AI liability and operator duties in the Netherlands: 2026 guide.

An operator running an AI agent in the Netherlands faces a supervision model that is designed but not yet enacted. The draft Uitvoeringswet AI-verordening, in internet consultation from 20 April to 1 June 2026, proposes ten market surveillance authorities, two of which jointly coordinate and one of which acts as central contact point. The two coordinating bodies are the Rijksinspectie Digitale Infrastructuur and the Autoriteit Persoonsgegevens. Three layers of obligation are already in force: the Article 5 prohibited practices (since 2 February 2025), the Article 4 AI literacy duty (since 2 February 2025), and the obligations on providers of general purpose AI models (since 2 August 2025). Article 50 transparency obligations apply from 2 August 2026. Annex III high-risk obligations apply from 2 December 2027 and Annex I from 2 August 2028, under the Digital Omnibus in force since 27 July 2026 as Regulation (EU) 2026/1744. The EU AI Act applies directly whether or not the Dutch bill passes.

The Dutch supervisory architecture: proposed, not yet enacted

The EU AI Act's governance requirements became applicable on 2 August 2025. The Netherlands did not answer them by creating a dedicated AI agency or by handing the portfolio to one body. It answered them by spreading market surveillance across the supervisors that already hold the sectors, and by putting two of them in charge of holding the structure together. That answer is written into a bill that has not yet been passed, which is the single most important fact for any operator reading a compliance plan dated before this correction.

The draft Uitvoeringswet AI-verordening went to internet consultation at internetconsultatie.nl on 20 April 2026 and closed on 1 June 2026. The consultation text describes a system of ten market surveillance authorities, of which two jointly provide the coordination of supervision and one acts as the central point of contact. It names them: the Autoriteit Persoonsgegevens (AP), the Rijksinspectie Digitale Infrastructuur (RDI), the Inspectie Leefomgeving en Transport (ILT), the Inspectie Gezondheidszorg en Jeugd (IGJ), the Nederlandse Voedsel- en Warenautoriteit (NVWA), the Nederlandse Arbeidsinspectie (NLA), the Autoriteit Financiele Markten (AFM), De Nederlandsche Bank (DNB), the procureur-generaal of the Hoge Raad and the president of the Afdeling bestuursrechtspraak of the Raad van State.

The two coordinating bodies are the RDI and the AP. The RDI puts it plainly on its own AI Act page: "De RDI en de AP coördineren dit. Vanuit een expertrol ondersteunen en adviseren zij en faciliteren ze de samenwerking." Rijksoverheid, announcing the consultation on 20 April 2026, described the same pairing and added that the AP is proposed as the supervisor for the domains that have no existing supervisor, with a dedicated AI officer. Neither source describes the AP as the lead market surveillance authority, and neither gives the RDI a coordinating role that the AP does not share. Which of the ten holds the central point of contact is not stated in the consultation text and has not been confirmed at source.

The consultation text also states that the bill designates the authorities responsible for notifying the bodies that carry out conformity assessments, and that it provides for the national AI regulatory sandbox. The Raad voor Accreditatie is the Dutch national accreditation body, but nothing on rva.nl sets out its position under the AI Act, so no claim is made here about how accreditation of Dutch notified bodies for AI will run in practice.

What this means for an operator working out who to talk to

Practically, the design points a Dutch business at the supervisor it already knows. Rijksoverheid puts the principle this way: businesses can keep dealing with the sectoral supervisors they already know. Applying that principle to the ten named bodies, a bank or an insurer would stay with DNB and the AFM, a clinical deployment with the IGJ, workplace deployments with the Nederlandse Arbeidsinspectie, food and consumer product safety with the NVWA, transport and environment with the ILT, and whatever is left over with the AP, which rijksoverheid describes as the proposed supervisor for the domains that have no existing supervisor. The consultation text does not publish that mapping domain by domain, so treat the detail as a working assumption until the bill or its explanatory memorandum is enacted.

