How to Get a Works Council to Yes

A works council approves a workforce AI tool when it can see, in writing, what the system infers, who reads individual answers, how a worker declines, and what happens to the data afterwards. The vendors who lose the meeting treat those questions as obstacles. The ones who win walk in with the answers as documents rather than promises, and the yes, done properly, is not a concession. It is the strongest reference in European enterprise sales. The council is not the wrong audience for your listening programme. It is the audience that asks the exact questions your leadership should have asked first.

One scene first, because it happens weekly somewhere in Germany. The HR director presents the new listening programme. The slides are good. Then a council member asks what the system detects in people’s voices, a vendor screenshot with an emotion radar goes up, and the project dies in the room, politely, with a request for further information that never quite concludes. Nothing illegal was proposed. The deployment simply never happens, which under co-determination is the same thing. If you own a rollout across Germany, France, the Netherlands or the Nordics, that meeting is already on your critical path whether or not it is on your plan. Most project plans mark it as a risk to manage. It is better understood as a design review you didn’t know you’d booked.

A quick honesty note before the steps. Consultation obligations differ sharply by country and sometimes by sector and site. What triggers formal codetermination for a German Betriebsrat is not identical to what a Dutch ondernemingsraad or a French CSE expects, and a European Works Council sits over the top once you cross 1,000 employees in the EEA with 150 in each of two member states. This piece stays at the pattern level on purpose. It is not legal advice, and your works council counsel should sign off on anything country-specific.

Why does the council hold the keys?

Because in much of Europe, employee representation over workplace technology is hard law with teeth, and the German version is the strongest. Section 87(1) No. 6 of the Works Constitution Act gives the works council a binding co-determination right over technical systems capable of monitoring the behaviour or performance of employees. Two words do all the work. Capable: the Federal Labour Court reads the provision on objective capability, so a system does not need to be intended as monitoring to trigger the right; a conversation platform that timestamps responses is capable, full stop. And binding: without agreement the employer cannot deploy, measures taken without proper involvement are invalid, and disputes go to a conciliation committee, not to a workaround.

The pattern repeats with local variation: Dutch councils hold consent rights over monitoring and personnel systems, French employers must inform and consult the CSE on new technology, and the AI Act’s Article 26 now adds an EU-wide floor: before a high-risk system goes into service in the workplace, deployers must inform affected workers and their representatives. The EU duty adds to the national ones rather than replacing them; in Germany you need both the AI Act paperwork and the council’s agreement, and the second is the one that can actually stop you.

Three more facts to hold before the meeting. Since the 2021 modernisation of the Works Constitution Act, German councils have an explicit right to bring in external AI expertise, so assume the person across the table has read more about the AI Act than your sales team has. The Hans Böckler Foundation found 68% of German works councils already report AI introduced in their organisations, so this is their operating environment, not a novelty. And while councils sit in a minority of German establishments overall, they concentrate in exactly the large employers where enterprise deployments happen, which is why the meeting finds you even though the average statistic suggests it might not.

One reframe before the five questions, because it changes the posture. The council’s list, what does it infer, who sees answers, were people told, where does the data go, can someone decline, is the same list a data protection authority asks, the same list the AI Act legislates, and the same list this series has argued every buyer should ask. Their instincts about emotion detection did not soften over the last three years; they became Article 5. The council is the exam the whole market is about to sit, arriving early, and a platform built on disclosure, restraint and evidence walks in already aligned.

What does the system infer about individuals?

The first question is almost never “does it collect data”. It is “what does it conclude”. A council wants the boundary between what a worker says and what the system decides about them, and it wants that boundary written down. The winning answer is a non-inference list: an explicit statement of what the tool does not derive, score, or flag at the individual level.

This is where most demos fall apart. The vendor talks about sentiment, then about flagging at-risk employees, then about surfacing trends, and the council correctly hears a system that turns a conversation into a judgement about a person. The moment inference is open-ended, the answer to “what does it infer” is “we don’t fully know”, and that is a no.

