Skip to main content
The AI Guys

Other industries · Public administration · Large organisation

A worked example: the organisation is invented; the rules, rulings and supervisors' statements are real and linked.

What a municipality must be able to show about the algorithm behind a benefit decision

Situation
The chief information officer and the municipal secretary of a municipality
Written by
Portrait of Tijn van der ZantTijn van der Zant
Last checked

A resident applies for social assistance. Before a caseworker opens the file, a model has placed the application in a queue for extra checks, on features the caseworker cannot see. Some time later a councillor asks in committee which algorithms the municipality uses on its residents, and the answer has to be collected by telephone from several departments.

The Dutch state has measured how common that position is. The Algemene Rekenkamer, the Netherlands Court of Audit, examined 70 organisations of central government and reported on 16 October 2024 that together they had named 433 artificial intelligence systems. For 35 percent of the systems in use the organisation did not know whether the system performed as expected. Only 5 percent could be found in the public Algoritmeregister, and the auditors noted an incentive to rate risks low. That study covered central government. Municipalities publish in the same register, which on 5 October 2026 held 1,577 algorithm descriptions. Publication is still voluntary, and the register's own page says that a legal obligation is coming.

The European Union Artificial Intelligence Act, Regulation (European Union) 2024/1689, turns the subject into dated obligations. Annex III lists as high risk any system used by public authorities, or on their behalf, to evaluate whether a person is eligible for essential public assistance benefits and services, or to grant, reduce, revoke or reclaim them. The requirements apply from 2 December 2027. For systems that were put into service before that date the Act bites when their design changes significantly, and the amending regulation of 8 July 2026 sets 2 August 2030 as the final date by which providers and deployers of high risk systems intended for public authorities must have taken the steps to comply.

Four duties then rest on the municipality as deployer. Before first use it carries out a fundamental rights impact assessment under Article 27. That assessment describes the process in which the system is used, how long and how often, which groups of people are affected, the specific risks of harm to them, the human oversight, and what happens when a risk materialises, including the complaint mechanism. The result goes to the market surveillance authority. The municipality registers its use in the database of the European Union, and Article 26 forbids a public authority to use a high risk system that its supplier has not registered there. It tells the people concerned that such a system is used in decisions about them. And under Article 86 a person affected by a decision based on the output of the system has the right to a clear and meaningful explanation of the role the system played.

The usual response is to treat all this as a publication exercise: fill in the register entry, attach the data protection impact assessment and close the file. A register entry describes the system on the day it was written and is silent on who may override the model, whether anyone has compared the queue with the outcomes, and what the supplier does with the data. The data protection assessment covers personal data, and Article 27 says that the fundamental rights assessment complements it.

There is a newer problem beside the old one. In this municipality staff use general assistants to summarise objections and draft letters, and a chatbot answers residents on the website. Since 2 August 2026 Article 50 requires providers to build systems that interact directly with people in such a way that those people are told they are dealing with an artificial intelligence system, unless that is obvious. The municipality has to check that its supplier has done so. Since 15 August 2026 municipalities also fall under the Cyberbeveiligingswet, with the Rijksinspectie Digitale Infrastructuur as supervisor and with a duty of care that, following the European directive behind it, includes the security of suppliers.

So the route decision comes first. For each type of work the municipality decides whether data may go to a model provider's online interface under contract, to a private deployment, or has to stay on its own premises or those of its shared service centre. A letter template can travel further than a debt file.

On every route the same things have to exist. There is an inventory that matches the register, and for each system the municipality's role. Supplier contracts are read for location and for reuse of data. The system is tested on the municipality's own past cases. There are logs, a caseworker with the authority to deviate from the model, trained staff, and a route for incidents and complaints. A model hosted in the municipal data centre leaves that list as long as it was.

The AI Guys supports public bodies in this preparation, from the inventory to the working procedure. Next month the municipal secretary should be able to put four things before the executive board: a complete inventory set against the register entries, for each system that touches benefits a decision on whether it stays in use and who owns the impact assessment, one page per supplier on where the data goes, and an answer for the councillor that fits on a single sheet.

What we would build here

We would build each kind of work on the route its data allows: letter templates on an online service under contract, files about residents on a private deployment or the municipality's own premises. It holds up because a caseworker may deviate from the model, it is tested on past cases, and a resident can be told what part it played.

The solutions we build

What this situation risks

The most the law allows here

€15 million or 3%

Source and caveat

Breaking the other duties the Act lists, such as the transparency duties: up to €15 million or 3% of worldwide annual turnover, whichever is higher.

European Union Artificial Intelligence Act (Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744), Article 99(4). Applies from 2 August 2026.

This is the most the law allows, not what is usually imposed. A fine can only follow a duty that applies: the transparency duties from 2 August 2026, the high-risk duties from 2 December 2027 and 2 August 2028.

Source: EU AI Act, Article 99(4), consolidated version of 27 July 2026, EUR-Lex, accessed 5 October 2026.

All fines

Sources

Show the sources
  1. Algemene Rekenkamer, "Focus op AI bij de rijksoverheid", 16 October 2024 (70 organisations; 433 systems; 35 percent of systems in use with unknown performance; 5 percent in the public register; an incentive to rate risks low)
  2. Algoritmeregister: 1,577 algorithm descriptions on 5 October 2026, with municipal entries on the front page
  3. Publication not yet obligatory, "Die verplichting komt er wel aan"; the Autoriteit Persoonsgegevens coordinates the control of algorithms
  4. Annex III, point 5(a)
  5. Article 27 (fundamental rights impact assessment; elements (a) to (f); notification of the market surveillance authority; complements the data protection impact assessment)
  6. Article 26(8) (public authorities register and do not use an unregistered system) and 26(11) (inform natural persons)
  7. Article 49(3) (registration of use by public authorities in the European Union database)
  8. Article 86 (right to explanation)
  9. Article 50(1) (people are informed that they interact with an artificial intelligence system; a duty on providers), applying from 2 August 2026
  10. Cyberbeveiligingswet since 15 August 2026; municipalities listed under "Overheid (gemeenten)" with the Rijksinspectie Digitale Infrastructuur as supervisor
  11. Supply chain security in the duty of care (Article 21(2)(d) of Directive (EU) 2022/2555)

All examples