Does a works council really have to approve an AI tool?

In Germany, the right is genuine co-determination rather than consultation. Section 87(1)(6) of the Works Constitution Act gives the works council a co-determination right over the introduction and use of technical devices designed to monitor the conduct or performance of employees, and the settled reading is that the capability to monitor is enough — the employer's intention not to use it that way does not remove the right. Without agreement, the employer cannot lawfully introduce the system, and the council can seek an injunction. In France the works council equivalent must be consulted and give an opinion before deployment, and in the Netherlands the works council holds a consent right over processing of personnel data and monitoring systems. The label differs; the sequencing does not.

Which AI tools trigger the requirement?

Far more than employers expect, because the test is technical capability, not purpose. An AI meeting assistant that attributes statements to named speakers, a coding assistant that reports per-developer acceptance rates, a customer support copilot that scores handled tickets, a sales tool that ranks reps by AI-derived activity, and a security product that models individual user behaviour all produce individually attributable performance data. So does a generic chat assistant deployed with per-user logging and an admin analytics dashboard. The question to ask about any tool is narrow: can an administrator see output attributable to one named employee? If yes, assume the right is triggered.

We are a US company with a small EU subsidiary. Does this apply to us?

Yes, where a works council exists at the local entity, and the size threshold is lower than most US employers assume — a German establishment with five permanent employees over eighteen can elect one, and once elected the rights attach regardless of how the parent runs global IT. A global rollout decided at headquarters and pushed through a single-tenant SaaS admin console is the classic failure mode: the tool is enabled for every seat in one action, the German seats are inside it, and the co-determination right was never exercised. Group-level IT decisions do not displace local employment law, and the local managing director carries the exposure.

What happens if we deploy first and consult afterwards?

Three consequences, in ascending order of cost. The works council can demand the system be switched off and seek interim relief to enforce that, which is disruptive precisely because the tool is already embedded in workflows. Data generated by an unlawfully introduced monitoring system is frequently held unusable in employment proceedings, meaning the performance evidence you gathered cannot support the dismissal it was gathered for. And separately, the processing may lack a lawful basis under GDPR as applied through national employment rules, which is a supervisory-authority matter with its own penalty regime. Retroactive consultation resolves none of the first two cleanly.

How does the EU AI Act interact with works council rights?

It adds a duty and does not replace the existing ones. Deployers of high-risk AI systems in an employment context must inform worker representatives and affected workers before putting the system into use, and employment and worker-management uses are expressly listed as high risk. That obligation sits on top of national co-determination law rather than harmonising it, so a German employer owes both the AI Act notice and the section 87 agreement. The practical effect is that the information a council would have had to extract through negotiation is now partly owed to it by statute, which tends to make the negotiation shorter, not longer.

What should a works council agreement for an AI tool contain?

The recurring clauses are: an exhaustive description of the data fields collected per user; an explicit prohibition on using the output for individual performance evaluation or disciplinary purposes, or defined limits if such use is agreed; retention periods with automatic deletion; a list of who may access individual-level data and under what approval; a rule that aggregated reporting must not be re-identifiable below a stated group size; vendor and subprocessor identification with a change-notification duty; and a review or sunset date tied to feature changes. The last clause is the one most often omitted, and it is why an agreement negotiated for a tool's 2025 feature set becomes contested the moment the vendor ships analytics in 2026.

Your EU Staff Can Veto the AI Rollout

Six Jurisdictions, Three Different Verbs

"Consult the works council" is a useful shorthand and a dangerous one, because the word hides the difference between an opinion the employer may disregard and a veto it cannot. Project planning that assumes the lighter version everywhere will discover the heavier one in the jurisdiction where the rollout is already live.

Jurisdiction

Body

Right

What it means in practice

Germany

Betriebsrat (works council)

Co-determination

Section 87(1)(6) BetrVG covers technical systems suited to monitoring conduct or performance. Capability alone triggers it; no agreement means no lawful introduction.

France

Comité social et économique

Consultation

Prior consultation and an opinion are required before introducing technology that changes working conditions or enables monitoring. Deployment before the opinion is a procedural defect.

Netherlands

Ondernemingsraad

Consent

Consent right over decisions on personnel data processing and systems monitoring presence, behaviour or performance. A decision taken without it can be nullified.

Austria

Betriebsrat

Co-determination

Works agreement required for control measures affecting human dignity; systems producing individual performance data are treated conservatively.

Belgium / Italy

Works council / union agreement

Consultation or agreement

Italy conditions remote-monitoring tools on a union agreement or labour inspectorate authorisation. Belgium requires information and consultation via collective instruments.