A platform that can win the room draws the line the other way. It says, in a document the council keeps: this tool does not produce individual performance scores, does not infer protected characteristics, does not build a per-worker risk profile, does not feed line managers a ranking. Whatever quality signal it produces about a response describes the response, not the person. A quality score on a session tells you whether an answer carried real signal or was given on autopilot, and that distinction lives at the level of the input, not the individual’s record.

The reasoning the council follows is simple. If the system cannot infer a judgement about a named person, then most of the surveillance concern collapses, because there is nothing personal to misuse downstream.

Who can see an individual’s answers?

Councils assume individual answers will eventually be read by a manager, because in most tools they can be. The answer that wins is an anonymity threshold enforced in the system, not promised in a policy. Below a set group size, individual responses are never shown, and the threshold is a setting the council can inspect, not a line in a privacy notice.

Here’s the distinction that matters. A policy says “we will only report in aggregate”. A threshold says the software will not render a breakdown for any group smaller than, say, five people, so a team of three cannot be resolved into three identifiable answers no matter who runs the query. The first is a commitment that can be quietly broken. The second is a property of the system.

When a council asks who sees individual answers and the honest reply is “the raw responses are anonymised at ingestion and the reporting layer refuses to slice below the threshold”, the conversation changes. You are no longer asking them to trust your intentions. You are showing them a mechanism that holds even if intentions change, even after the sponsor who promised good behaviour has left the company.

How does a worker decline without being noticed?

Consent is only real if declining is safe, and a council knows that a “voluntary” survey with a visible non-response is not voluntary at all. The answer that wins is genuine abstention: a worker can skip a question or the whole conversation, and that choice is indistinguishable, in the data, from someone who simply hasn’t got to it yet. No flag, no follow-up list, no manager notification that Person X opted out.

This is the point where completion-rate thinking does real harm. A platform built to maximise completion treats a decline as a gap to close, so it nudges, reminds, and eventually surfaces the non-responders. From the council’s seat, that is coercion with a progress bar. The pressure to finish is exactly what produces the answer given to make the screen go away, which is worthless data wearing a green tick.

A tool designed around abstention inverts that. Declining is a first-class outcome. It is logged as a choice, not a failure, and it never resolves to a named person a manager can lean on. When you can tell a council “a worker who says nothing faces no consequence the system can even generate”, you have removed the objection that sinks most voluntary programmes.

Where is the proof that people were told?

A council will ask what the tool disclosed to workers and when, and “we tell them it’s AI” is not an answer. The answer that wins is a record of disclosure the council can hold in its hand: proof that each participant was told they were speaking to an AI system, what it was for, and that they could decline, captured per session. Disclosure you cannot evidence is disclosure that, for audit purposes, did not happen.

The stakes on this question are no longer emerging; they are dated. From 2 August 2026, Article 50 of the AI Act requires AI systems that interact with people to identify themselves at first contact, in the language the conversation runs in, and a council reading that rule will want to know not that you intend to disclose but that you can show you did, for any given conversation, months later.

This is where the exportable record earns its place. A conversation-level export that captures the disclosure text shown, the abstention option offered, and the timestamp turns “we comply” into “here is the file”. The council does not have to believe you. It has to open the export.

What happens to the data, and for how long?

The last question is retention, and vagueness here reads as risk. The answer that wins is a written retention schedule: what is kept, in what form, for how long, and when it is deleted, with anonymised aggregates and identifiable raw data on separate clocks. A council rarely objects to data existing. It objects to data existing forever with no stated purpose.

The schedule should be specific enough to act on. Raw responses anonymised at ingestion. A stated retention period for the aggregate signal. A deletion point for anything session-level, with the deletion itself evidenced. When the council can see that identifiable material has a short, defined life and that the long-lived data is aggregate by construction, the fear of a slowly accumulating dossier goes away, because the dossier cannot form.

The meeting on one page

Every question has a losing answer, which is a promise, and a winning answer, which is a document. The difference decides the meeting, because in Germany the yes takes the form of a works agreement, and agreements can only be drafted from things that can be written down.

