AI in Dutch employment law: what employers must arrange

As an employer in the Netherlands, you may use AI in recruitment, performance management and workforce planning, but four sets of rules apply at once: the EU AI Regulation, the GDPR, Dutch equal treatment law and the Works Councils Act (Wet op de ondernemingsraden, WOR). You, not the software supplier, are liable for the outcome, and most AI tools that monitor or assess staff need the prior consent of your works council (ondernemingsraad).

This article explains what each regime requires, when the obligations apply and what you should have in place before an AI tool influences a hiring, pay or dismissal decision.

Which rules apply when you use AI in employment decisions?

Four regimes apply side by side, and none replaces the others. The AI Regulation sets product and use rules, the GDPR governs the personal data, the equal treatment acts govern the outcome, and the WOR gives employee representatives a say in the introduction.

What matters legally is not how a supplier describes the tool. The question is whether the system produces or materially influences a decision about an identifiable person, and whether a human can genuinely reach a different conclusion. Typical examples are CV screening and ranking, scoring of video interviews, productivity dashboards that feed into pay, shift-allocation software and biometric time registration. Our article on AI in the workplace covers the use of generative tools by staff.

When do the AI Regulation’s obligations apply?

The prohibitions of the EU AI Regulation already apply; the obligations for high-risk systems have been postponed. Under the AI omnibus, Regulation (EU) 2026/1744 (in force since 27 July 2026), the high-risk rules for systems listed in Annex III, which includes recruitment, task allocation, promotion, termination and performance monitoring, apply from 2 December 2027.

Already prohibited are, among other things, AI systems that infer the emotions of workers (except for medical or safety reasons), social scoring and biometric categorisation that infers sensitive characteristics such as race, political opinion, trade union membership or sexual orientation. Much of what is sold as “sentiment analysis” or “engagement monitoring” falls within the ban on emotion recognition at work.

Once the high-risk rules apply, you are usually the deployer, not the provider. As deployer you must use the system according to the instructions, assign human oversight to trained staff with authority to intervene, keep the logs the system generates, check that your input data is relevant, and inform workers and their representatives before the system is put into use. If you put your own name on a system or modify it substantially, you can become the provider and take on the full conformity assessment.

The highest fines under the AI Regulation, for prohibited practices, are €35 million or 7% of worldwide annual turnover, whichever is higher. The two-year delay is time to prepare, not a reason to wait: the GDPR, equal treatment law and the WOR apply today.

How does equal treatment law apply to algorithmic selection?

A decision produced by an algorithm is judged exactly like a decision taken by a manager. The General Equal Treatment Act (Algemene wet gelijke behandeling, AWGB) prohibits distinction on grounds such as religion, belief, political opinion, race, sex, nationality, sexual orientation and civil status; separate acts cover age, disability and chronic illness, working hours and fixed-term contracts.

Both direct and indirect distinction are prohibited. Most algorithmic problems are indirect: a model trained on past hiring data reproduces old patterns, and proxies such as postcode, graduation year, a career gap or language use do the work of a prohibited criterion. Indirect distinction is only lawful if it is objectively justified by a legitimate aim, with appropriate and necessary means.

The burden of proof decides most cases. Under Article 10 AWGB, once an applicant or employee puts forward facts that suggest distinction, you must prove that you did not discriminate. A statistical disparity in the outcomes of an automated selection can be enough, and an employer that cannot explain what the system weighed will struggle to discharge that burden. Complaints go to the Netherlands Institute for Human Rights (College voor de Rechten van de Mens) or to the civil courts.

There is no statutory duty to use a documented recruitment method: the Senate (Eerste Kamer) rejected the bill on equal opportunities in recruitment (bill 35673) on 26 March 2024. That does not reduce your liability for the outcome. It means you choose your own method and carry its risk, so document your selection criteria and test the outcomes across groups at regular intervals.

What does the GDPR require for automated decisions?

The GDPR applies now, and the Dutch Data Protection Authority (Autoriteit Persoonsgegevens) enforces it. Three rules do most of the work in an employment setting.

• No decisions based solely on automated processing. A person has the right not to be subject to a decision based solely on automated processing with legal or similarly significant effects. Automatically rejecting an applicant, or setting a pay-related rating without a real human assessment, falls within this rule. The reviewer must have the authority and the information to decide differently.
• Transparency. You must give meaningful information about the logic involved and the consequences of the processing. That is not source code, but you must be able to say which factors the system weighs.
• Data protection impact assessment (DPIA). Systematic evaluation of employees by automated means and systematic monitoring at work require a DPIA before the system goes live. Involve your data protection officer and record the alternatives and safeguards you considered.

Consent is rarely a valid legal basis in employment, because the dependency of the employee means consent is seldom freely given. You will normally rely on the performance of the contract, a legal obligation or a legitimate interest, and a legitimate interest needs a documented balancing test. Where the supplier processes data on your instructions, a processing agreement is required; see our page on controller and processor roles. Breaches can lead to fines of up to €20 million or 4% of worldwide turnover and to compensation claims by employees.

