AI in government: what Dutch municipalities and central government are doing now, and what the AI Act asks

Dutch government and AI today

Mostly internal work: around two thirds of central government’s AI systems are aimed at internal processes.

In October 2024, the Netherlands Court of Audit (Algemene Rekenkamer) mapped for the first time how much AI central government uses. In the study Focus on AI in central government, 70 government organisations together reported 433 AI systems. 167 of them were still ongoing experiments. Around two thirds of the systems are aimed at internal processes without a direct impact on citizens and businesses, such as turning conversations into text or anonymising documents.

Generative AI is on the rise too, as far as the figures can show. According to the Government-wide Monitor on Generative AI, which the research organisation TNO published in December 2025 for the Ministry of the Interior and Kingdom Relations (BZK), the number of applications found rose from 8 in June 2024 to 81 in June 2025. Of these, 37 were in use and 29 were still an experiment. In more than 75 percent of the applications, civil servants are the users, not citizens. TNO cannot establish how representative this picture is.

Government on generative AI

Since April 2025, civil servants have more room to use generative AI, under conditions.

In January 2024, the Dutch cabinet presented a government-wide vision on generative AI. It sets out four principles: generative AI must be safe and fair, serve human well-being and human autonomy, and contribute to sustainability and prosperity. One of the six lines of action: government experiments with generative AI itself, safely.

In April 2025, a government-wide position followed, agreed with municipalities, provinces, water authorities and implementing agencies. It gave civil servants more room to use generative AI in their work, for example to speed up permit applications or answer questions from citizens faster. But under conditions, such as risk analyses. Central government also encourages applications developed in Europe and open-source solutions. According to the Dutch government, these are the first rules for AI that apply to the whole of government.

The Dutch Algorithm Register

The register is growing fast, but registration is not yet a legal requirement.

In the Algorithm Register (Algoritmeregister), Dutch public authorities publish information about the algorithms they use. The register focuses on algorithms with a high impact, including high-risk AI. According to a letter to the Dutch House of Representatives of 9 September 2025, it then held 1,000 algorithms, against 675 on 1 January 2025. On 31 October 2026, the register counted 1,613. Registration is not yet a legal requirement. In December 2022, the cabinet did announce a legal obligation of transparency about algorithms, but in September 2025 the State Secretary for the Interior wrote that he did not (yet) consider such an obligation opportune at this time.

Three recent examples from the register show what it looks like:

• Wmo Agent, municipality of Almere
  AI writes a draft of the assessment report for an application under the Social Support Act (Wmo), the Dutch law on care and support at home. The agent does not decide on support; the consultant checks everything and a colleague reviews the report. According to the registration, the agent has been put on hold temporarily for technical reasons.
• AI desk, social services department of Noardwest Fryslân
  staff query a knowledge base on topics such as the Participation Act and the Social Support Act. A staff member always checks the answer, and the AI desk never talks to residents directly.
• Service desk assistant, province of Noord-Holland
  AI drafts answers to questions from residents and businesses to the provincial service desk.

How do municipalities use AI?

In the TNO monitor, municipalities are the largest group, and the VNG focuses on working together.

Of the 81 generative AI applications in the TNO monitor, 34 were at municipalities, mostly chatbots for civil servants and for contact with residents. One example is GEM, a virtual assistant on the websites of 25 municipalities, started in 2019 by Utrecht and Tilburg. The municipalities share the costs and a single privacy assessment (DPIA, data protection impact assessment).

The VNG, the Association of Netherlands Municipalities, focuses on working together. Two guides appeared in October 2026. The guide to scaling up AI (27 October) helps public authorities roll out a successful pilot across more organisations. The guide on AI in software updates (8 October) helps buyers now that municipalities increasingly receive software updates with AI built in.

What does the AI Act ask?

The AI Act sets extra requirements, above all for AI that assesses eligibility for essential public assistance benefits and services.

The AI Act affects public authorities on a few extra points. We follow the consolidated text on the European Commission’s AI Act Service Desk, after the amendment by the Digital Omnibus.

• Already in force
  since 2 February 2025, the prohibitions apply, such as social scoring that leads to unjustified detrimental treatment of people. Organisations must also ensure the AI literacy of their staff.
• High risk
  AI that a public authority uses to evaluate whether someone is eligible for essential public assistance benefits and services, including healthcare services, or to grant, reduce, revoke or reclaim them, is listed as high risk in Annex III. The strict rules for these systems apply from 2 December 2027.
• Fundamental rights assessment
  before using such a system, public authorities must in principle carry out an assessment of the impact on fundamental rights.
• Registration
  a public authority may not use a high-risk system that is not registered in the EU database (Article 26). For critical infrastructure (Annex III, point 2), registration takes place at national level instead (Article 49).
• Existing systems
  high-risk systems intended for use by public authorities must comply by 2 August 2030 at the latest (Article 111).

