The union question, the works council, and consultation done right
The short answer
Involving the workforce is a ladder, not a step
Information, consultation, and co-determination (consent) are three legally distinct rungs, and an AI decision can sit on any of them depending on the system and the country. The first analytic move, every time, is which rung, because a consent-level monitoring tool treated as a mere information notice is an unlawful deployment a court can switch off.
What you will be able to do
- Distinguish the three legal levels of employee involvement (information, consultation, and co-determination or consent), and analyze which one a given AI decision triggers, because treating a consent-level duty as if it were mere information is unlawful and treating everything as consent stalls you needlessly.
- Analyze which duty you owe and to whom, mapping each AI system and workforce change on your workforce map to the representative body and the jurisdiction that governs it, so you never route an AI decision around a body that had a legal right to be consulted (see Topic 9.1).
- Determine when the duty triggers, recognizing the events that turn an AI plan into a consultable one (introducing a new technology, monitoring behavior or performance, a substantial change to work organization, deploying a high-risk system in the workplace), and consult before you act rather than after.
- Run consultation that is genuine, not theater, sharing real information in time for a real study, listening to the representatives' opinion, and being willing to change the plan, distinguishing that from a rubber-stamp meeting held after the decision is already made.
- Build a consultation plan and record, the dated artifact that names who must be involved at what level, when, what you shared, what they raised, and what changed, so the involvement is provable and feeds your evidence annex and dossier (see Topic 10.6) (see Topic 13.1).
- Read collective resistance as information and leverage, connecting the union question to the discipline of Topic 9.3, so a workforce's organized pushback improves the deployment rather than being fought as an obstacle (see Topic 9.3).
- Defend the consultation against the charge that it was a box-ticking exercise done around the people it affected, proving with the record that the representatives were genuinely involved before the decision and that their input could and did shape it.
The ideas, one by one
Consult before you decide, or you have not consulted
Genuine consultation happens before implementation, at a point when the plan can still change, with real information shared in time to study, a genuine hearing of the representatives' opinion, and openness to changing course. A meeting held after the decision is notification in consultation's clothes, and French courts have suspended live AI for exactly that.
The duty is local; do not export a home-country assumption
The same AI system can owe a German co-determination, a Dutch consent, a French CSE consultation, and a US bargaining duty at once. "We will just inform them" breaches continental law; "there is no works council" misses continental duties; a works council is not universal. Analyze each population against its own law.
Monitoring triggers the strongest rights, on capability not intent
A system suitable for monitoring behavior or performance engages co-determination in Germany and consent in the Netherlands, and the trigger fires on what the system can do, not on your promise about how you will use it. If it can monitor, the right is engaged; the capability line, not your intent, is what a court reads.
The AI Act's worker notice is a floor, not the whole duty
Article 26(7) requires informing workers and their representatives before a high-risk system goes live, but it sits on top of national consultation and co-determination rights, not in place of them. Discharging Article 26(7) does not discharge a German or Dutch consent duty.
Consultation done right is a source, not a cost
The works council's expert and the union know the work in ways the deploying team does not; they find the flag that punishes the wrong behavior and the accountability gap the internal design missed. SAG-AFTRA's performers produced the workable mechanisms the companies could not. Genuine consultation is the cheapest, best-informed critique your deployment will get (see Topic 9.3).
Write the record as you go, and build it for a hostile reader
A consultation you cannot prove is one a court and an auditor treat as never having happened. The dated consultation plan and record must show what was shared, when, and what changed, and it must survive a works council lawyer reading it for the tell of a pre-made decision, because that reader is coming (see Topic 10.6) (see Topic 13.1).
The strongest jurisdiction sets the pace for a uniform system
If one AI system must be identical everywhere, the consent-level countries gate the timeline. Planning the rollout around the thin-duty countries and colliding later with the veto countries just moves the delay from your schedule into a courtroom, which is the more expensive place to have it.