Some bodies that appear in other write-ups of the Dutch model do not appear in the consultation text at all. The Nederlandse Zorgautoriteit, the Inspectie van het Onderwijs, the College voor de Rechten van de Mens and the Autoriteit Consument en Markt are not among the ten. An operator planning its regulatory contact map should not assume they hold AI Act market surveillance competence.

For financial services operators there is one supervisory expectation that is already published and readable. The EIOPA Opinion on Artificial Intelligence governance and risk management, dated 6 August 2025, is addressed to national supervisors and clarifies key principles and requirements in insurance sector legislation for the use and supervision of AI systems, on a risk-based and proportionate approach. It sits within the Solvency II and Insurance Distribution Directive frameworks. It creates no new binding rules, and it is a reasonable guide to the questions a Dutch insurer should expect.

Where the bill stands on 17 August 2026

It is still a bill. A search of the Tweede Kamer wetsvoorstellen database on 17 August 2026 for AI-verordening returned no results, and a search for kunstmatige intelligentie returned no results either. The European Commission's own national resources page for the AI Act lists no Dutch designated market surveillance authority, notifying authority or single point of contact. To become law the proposal would still need to run its parliamentary course and be published in the Staatsblad.

Two consequences follow. The first is that nobody in the Netherlands should be waiting for the national act before building an AI governance programme, because the Regulation applies directly and its deadlines do not move with the Dutch legislative calendar. The second is that the penalty ceilings people quote in Dutch board papers are European, not Dutch. Article 99 of the AI Act sets administrative fines of up to EUR 35,000,000 or 7 per cent of total worldwide annual turnover for the preceding financial year, whichever is higher, for breach of the Article 5 prohibitions; up to EUR 15,000,000 or 3 per cent for breaches of the operator obligations listed in Article 99(4), which include Article 26 and the Article 50 transparency rules; and up to EUR 7,500,000 or 1 per cent for supplying incorrect, incomplete or misleading information to notified bodies or national competent authorities. Member States lay down the penalty rules and national bodies impose them. Until the Dutch designations are law, which Dutch body would impose which fine is a question the bill answers and the statute book does not.

What is already in force: the three active obligation layers

Article 5: prohibited AI practices (in force since 2 February 2025)

Article 5 of the EU AI Act prohibits a set of AI practices that the legislature deemed incompatible with fundamental rights. These prohibitions apply to any operator deploying AI in the Netherlands, whether or not the Uitvoeringswet AI-verordening is in force.

The prohibited practices are: subliminal techniques beyond a person's awareness, or purposefully manipulative or deceptive techniques, that materially distort behaviour and cause or are likely to cause significant harm; exploitation of vulnerabilities due to age, disability or a specific social or economic situation, to the same effect; social scoring, meaning the evaluation or classification of natural persons or groups based on social behaviour or personal characteristics leading to detrimental treatment, which the Article does not confine to public authorities; assessment of the risk that a person will commit a criminal offence based solely on profiling or on assessing personality traits; creating or expanding facial recognition databases through untargeted scraping of facial images from the internet or CCTV footage; inferring emotions in the workplace and in education institutions, except for medical or safety reasons; biometric categorisation to deduce race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation; and real-time remote biometric identification in publicly accessible spaces for law enforcement, permitted only for three narrow objectives.

Two further prohibitions, covering AI systems that generate non-consensual sexual deep fakes and child sexual abuse material, start to apply on 2 December 2026 according to the European Commission implementation timeline.

What the Autoriteit Persoonsgegevens has said about its own supervisory priorities under Article 5 could not be established for this correction. The AP's website returns HTTP 403 to automated retrieval, so any account of AP guidance, AP reports on algorithmic risk or AP enforcement would be repetition rather than verification, and it has been left out.

Article 4: AI literacy (in force since 2 February 2025)

Under Article 4, both providers and deployers must take measures to ensure a sufficient level of AI literacy among their staff and others who operate AI systems on their behalf. The literacy requirement is calibrated to the role of the person and the nature of the AI system: a legal team approving an AI-generated contract needs different competencies than a warehouse worker interacting with an algorithmic task-allocation system.

