Back to the blog

Who will supervise AI in the Netherlands: the draft of 20 April

· By Bee Flow · 5 min read

On 20 April 2026 the Dutch government put its draft AI Act implementation act out for consultation. Ten regulators, the AP as the default, the RDI as contact point. What a company can expect, what is still open, and how to respond before 1 June.

A wall of ten letterboxes of different sizes, with a thin amber thread linking the large middle one to its neighbour

On 20 April 2026 the Dutch government opened a public consultation on the draft act that implements the EU AI Act in the Netherlands, the Uitvoeringswet AI-verordening. In a letter to the House of Representatives sent the same day, state secretary Aerdts (Economic Affairs and Climate) sets out how supervision of the AI Act is meant to work here. It is a proposal: the Council of State and both chambers of parliament come next, and the regulators get their formal powers only once the act enters into force.

For a privacy or compliance officer at a Dutch company it is still worth reading now: for the first time it is written down which regulator you will deal with on AI.

Ten regulators, with the AP as the default

The government proposes to designate ten market surveillance authorities, the letter says. According to the government's news item, existing regulators also supervise AI within their own domain, so companies deal with bodies they know. Where no clear regulator exists, the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) is proposed.

Annex 1 of the letter divides the work roughly like this (leaving out the courts, which get their own):

Part of the AI Act Proposed regulator
Prohibited AI practices AP, with the AFM (financial markets authority) for manipulation and exploiting vulnerabilities by financial institutions
Transparency obligations, such as for chatbots and AI-generated content AP, with the AFM and DNB (the central bank) for financial institutions
High risk: among others biometrics, education, employment and workers' management, essential services AP
High risk: creditworthiness, and premiums for health and life insurance AFM and DNB for financial services, the AP otherwise
High risk: critical infrastructure RDI or ILT, depending on the domain
AI in products such as machinery, toys, lifts and medical devices Today's product regulators: NVWA, NLA, ILT, RDI and IGJ

The European Commission will itself supervise general-purpose AI models such as those behind ChatGPT and Gemini, the letter notes.

What the AP and the RDI coordinate

The Dutch Authority for Digital Infrastructure (Rijksinspectie Digitale Infrastructuur, RDI) becomes the single point of contact that Article 70 of the AI Act requires: the contact for the Commission and for anyone here with a question about supervision. The RDI and the AP share a coordinating role. Annex 1 says this should give one reading of the regulation, shared good practice and joint investigations.

One detail matters to privacy officers. Inside the AP, AI Act supervision is to sit apart from GDPR supervision, independent of it and on an equal footing, with its own contact person on the board. The government's news item calls this an AI board member, and the letter says decisions under the AI Act need that person's consent. So you will often deal with the same organisation as for the GDPR (the AVG in Dutch), through a different line inside it.

What a company can expect already

For a company outside the financial sector that uses AI rather than builds it, supervision in this proposal mostly lands with the AP, where rules apply to that use at all. Using AI in HR, for example in recruitment, can fall under the high-risk area of employment and workers' management, which the government assigns to the AP. Financial institutions keep the AFM and DNB. Makers of products that already fall under EU product rules keep the regulator they know.

More important than who: the obligations do not wait for the Dutch act. The government expects the implementation act to enter into force later than the AI Act requires, but notes that the obligations apply regardless, because the regulation applies directly. The prohibitions have applied since February 2025. We wrote in February about the one closest to employers, emotion recognition at work. The consultation page names August 2026 and August 2027 for the remaining parts.

If you build AI systems yourself, there will be one Dutch AI regulatory sandbox run by all the regulators together. The letter says it gives legal and technical advice, no infrastructure or data, and that rules cannot be set aside in it.

What is still open

The regulators are still working out what the new tasks cost and whether they can carry them out. The mandatory assessments run alongside the consultation. For the high-risk areas that go to the AP, the letter calls cooperation agreements with the sector regulators the obvious step. The letter does not say any are in place yet.

Then there is Brussels. In November 2025 the European Commission proposed changes to the AI Act, known as the AI omnibus. The government writes that the dates on which several provisions start to apply will probably move back as a result. How that turns out is not known yet, and we would not plan around a delay before there is an adopted text.

What to do in the coming weeks

Make a list of where AI is used in your company, and next to each use note which part of the AI Act applies. A chatbot on your website, a tool that pre-sorts job applications and an AI chat for internal writing sit under very different rules. The same list shows which regulator goes with which use.

If you want to respond to the proposal, you can do so until and including 1 June 2026 at internetconsultatie.nl/uaiv. The consultation is in Dutch, and a response is only made public if you agree to that.

This article is not legal advice. It describes a draft that can still change.

Where Bee Flow helps

An inventory like that is easier when your team's AI use happens in one place, not across separate accounts. In Bee Flow the administrator chooses which models are available: in the cloud, at a European provider or on your own server. Privacy Shield blocks personal data, or replaces it with a token, before a message goes to a model. An automation can carry an approval step, so a person decides before anything is sent. Privacy and security explains how that works, and you can get in touch to see how it would fit your situation.

#compliance #privacy