Ireland / Nordics

Information and consultation forum

Consultation

Lighter procedural burden, but GDPR Article 88 national rules and transparency duties still apply to employee monitoring.

Capability Is the Test, and Almost Everything Has It

Employers reach for the same defence every time: the tool is a productivity aid, nobody intends to measure anyone with it, and the analytics tab will simply go unused. That argument loses on the statute's own terms, because the right attaches to systems suited to monitoring rather than to systems used for it. An unused dashboard is still a dashboard.

The list of triggering tools is therefore much longer than the list anyone budgets for. Meeting assistants that attribute statements to speakers. Coding assistants reporting per-developer acceptance rates. Support copilots scoring resolved tickets. Security products modelling individual behaviour. A general chat assistant with per-seat logging and an admin analytics view. Each was procured as a tool and each arrives carrying an employment-law obligation nobody read into the order form.

The Sequence That Cannot Be Reordered

Feature inventory before the pilotLast clean moment

List every field the tool can attribute to a named user, including admin dashboards and export APIs. This document is what the council will negotiate against, and producing it late signals that it was assembled defensively.

Notify worker representativesStatutory notice

For high-risk employment uses, the EU AI Act requires informing worker representatives and affected workers before the system is put into use. This is a floor, not the negotiation.

Open co-determination where it existsHighest

In Germany and Austria this is where the tool can be stopped. The council's leverage is highest before rollout and effectively unlimited if rollout happened without it.

Negotiate the works agreementTimeline risk

Data fields, purpose limits, access control, retention, re-identification thresholds and a review trigger. Weeks to months is normal; a global launch date set by headquarters is not a reason it will be faster.

Configure to match the agreementImplementation gap

The commitments have to be enforced in the admin console, not in a policy document. An agreement promising no individual analytics alongside a dashboard that shows them is the fact pattern that ends the relationship.

Re-open on vendor feature changesDrift

SaaS tools gain analytics capability silently through release notes. An agreement scoped to the tool as it was is out of scope the day the vendor ships per-user scoring.

The Penalty Nobody Models: Evidence You Cannot Use

Fines get the attention, but the consequence that actually bites is evidentiary. Data produced by a monitoring system introduced without the required agreement is frequently excluded in employment proceedings. The concrete version: a company builds a performance case against an underperforming employee using output from an AI tool deployed without works council agreement, moves to dismiss, and discovers at the labour court that the entire evidentiary basis is inadmissible. The dismissal fails, the employee is reinstated with back pay, and the tool's business case has now produced a net loss.

That outcome does not require a regulator to notice anything, does not depend on a supervisory authority's enforcement priorities, and arrives on the employer's own timeline — at the exact moment it needed the data most.

Six Controls Worth Having

Ask the capability question, not the purpose question

Procurement asks what the tool is for. Co-determination asks what it can do. Rewrite the intake question to 'can an administrator see output attributable to one named employee?' and route every yes to employment counsel before the pilot, not before the renewal.

Give headquarters a European launch date, not a global one

The single most expensive assumption in cross-border AI rollout is that EU seats can go live with the rest. Build the consultation window into the project plan at kickoff; discovering it two weeks before launch means either an unlawful deployment or a visibly slipped date.

Negotiate the analytics off before you need them off

Per-user dashboards are almost always optional modules. Disabling them at procurement is a configuration choice. Disabling them after a council objection is a concession that the monitoring capability was live, and it starts a conversation about the data already collected.

Write the sunset clause the vendor will trigger

Tie the agreement to a described feature set and require renegotiation on material change. Without it, every vendor release note becomes an argument about whether the existing agreement still covers the product.

Separate aggregate reporting from individual access

Councils rarely object to team-level throughput; they object to per-person scores. Agreeing a minimum group size for any report, enforced in the tool, resolves most of the negotiation and preserves the metrics the business actually wanted.

Do not rely on employee consent

Consent in an employment relationship is treated as unlikely to be freely given, and building a rollout on it invites the whole basis to collapse. The workable route is a works agreement or a national employment-law basis, which is also what the council is asking to negotiate.

Your Own Marketing Describes the Monitoring

If you sell an AI tool into European employers, the works council will read your product pages before it reads your DPA. Phrases like "see who is adopting it", "per-seat usage insights" and "manager visibility" are the exhibits that decide whether your customer's rollout needs an agreement.

See what your public pages are claiming in one pass. Run a free scan and reconcile every monitoring claim against what you tell buyers in diligence.

This article is general information and not legal advice. Employee representation law differs substantially between EU member states in scope, thresholds, procedure and remedy, and the interaction with the EU AI Act is still being worked out in national practice. Consult qualified local employment counsel before deploying any AI system to employees in Europe.

Your EU Staff Can Veto the AI Rollout.