Article 4 was not deferred by the Digital Omnibus. It has been in application since 2 February 2025, on the European Commission's own implementation timeline, and remains so.

In the Dutch context, the AI literacy obligation sits next to the data protection regime administered by the AP under the Uitvoeringswet Algemene verordening gegevensbescherming, which is the Dutch implementing act for the GDPR and is in force. It is not the Wet bescherming persoonsgegevens, which was repealed on 25 May 2018 and still appears in older compliance material. Operators who already run a data protection training programme should extend it to cover AI literacy rather than treating Article 4 as a separate compliance stream.

General purpose AI model obligations (in force since 2 August 2025)

On 2 August 2025 the AI Act obligations for providers of general purpose AI models entered into application, and the governance requirements took effect with them. Supervision of those model providers sits with the European AI Office rather than with any national market surveillance authority, and the AI Office holds a separate penalty regime for them under Article 101, with a ceiling of EUR 15 million or 3 per cent of worldwide annual turnover.

For deployers in the Netherlands, the practical consequence is that any AI system built on a general purpose model must be deployed consistently with the model provider's usage policies and documentation. Where a business changes the intended purpose of a system so that it becomes high-risk, Article 25 makes it a provider of that system with the full provider obligation set. Dutch operators building AI agents on top of foundation models from non-EU providers should read the documentation those providers make available and check that their use case sits inside the documented intended purpose.

Article 50: transparency obligations from 2 August 2026

Article 50 applies from 2 August 2026 and was not deferred by the Digital Omnibus. Its duties are split between providers and deployers, and getting that split wrong is the most common way a compliance plan misallocates work.

Article 50(1) puts the obligation on providers: an AI system intended to interact directly with natural persons must be designed and developed so that the persons concerned are informed that they are interacting with an AI system, unless that is obvious from the context. Article 50(2) also puts the obligation on providers, including providers of general purpose AI systems, to ensure that synthetic audio, image, video or text outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Deployers hold Article 50(3), informing the natural persons exposed to an emotion recognition or biometric categorisation system of its operation, and Article 50(4), disclosing that image, audio or video content constituting a deep fake has been artificially generated or manipulated.

For a Dutch business buying a customer-facing chatbot, the practical consequence is contractual as much as technical. The disclosure and marking duties belong to the party that built the system, so a deployer's leverage is the procurement document. Ask for the Article 50(1) disclosure behaviour and the Article 50(2) marking to be warranted in writing, and keep the evidence. Where the business generates deep fakes or runs emotion recognition, the duty is its own and cannot be pushed up the chain.

The European Commission implementation timeline records 2 December 2026 as the transitional deadline for certain providers of AI systems to comply with Article 50(2). This is in force, not proposed.

High-risk AI obligations: the timeline as it now stands

The Digital Omnibus entered into force on 27 July 2026 as Regulation (EU) 2026/1744. It moved the standalone Annex III high-risk obligations to 2 December 2027 and the Annex I obligations, for AI embedded in regulated products, to 2 August 2028. The European Commission's implementation timeline carries both dates. Nothing else moved. The Article 5 prohibitions, the Article 4 AI literacy duty, the general purpose AI obligations and the Article 50 transparency rules were not deferred.

This is a deferral of the conformity work, not a reprieve from the Regulation. Any guide, policy or board paper still describing 2 August 2026 as the binding high-risk date, or describing the deferral as awaiting adoption, is out of date. An earlier version of this guide said both things and contradicted itself doing so.

