine co-decision only in narrow areas like occupational safety.
That means there's no enforceable consent requirement for rolling out AI recruiting software, unlike in Germany. That doesn't mean Swiss participation is toothless – but it's consultation, not negotiating power with a veto. Anyone rolling out tools across both countries would do well to treat the German process as the stricter benchmark, and not assume the Swiss rollout is automatically simpler the moment employee or candidate data from the EU is processed alongside it.
What This Means for Rolling Out a Tool
The practical takeaway is unspectacular but effective: bring the works council in before choosing a vendor, not after. Anyone who only informs them once the contract is ready to sign is negotiating from the weakest possible position and risking exactly the delay they were trying to avoid. Bring them in early, and you often get the opposite: a council that thinks through technical questions because it actually understands them – and a works agreement that, if push comes to shove, is your best defence rather than your biggest risk.
Sources
Skill-Sprinters: "BAG 2026 – Betriebsrat hat Mitbestimmung bei praktisch jeder KI"
ad-hoc-news: "Betriebsräte 2026: KI in der Personalarbeit erfordert Mitbestimmung"
ad-hoc-news: "Betriebsratswahlen 2026: Mitbestimmung im KI-Zeitalter"
Skill-Sprinters: "KI-Betriebsvereinbarung 2026: Muster, Mitbestimmung, Pflichten"
Bitkom: "Künstliche Intelligenz und Mitbestimmung" (Leitfaden 2026)
HUMAINE: Muster-Betriebsvereinbarung für den Einsatz von künstlicher Intelligenz
betriebsrat.de: "Still, aber wirksam: Arbeitnehmervertretung auf Schweizer Art"
Transparency Note on the Use of AI
AI-based tools were used in a supporting capacity in the creation of this article. They serve in particular for research and structuring of information, linguistic and grammatical revision, translation, and in part for the creation or editing of illustrations and image material.
The concept, substantive statements, assessments and conclusions come from the author and reflect his personal thoughts, professional experience and convictions. All content is editorially reviewed before publication and published under the author's responsibility.
Articles on this website are not created or published automatically by AI. AI is used as a supporting tool; responsibility for content and editorial decisions remains entirely with the author.


Comments