The union question is the module's law made collective
Resistance is information (see Topic 9.3); organized resistance is information with power behind it. Consultation done right is how you get that information early and cheaply and turn it into a deployment the workforce helped shape, instead of getting it late and expensively as a strike or a suspension order.
Run two strands on every worker-facing system
A monitoring or scoring AI system engages a data-protection duty (a data protection impact assessment, individual rights over automated decisions) alongside the worker-involvement duty. They are separate obligations to separate holders under separate law; discharge and record both, because completing one never discharges the other (see Topic 10.4).
The analyst's job is a decomposition, not a single answer
Separate the populations, lead with capability not label, place each duty on the right rung with a stated reason, fix the timeline to the strongest duty, and build the record for a hostile reader. A one-line "get it live" instruction becomes a per-jurisdiction, per-rung, per-trigger analysis with a defensible record, which is exactly what lets you deploy among represented people without being stopped by a court or a strike.
Listen to it
The same lesson, talked through by two hosts as episode 72 of the podcast.
Real cases
These are documented cases used to illustrate consultation, not to predict your organization. Each is cited and used for a specific point.
Example 1: The SAG-AFTRA video game strike over AI (this topic's anchor). From 26 July 2024 to 9 July 2025, SAG-AFTRA struck eleven major video game companies over the use of AI to create digital replicas of performers' voices and likenesses, ending when members ratified a new Interactive Media Agreement by 95.04 percent that requires informed consent and disclosure for digital-replica use, pays session fees for AI-generated performance rather than allowing an unlimited buyout, and lets performers suspend consent for new AI material during a strike (SAG-AFTRA, "SAG-AFTRA Strikes Video Games Over A.I.," and the 2025 Interactive Media Agreement). The value for this topic is precise: it shows a workforce with collective power forcing the AI question to the table and producing specific, enforceable, workable mechanisms that a purely internal design would never have generated. It is consultation not as a courtesy but as the thing that made the technology's use legitimate and durable, and it is the standing rebuke to any plan to route AI around the people it affects.
Example 2: A French court suspends an AI deployment for skipping consultation. In 2025, French courts applied Article L2312-8 of the Code du travail to AI, holding that deploying an AI tool that can affect working conditions is the "introduction of a new technology" requiring prior information and consultation of the CSE, and suspending a rollout that had begun without it; the courts held that even an experimental or pilot deployment does not escape the duty (Code du travail Article L2312-8; French rulings reported by CMS and French unions, 2025). The point for a governance lead: "we were just piloting it" is not a defense, and a court can order your live AI system switched off until you consult properly, which is the concrete cost of consultation-as-theater.
Example 3: German co-determination over monitoring AI. Under the German Works Constitution Act, the works council holds a genuine co-determination right (a veto) over the introduction and use of technical equipment suitable for monitoring employee behavior or performance (BetrVG Section 87(1) No. 6), and the 2021 Works Council Modernization Act added explicit information-and-consultation rights at the planning stage of AI (Section 90) and a presumption that the council may hire an external AI expert at the employer's cost (Section 80(3)) (BetrVG; established). The illustrative point: in Germany an AI system that can monitor workers is not something you consult about and then decide; it is something you cannot deploy at all without the works council's agreement, which places it on the top rung of the ladder and reshapes the whole rollout timeline.
Example 4: The capability line, drawn by a Hamburg court. A Hamburg labor court held in early 2024 that permitting employees to use a chatbot via their own private accounts did not trigger the monitoring co-determination right, because the employer had no access to usage data and thus no monitoring capability, while making clear that company accounts or deployment through company systems, where use can be tracked, would trigger it (Hamburg Labour Court, reported by Bird and Bird, 2024; established as reported). The point: the co-determination trigger is real monitoring capability, not the mere presence of AI, which is exactly the kind of fact a consultation disclosure must state honestly.