Do you need the works council’s consent?

Usually yes. Under Article 27(1)(l) WOR, any arrangement on facilities that are aimed at or suitable for monitoring the presence, conduct or performance of employees requires the works council’s prior consent, and so do rules on staff appraisal under Article 27(1)(g) WOR.

Almost every AI tool for performance measurement or monitoring falls within one of those categories. A decision taken without consent is void if the works council invokes the nullity. If the council refuses, you can ask the subdistrict court (kantonrechter) for substitute consent under Article 27(4) WOR, which is only granted if the refusal is unreasonable or the decision is required by weighty business reasons.

In addition, the works council has an advisory right on the introduction or modification of an important technological facility (Article 25(1)(k) WOR). If you do not follow its advice, you must suspend implementation for one month (Article 25(6) WOR), during which the council can appeal to the Enterprise Chamber (Ondernemingskamer) of the Amsterdam Court of Appeal under Article 26 WOR. In practice: involve the council at the selection stage, share the technical documentation and record agreements on testing and evaluation.

Can AI output support a dismissal?

Only as evidence, never as the ground itself. Dutch dismissal law is a closed system: under Article 7:669(1) of the Dutch Civil Code (BW) you need a reasonable ground and must consider redeployment, and only the statutory grounds in Article 7:669(3) BW count.

For poor performance (ground d), you must have told the employee in good time what was wrong, given a real opportunity to improve, and the unsuitability must not result from insufficient training or poor working conditions. A productivity score does not prove any of these points, and the employee can challenge how and over what period the system measured. During illness, the prohibition on termination in Article 7:670(1) BW applies, so scoring an employee who is ill or on adapted duties against the normal standard is risky.

Where automation removes positions, the redundancy route runs through the UWV and the duty to consider redeployment, with or without retraining, applies. Use automated output as one input to a documented human assessment, record the reasoning of the decision-maker and keep the improvement process on paper. Employers who get this wrong usually lose on process, not on substance; see our overview of common employment law compliance mistakes.

Who owns what AI produces for your business?

Under Article 7 of the Dutch Copyright Act (Auteurswet), you as the employer are regarded as the author of works your employees create in the course of their employment, unless agreed otherwise. That rule does not cover freelancers: their rights must be assigned by deed.

Copyright only protects a work that reflects human creative choices. Output generated by a prompt with little human input is therefore probably not protected, and competitors may copy it. Also read the terms of the AI platform: they determine what rights the provider claims over your input and output, and entering confidential information or personal data into a public tool can breach both confidentiality obligations and the GDPR. Our guide to intellectual property law in the Netherlands explains the wider rules.

What should you check as an employer?

• Make an inventory of every system that touches an employment decision, including tools a department bought on its own, and record which decision each one influences.
• Check per system whether it falls under a prohibition, is likely high-risk under Annex III, involves solely automated decisions under the GDPR, and requires works council consent under Article 27(1) WOR.
• Carry out and document a DPIA before the system goes live.
• Obtain works council consent or advice before implementation, not after the supplier contract is signed.
• Require from the supplier technical documentation, bias test results, a processing agreement, audit rights and incident notification.
• Write down which decisions always need a human assessment and who may override the system.

What should you check as an employee or applicant?

• Ask whether a decision about you was taken solely by a system and request meaningful information about the logic involved.
• Use your GDPR right of access to obtain the personal data and scores the system holds about you.
• Check whether the works council consented to the monitoring or appraisal system that was used.
• If you suspect discrimination, collect facts that point to it; under Article 10 AWGB the employer must then prove it did not discriminate.
• In a performance-based dismissal, check whether you were warned in time and given a real opportunity to improve, as Article 7:669(3)(d) BW requires.

What can we do for you with AI in employment?

Our employment lawyers advise employers, works councils and employees on the legal side of workplace AI.

• We classify your AI tools under the AI Regulation, the GDPR and the WOR and set out which steps apply when.
• We draft or review the DPIA, the processing agreement and the supplier contract.
• We prepare the request for consent or advice to the works council and guide the consultation.
• We draft an internal AI policy on human oversight, testing and complaints.
• We assess a contested automated decision and conduct the procedure before the subdistrict court or the Netherlands Institute for Human Rights.

Summary

• AI in employment is governed by the AI Regulation, the GDPR, equal treatment law and the WOR at the same time; the employer bears the risk.
• Emotion recognition at work is already prohibited; the high-risk rules for HR systems apply from 2 December 2027.
• Solely automated decisions with significant effects are not allowed; a real human assessment is required.
• Monitoring and appraisal tools need prior works council consent under Article 27(1) WOR.
• An AI score is evidence, not a dismissal ground; Article 7:669 BW still applies in full.

Frequently asked questions