The Annex III areas most likely to affect Dutch businesses are point 4, employment, workers management and access to self-employment, which covers recruitment, selection, decisions on work relationships, task allocation and performance monitoring; point 5, access to and enjoyment of essential private services and essential public services and benefits, where point 5(b) names creditworthiness evaluation and credit scoring with an exception for financial fraud detection; and point 3, education and vocational training. Point 6 is law enforcement and point 8 is administration of justice and democratic processes. None of the eight areas covers medical devices, so an earlier version of this guide was wrong to place software as a medical device in Annex III point 6. AI embedded in regulated products runs on the Annex I track and the later 2 August 2028 date.

The extra fourteen months are best spent on the parts of high-risk compliance that take organisational time rather than paperwork: knowing which systems you actually run, establishing who owns each one, and building a risk management and logging practice that a supervisor could inspect without a special project being convened first.

Deployer obligations under Article 26: what Dutch businesses must implement

Article 26 of the EU AI Act sets out the obligations that apply to deployers of high-risk AI systems. These apply to any Dutch business that puts a high-risk AI system to use in a professional context, regardless of whether the AI system was developed in the Netherlands. The deployer does not need to have built the system. Using it is sufficient to engage the Article 26 obligations.

The core Article 26 obligations for Dutch deployers are as follows. They must use the system in accordance with the instructions for use. They must assign human oversight to natural persons who have the necessary competence, training and authority. They must monitor operation in line with the instructions and keep the logs the system generates automatically, for a period appropriate to the purpose and in any event at least six months. On identifying a serious incident they must inform first the provider, then the importer or distributor and the relevant market surveillance authorities. They must inform the natural persons who are subject to the use of the system, and an employer deploying such a system at work must inform workers' representatives and the affected workers first. Deployers that are public authorities must comply with the registration obligations referred to in Article 49 and register the system in the EU database. Article 26(9) tells deployers of Annex III systems to use the information provided under Article 13 when carrying out their data protection impact assessment.

Article 27 sits on top for a defined group: deployers that are bodies governed by public law, private entities providing public services, and deployers of the systems in Annex III points 5(b) and 5(c), the point 5 area covering creditworthiness and insurance risk assessment. They must carry out a fundamental rights impact assessment before putting the system into use. Systems in the Annex III point 2 area are excluded from that duty.

A deployer becomes a provider under Article 25, with the full provider obligation set including conformity assessment, if it puts its name or trademark on a high-risk AI system already placed on the market, makes a substantial modification to such a system in a way that leaves it high-risk, or modifies the intended purpose of a system so that it becomes high-risk. Reselling an agent under your own brand is the version of this that catches Dutch resellers and integrators most often.

The Algoritmeregister: Dutch public sector transparency

The Dutch government runs the Algoritmeregister at algoritmes.overheid.nl, a public register in which government organisations publish information about the algorithms they use. The register describes its own scope in one line: "Het register focust op impactvolle algoritmes (waaronder hoog risico AI-systemen) en geeft de bezoeker inzicht in hoe deze algoritmes werken." On 17 August 2026 it listed 1,536 algorithms.

Its scope is wider than the AI Act's high-risk registration duty, because it reaches impactful algorithms that would not meet the Annex III threshold, and it covers systems already in use rather than only new deployments. Whether publication is legally compulsory could not be established at source. The register's own guidance for government bodies is phrased as an invitation, "Werk je bij een overheid en wil je ook een algoritme registreren?", and no statutory duty is stated on the site. A private operator delivering AI into Dutch public administration should therefore treat register disclosure as a procurement question in the first instance, and read the contract, rather than assume a general legal obligation either way.

The revised Product Liability Directive: what Dutch operators should prepare for

Directive (EU) 2024/2853 on product liability came into force on 8 December 2024. EU countries have until 9 December 2026 to transpose it. The European Commission states that the new rules apply to products placed on the market as of 9 December 2026, and that the 1985 directive remains applicable for products placed on the market before that date.

The Commission also states that the rules "specifically clarify that all types of software are covered by the new directive, including applications, operating systems and AI-systems". That is the sentence that matters for an AI agent operator. From December 2026, an agent that causes damage sits inside a strict liability regime in which the claimant does not have to prove negligence, and the Netherlands will need implementing legislation in place by the transposition deadline.