The question	The losing answer	The winning artefact
What does it infer about our people?	”The emotion features can be turned off”	The signed non-inference list, with a per-conversation attestation in every exported record
Who sees individual answers?	”Only authorised users”	The anonymity threshold as an inspectable setting, plus the access log inside each record
Can a worker decline safely?	”Participation is encouraged”	Abstention as a first-class outcome, indistinguishable from not-yet-answered
Were people told it was an AI?	”It’s in the launch comms”	The disclosure field: exact text, language, timestamp, per person, exportable
What happens to the data?	”We follow GDPR”	The retention schedule as per-record fields, with deletion evidence when it expires

Read the right-hand column as a set and notice what it is: the Compliance Conversation Record, published earlier in this series, doing its second job. It was designed as audit evidence. It turns out to be works council evidence, because a council and an auditor are asking the same question from different chairs: what happened to this person, and can you prove it?

The pack to bring to the meeting

Walk in with the documents printed, not screen-shared. A council remembers what it can annotate.

The conversation record export. A per-session file showing the disclosure shown, the abstention offered, the timestamp, and the anonymised response. This answers the disclosure and consent questions in one artefact, and it is the same record an auditor would ask for later, so you are building the audit file and the works council file at once.
The non-inference list. One page. What the system does not derive, score, flag, or rank at the individual level. Signed by someone who can be held to it. This answers the inference question and pre-empts half the follow-ups.
The retention schedule. A table of data categories against retention periods and deletion triggers, with identifiable and aggregate data on separate rows.
The intended-purpose statement. What the system is for and what it must not be used for, which is the instructions-for-use document the AI Act will require by December 2027 anyway. Publishing it to the council early converts a future obligation into present trust.
The disclosure text in every language the workforce actually speaks. This lands especially well where the council represents a multilingual floor, and it is the difference between claiming coverage and demonstrating it.
The drift clause. A written commitment that if the system’s use ever extends toward individual performance evaluation or any purpose beyond the stated one, the agreement reopens. This is the trust multiplier: it converts the council’s biggest fear, scope creep, into their own contractual trigger, offered before they ask.

Bring the anonymity threshold setting too, ideally visible in the tool itself, so “we can’t slice below five” is demonstrated rather than described. And bring the vendor into the room for the technical session, because a vendor who will not face a works council is answering a question the council has not asked yet.

What a works council should still push back on

A good council will not stop at the five questions, and you should not want it to. The honest section of this pitch is naming the uses the tool should not be put to, before they ask. The strongest position in the room is the vendor or programme owner who lists the guardrails unprompted.

Push back is warranted on scope creep. A listening tool justified as employee voice should not quietly become a management-by-exception dashboard, where the aggregate is fine but someone starts asking for the outliers. It should not be used to build a case against a named individual, to monitor union activity, or to substitute for the consultation the council is legally owed. Any inference about emotional or mental state is prohibited territory in the workplace regardless of what anyone agrees. And a quality score on a response, which exists to tell you whether an answer carried signal, must never be repurposed into a judgement about the worker who gave it. That is the one line that, if crossed, turns the whole design against its purpose.

Demands worth making beyond the misuse list: the AI Act classification memo, because “we assume it is not high-risk” deserves the follow-up “show me the reasoning”; a review clause with a date on it, not just a trigger; and an answer to the question that outranks all the technical ones: what will the organisation do with what it hears? A channel that gathers candour and changes nothing burns the trust of exactly the people the council represents, and a council is right to treat the action plan as part of the deployment, not an aftermath.

Say all of this out loud. The council’s job is to assume the tool will be misused eventually, because tools outlive the people who introduce them. When you name the misuses and show the mechanisms that prevent them, you are speaking their language instead of asking them to lower their guard.

Why this turns representatives into references

There is a strategic point underneath the tactics. Research on why works councils sometimes help and sometimes hinder keeps landing on the same prerequisites: trustful cooperation, unhindered information flow, and voluntary agreement rather than imposition. A programme that answers the five questions with documents is doing exactly that: giving the council information, cooperating on design, and letting them agree rather than acquiesce.

The council that helped shape the guardrails tends to defend the rollout to the people it represents, and it will say so to the next site, the next country, the next council in the group. Works agreements travel between councils faster than marketing travels between buyers. A works agreement negotiated in the open, with the non-inference list attached and council observers in the pilot, is proof of a kind no case study reaches: the people whose job is scepticism examined the system and signed. A vendor with two or three such deployments holds a position in codetermined Europe that a US-first competitor with an emotion radar in its demo cannot enter at any price. The meeting most vendors dread is, handled honestly, the moat.