Example 5: The Dutch consent right over personnel and monitoring systems. Under the Dutch Works Councils Act, the works council has an instemmingsrecht (consent right) over decisions to adopt, change, or withdraw personnel systems that process employee data or that can register or monitor performance or behavior (WOR Article 27), alongside an adviesrecht (advice right) on major organizational decisions (WOR Article 25) (WOR; established). The point: the Netherlands, like Germany, can place an AI personnel or monitoring tool on the consent rung, so a pan-European rollout faces more than one veto jurisdiction, not just one.
Example 6: The EU AI Act's own worker-information hook. The EU AI Act requires that before an employer puts a high-risk AI system into service or use at the workplace, it must inform the affected workers and their representatives that they will be subject to it, in accordance with national worker-information rules (EU AI Act, Regulation (EU) 2024/1689, Article 26(7); established). The point: the AI Act adds a specific, floor-level information duty on top of national consultation and co-determination rights; it does not replace them, and treating the Article 26(7) notice as the whole of your obligation would breach the stronger national duties that sit above it. As of 2026 this specific AI Act duty is itself deferred, to 2 December 2027 for stand-alone high-risk systems under the Digital Omnibus timeline (Council of the EU, 29 June 2026), but the underlying lesson is not: the national duties Article 26(7) sits on top of are not deferred and already apply today.
Example 7: The EU information-and-consultation floor. Directive 2002/14/EC of 11 March 2002 establishes a general framework requiring, in undertakings above a size threshold, information and consultation on decisions likely to lead to substantial changes in work organization or contractual relations, with information given in time for an adequate study and consultation conducted with a view to reaching agreement (Directive 2002/14/EC; established). The point: even where no strong national works council exists, an EU baseline consultation duty can still catch a substantial AI-driven change to how work is organized, so "there is no works council here" does not automatically mean "there is no duty."
Example 8: The soft-law expectation of stakeholder involvement. Beyond hard law, the OECD AI Principles (updated 2024) call for responsible stakeholder engagement across the AI lifecycle, and international labor institutions treat consultation as central to fair technological transition (OECD AI Principles, 2024; established as principles, not enforceable law). The point: even in jurisdictions with thin statutory duties, the governance standard the world is converging on expects genuine involvement of affected workers, so an organization aiming at the elite standard consults because it is right and durable, not only where a statute forces it.
Example 9: The German planning-stage and external-expert rights, expanded for AI. The 2021 Works Council Modernization Act (Betriebsraetemodernisierungsgesetz) amended the German Works Constitution Act to make explicit that the works council's information-and-consultation right at the planning stage of technology deployment covers AI (BetrVG Section 90), and reinforced that the council may bring in an external expert to assess AI, presumed necessary, at the employer's cost (BetrVG Section 80(3)) (established). The point for a governance lead: in Germany the duty attaches early, at planning, not at go-live, so a system quietly designed to completion and only then shown to the works council has already skipped the stage where the council's input was owed; and the disclosure you give will be read by a specialist the workforce hired on your budget, which rewards honesty and punishes marketing gloss.
Example 10: The distinction between advice and consent in the Dutch model. The Dutch Works Councils Act separates two rights that are easy to confuse: the advice right (adviesrecht, Article 25) over major organizational or financial decisions, where the works council must be asked for advice in time to influence the decision and can appeal to a court if it is ignored, and the consent right (instemmingsrecht, Article 27) over personnel systems including those that process employee data or monitor performance, where the council's agreement is required (WOR Articles 25 and 27; established). The point: even within one country the rung differs by what the AI system is and does, so an AI tool that is both a major organizational change and a monitoring personnel system can engage both an advice right and a consent right at once, and analyzing "which right, over which aspect" is part of the Analyze skill, not a detail to gloss.