Government does not need to be the fastest with AI, but it does need to be the most trustworthy. There is a lot at stake. Start with a vision, then map, process by process, what AI can do and where a civil servant makes the difference.

Where does it get stuck?

It gets stuck on knowing whether AI works, on knowledge and resources, and on trust.

Knowing whether it works. According to the Court of Audit, for 35 percent of the AI systems central government had in use, it was unknown whether they met expectations. For more than half of the systems, the opportunities had not yet been weighed against the risks. Only 5 percent of the systems were in the register at the time.

Knowledge and resources. In TNO’s interviews, public authorities name, among other things, limited AI literacy, data that is not always in order, unclear rules and too little budget and time. Staff also sometimes use AI tools on their own initiative.

Trust. The Dutch childcare benefits scandal shows what is at stake. In December 2021, the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) fined the Dutch Tax and Customs Administration 2.75 million euros. The tax authority used the nationality of applicants for childcare benefit in, among other things, a system that automatically flagged applications as risky. According to the AP, this was unlawful and discriminatory. We wrote earlier about algorithms gone wrong.

How does a public body start?

Start with internal work, set the goal in advance and keep a human in charge.

1. Start with internal work.
   According to the Court of Audit, around two thirds of the AI systems are aimed at internal processes, and according to TNO a staff member is the user in more than 75 percent of the generative AI applications. Then map process by process, from a Wmo report to answering questions from residents, which tasks AI can prepare and which decision stays with the civil servant.
2. Set the goal in advance.
   The Court of Audit found that it remains unknown whether AI works if the goal has not been set beforehand.
3. Keep a human in charge.
   The examples from the register show how: the AI writes a draft, a staff member checks and decides.
4. Check whether it is high risk.
   If the system assesses eligibility for essential public assistance benefits and services, for example (Annex III), it is in principle high-risk: stricter rules apply from 2 December 2027, and as a rule a fundamental rights assessment. An exception applies if the system only performs a narrow procedural or preparatory task without materially influencing the decision (Article 6(3)).
5. Register.
   Put high-impact algorithms in the register, even though that is not yet mandatory.
6. Teach your people to work with it.
   AI literacy has been an obligation since February 2025.
7. Work together.
   Build on existing solutions and on the VNG guides, instead of building everything yourself.

The government of the future

The number of applications found is growing fast, but whether AI works and is fair remains an open question.

According to TNO, many generative AI applications are already in use, and the VNG is now helping to scale up pilots. The number of registrations and applications found is growing fast. At the same time, the questions of the Court of Audit and the lesson of the childcare benefits scandal have not yet been answered: does it work, and is it fair to citizens? For high-risk systems in Annex III of the AI Act, such as systems that assess eligibility for essential public assistance benefits and services, stricter requirements apply from 2 December 2027. Before using them, public authorities must in principle also assess the impact on fundamental rights. An ethical perspective belongs there from the start.

AI is not the first new technology in government, and certainly not the last. My advice: let people play and experiment with it themselves, in a safe environment and on a real problem of residents. That is how you learn how it works. It takes leaders who activate their civil servants and give them room, and who then look critically at the results together. Enthusiasm for change and a critical eye suit a government that wants to win back trust. That side of change, the people and the organisation, is the subject of the keynote on the organisation of the future.

Frequently asked questions

What is the Dutch Algorithm Register?

A public government website, algoritmes.overheid.nl, where Dutch public authorities describe which algorithms they use, for what and with what oversight. The register focuses on algorithms with a high impact. On 31 October 2026, it held 1,613 algorithms. Registration is not yet a legal requirement.

Can civil servants use ChatGPT?

Since the government-wide position of April 2025, civil servants have more room to use generative AI, under conditions such as a risk analysis and reliable models. Some organisations offer their own assistant, such as the PZH assistant of the province of Zuid-Holland, only for their own civil servants.

When do the AI Act rules apply?

The prohibitions and the AI literacy obligation have applied since 2 February 2025. For high-risk AI in Annex III, such as AI that assesses eligibility for essential public assistance benefits and services, the strict rules apply from 2 December 2027. Existing high-risk systems intended for public authorities must comply by 2 August 2030 at the latest.

What is the lesson of the benefits scandal?

That a system which automatically flags people as a risk can discriminate if it uses the wrong data. The Dutch Data Protection Authority fined the tax authority for this in 2021.