If you own the rollout, do this before the meeting

Draft the non-inference list yourself first, then ask the vendor to sign it. If they hesitate, you have found the problem before the council does.
Ask to see the anonymity threshold as a setting, not a policy line. If it lives only in a privacy notice, it is a promise, not a control.
Export one real conversation record and read it as the council will. If it doesn’t show disclosure, abstention and a timestamp, it won’t answer the disclosure question.
Write the retention schedule as a table and put identifiable and aggregate data on separate clocks before anyone asks.
List the uses you will refuse, add the drift clause, and bring both to the meeting unprompted.
Book the council before the launch date exists. Consultation beats announcement everywhere, but in codetermined Europe it is also simply the sequence the law expects.

Frequently asked questions

Does a works council have a legal veto over a workforce AI tool?

It depends on the country and the specific use. In Germany, a system objectively capable of monitoring behaviour or performance, and a timestamped conversation platform is, triggers binding co-determination under Section 87(1) No. 6, and deployment without agreement is invalid. Other jurisdictions have consent or consultation rights that are strong but stop short of a veto, and the AI Act adds an EU-wide duty to inform workers and representatives before high-risk deployment. Treat every deployment as requiring genuine agreement and get country-specific counsel rather than assuming your home-market rules travel.

We already deployed without consulting. How bad is it?

In Germany, potentially very: measures introduced without proper works council involvement are invalid, and the council can demand the system stop. The recoverable path is the honest one: bring the full pack, acknowledge the sequence error, and negotiate the works agreement now rather than defending the fait accompli. The pack matters more in this scenario, not less, because trust arrives already spent.

What is the difference between a quality score and inferring something about an employee?

A quality score describes a single response, telling you whether an answer carried real signal or was given on autopilot to dismiss the prompt. Inference builds a conclusion about the person, such as a performance rating or a risk flag. The first improves the trustworthiness of the data. The second creates exactly the individual judgement a works council exists to guard against, which is why the two must stay separate.

Why not just promise anonymity in the privacy notice?

A promise can be broken quietly, usually long after the person who made it has moved on. An anonymity threshold enforced in the software refuses to render any group below a set size, so individual answers cannot be resolved regardless of who runs the query. Councils trust mechanisms over commitments because mechanisms outlast intentions.

What should we never use an employee listening tool for?

Building a case against a named individual, monitoring union activity, ranking people, inferring emotional or mental state, or substituting for the consultation a representative body is legally owed. A tool introduced as employee voice should not become a management-by-exception dashboard that surfaces outliers by name. Naming these limits before the council does is the fastest way to earn its confidence.

Should the vendor be in the room?

Yes, for the technical session. A council interrogating a system deserves the people who built it, relayed answers lose precision in both directions, and a vendor’s willingness to sit across from employee representatives is itself evidence about the product. Any vendor who resists that meeting has told you something about what the demo dashboard shows.

Sources

Betriebsverfassungsgesetz, Section 87(1) No. 6, Section 80(3) and Section 95(2a), on co-determination over technical monitoring systems and works council rights regarding AI, including the 2021 Works Council Modernisation Act amendments.
GermanLaw International, Artificial intelligence and employee co-determination, on the binding nature of Section 87 rights and the invalidity of measures taken without proper involvement.
Hans Böckler Foundation, 2024, on the 68% of German works councils reporting AI systems introduced in their organisations.
European Works Councils, European Commission, thresholds and process for establishing an EWC.
Do works councils raise or lower firm productivity?, IZA World of Labor, on the prerequisites for cooperative works council relationships.
Proliferation of works councils and the desire for interest representation, German Economic Institute (IW), on works council prevalence and concentration in larger establishments.
Regulation (EU) 2024/1689, the EU AI Act: Article 26(7) on informing workers and their representatives before high-risk deployment, Article 50 on disclosure, and Article 5 on prohibited workplace emotion inference.
The Compliance Conversation Record: A Proposed Specification (v0.1), earlier in this series, on the evidence artefacts referenced throughout.