Where people go wrong
- Treating all worker involvement as one thing. The deepest error. Information, consultation, and co-determination are three legally distinct rungs, and an AI decision can sit on any of them depending on the system and the country. Treating a consent-level monitoring tool in Germany as if a quick "we informed them" satisfies the duty is unlawful and can get the system ordered off. Analyze which rung first, every time.
- Consulting after the decision. A "consultation" held once the system is bought, built, and scheduled to go live is not consultation; it is notification. Genuine consultation happens before implementation, at a point when the plan can still change. French courts have suspended live AI deployments precisely because the employer moved before consulting (Code du travail Article L2312-8; 2025 rulings). Timing is not a formality; it is the essence.
- Sharing too little, too late to study. Handing representatives a thin summary at the meeting itself defeats consultation, because they cannot study what they were not given in time (Directive 2002/14/EC, Article 4(3)). Real consultation requires real information early enough for an adequate study, especially where the workforce can bring its own expert to check it (BetrVG Section 80(3)).
- Assuming your home country's rules are the world's. An American lead who assumes "we will just inform them" breaches German and Dutch consent rights; a British lead who assumes "there is no works council" misses continental duties; a German lead who assumes a works council exists everywhere is surprised by the US union model. The duty is irreducibly local; analyze each population against its own law.
- Believing the AI Act's Article 26(7) notice is the whole duty. Article 26(7) is a floor-level information duty for high-risk systems, sitting on top of national consultation and co-determination rights, not replacing them. Telling workers they will be subject to the system does not discharge a German co-determination duty or a Dutch consent duty. The AI Act adds to the national duties; it does not absorb them.
- Mistaking capability for intent on monitoring. "We promise we will not use it to monitor" does not avoid the co-determination trigger, which fires on a system's suitability for monitoring behavior or performance, not on your stated intent (BetrVG Section 87(1) No. 6; the Hamburg 2024 capability line). If the system can monitor, the right is engaged; analyze capability, not promises.
- Running consultation as theater with the decision pre-made. Entering the room having already decided everything, receiving the representatives' opinion as a formality, and answering it with why-you-will-not-change is a staged process the law and the workforce both see through. Consultation must be conducted with a genuine view to agreement and a real openness to changing the plan, or it fails in court and in trust (see Topic 9.6).
- Treating representatives as an obstacle rather than a source. The works council's expert who finds the flag that punishes the wrong behavior, the union that names the accountability gap, are giving you the exact information Topic 9.3 taught you to prize (see Topic 9.3). Fighting them wastes the cheapest, best-informed critique your deployment will ever get. SAG-AFTRA's performers produced the workable mechanisms the companies could not; consultation is a source, not a cost.
- Forgetting the non-council routes to a duty. Even without a strong works council, the EU baseline (Directive 2002/14/EC) can require consultation on a substantial change to work organization, a union can hold a bargaining right, and data-protection law can require involvement over monitoring. "No works council" is not "no duty"; check the other routes.
- Not writing it down as you go. A consultation you cannot prove is one a court and an auditor treat as not having happened. Without a dated consultation record showing what was shared, when, and what changed, you cannot defend against the charge that you moved around the workforce, and your Module 13 dossier will show a documented system with a legal hole where its consultation should be (see Topic 13.1).
- Letting the fastest jurisdiction set the pace for a uniform system. If one AI system must be identical everywhere, the consent-level jurisdictions (Germany, the Netherlands) gate the timeline, not the easiest one. Planning the rollout around the thin-duty countries and colliding later with the veto countries just relocates the delay from your schedule to a courtroom.
- Confusing consultation with a rubber stamp because it is "not a veto." Consultation not being a veto (you can still decide after genuine consultation) does not make it optional or cosmetic. The obligations to share, to hear, to respond, and to be genuinely open remain binding, and skipping them is unlawful even though the final decision is yours. "Not a veto" is not "not required."