The detail of the Directive's evidentiary presumptions and of the damage categories it recognises was not verifiable at source for this correction, because EUR-Lex returned empty documents to every request, so the specifics that appeared in the earlier version of this section have been removed. Operators should read the transposing Dutch bill when it appears rather than rely on secondary summaries, this one included.

Dutch operators who deploy AI agents for third parties or consumers should begin reviewing their indemnification provisions in commercial contracts and their liability caps in terms of service. Contracts that were drafted before the revised PLD was adopted may not adequately allocate the new strict-liability exposure.

AI liability and insurance in the Dutch market

The Dutch insurance market is one of the most developed in Europe, with strong penetration of professional indemnity, directors and officers, and cyber policies among large and mid-sized businesses. However, the standard forms of these policies were written before AI agents became common operational tools, and most carry exclusions or gaps that leave AI-generated errors uninsured.

The practical insurance issues Dutch operators face in 2026 are four. First, general liability and professional indemnity policies typically exclude or limit coverage for losses arising from the outputs of automated systems. Operators should review their policy wordings for exclusions referencing automated decision-making, algorithms, or software-generated outputs. Second, cyber policies in the Netherlands generally cover data breaches and ransomware but do not contain affirmative coverage for third-party losses caused by an AI agent that provides incorrect advice, executes an unauthorised transaction, or generates misleading content. Third, the revised Product Liability Directive exposure from December 2026 onward creates a new category of strict product liability that existing products liability policies, where they exist, may not have priced for AI systems. Fourth, if the AP ends up holding a data protection mandate and an AI Act mandate at once, a single AI incident may produce a regulatory enforcement track and a civil damages claim in parallel, each of which needs separate cover.

The affirmative AI insurance market gives a few concrete reference points. ElevenLabs announced a policy backed by the AIUC-1 standard on 12 February 2026, with no carrier named at source. AIUC is the Artificial Intelligence Underwriting Company. Munich Re's aiSure is written through Mosaic with EUR, USD or CAD 15 million of initial capacity from 26 February 2026, and Munich Re describes it as a performance guarantee that settles on measurable performance data, with technical due diligence required before cover is written. Armilla offers AI liability cover with a limit of USD 25 million per organisation. Dutch operators should confirm availability and applicability with a specialist broker. The product-specific reviews at agentinsured.eu carry the structured analysis.

What operators in the Netherlands should do by category and timeline

The practical preparation sequence for a Dutch business deploying AI agents follows the dates on the European Commission implementation timeline.

Now: confirm that no deployed AI system engages the Article 5 prohibited practices; document your AI literacy programme under Article 4; identify every AI system you run that is built on a general purpose AI model; get written warranties from your providers on the Article 50(1) disclosure behaviour and the Article 50(2) marking; audit your insurance policies for AI exclusions and gaps; and if you may be in Annex III scope, begin the self-classification exercise.

From 2 August 2026, Article 50 in application: ensure the systems you deploy tell people they are dealing with an AI system, which is a provider duty you should have contracted for; put the required notices on any emotion recognition or biometric categorisation system you deploy; and disclose any deep fake content you generate or manipulate.

By 2 December 2026: the transitional deadline for certain providers to comply with the Article 50(2) marking duty, and the date on which the prohibitions on AI systems generating non-consensual sexual deep fakes and child sexual abuse material start to apply.

By 9 December 2026, Product Liability Directive transposition: review and update commercial contracts and terms of service to reflect the strict product liability exposure that attaches to products placed on the market from that date; engage your broker on product liability cover for AI systems supplied to consumers or third parties.

By 2 December 2027, Annex III high-risk: complete Annex III classification; run a risk management system, technical documentation, logging and human oversight that a supervisor could inspect; complete the Article 27 fundamental rights impact assessment where you fall in its scope; and register in the EU database where you are a public authority deployer, under Article 49. Annex I embedded systems follow on 2 August 2028.

Frequently asked questions