- Treating the data-protection duty as the worker-involvement duty. A monitoring or scoring AI system owes both a worker-involvement duty (to a works council or union) and a data-protection duty (to each affected worker). They come from different law and are owed to different holders, so a completed data protection impact assessment does not discharge a co-determination right, and a completed consultation does not discharge individual data-protection rights. Run and record both strands (see Topic 10.4).
- Collapsing an advice right and a consent right into one. In regimes like the Dutch model, a single AI decision can engage both an advice right over a major change and a consent right over a monitoring system. Treating the consent right as if it were only advice (proceed after asking) is as unlawful as treating consent as information; honor each right for the aspect of the decision it governs.
- Mapping only the national bodies on a multi-country EU rollout. A single AI system deployed across employees in more than one EU member state above the size thresholds can owe a transnational European Works Council consultation (Directive 2009/38/EC) in addition to each country's national works council duty. A plan that lists the German, Dutch, and French bodies and stops there has consulted around the EWC layer, not through it.
- Waiting for "deployment" to start the clock. The duty in strong regimes attaches at the planning stage, not at go-live: Germany's BetrVG Section 90 gives the works council information and consultation rights when technology is being planned, which can mean before a vendor contract is signed. A specification locked into a procurement contract before the works council was consulted is already a breach, even though nothing has "deployed" yet.
- Treating consultation as a one-time event with no re-trigger. A significant retraining, a scope expansion, or a vendor patch that adds new monitoring capability to a system already in use can re-trigger the same duties that applied at first launch, because the change, not the calendar, is what the law watches. A consultation plan and record that only covers the initial go-live, with no note of what would re-open it, will accumulate unnoticed breaches as the system evolves.
- Using genuine urgency as a permanent excuse to skip consultation. A real safety emergency, an AI system causing active harm that must be shut down or patched immediately, can justify compressing the timeline for the minimum action needed to address that specific risk. It does not erase the duty. The defensible move documents the emergency that justified acting first, limits the action to what the emergency required, and completes genuine consultation, or seeks retrospective agreement where the rung was consent, as soon as the emergency has passed; "it was urgent" used as a standing label to avoid consultation on an ordinary rollout does not survive scrutiny.
Questions people ask
- What is works council?
- A permanent, legally recognized body of elected employees that holds rights to be informed, consulted, and in some countries to co-determine, on decisions affecting the workforce, independent of whether a union is present. Strong forms exist in Germany (under the BetrVG), the Netherlands (under the WOR), and France (the CSE), and an AI system affecting workers commonly triggers one of their rights.
- What is union?
- An organization of workers that bargains collectively with an employer over terms and conditions of employment. In systems without works councils, such as the United States, the union is the main channel of worker power, and AI-driven changes to how people are monitored, disciplined, scheduled, or paid can be subjects of bargaining, as the SAG-AFTRA video game strike showed.
- What is information (as a legal duty)?
- The lowest level of worker involvement: telling the representatives something, in time and in enough detail to understand it, without a duty to seek their view or agreement. The EU AI Act's Article 26(7) duty to tell workers they will be subject to a high-risk system before it goes live is an information duty.
- What is consultation (as a legal duty)?
- The middle level: sharing real information in time for an adequate study, hearing the opinion the representatives are entitled to form, meeting them, giving a reasoned response, and doing so with a genuine view to reaching agreement, all before the decision is implemented. It is not a veto; after genuine consultation the employer can still decide, but the process is binding and must come first (Directive 2002/14/EC).
- What is co-determination (Mitbestimmung) / consent?
- The highest level: the representatives can refuse, so the employer cannot proceed without their agreement (or a ruling from a conciliation body or court). In Germany the works council co-determines the introduction of monitoring-capable technology (BetrVG Section 87(1) No. 6); in the Netherlands the works council's consent is required for monitoring personnel systems (WOR Article 27). An AI monitoring tool commonly lands here.
Keep going
This lesson builds AI in hiring and employment decisions, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
Watch the lesson and prove it
Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.
Watch the lesson and prove it →