At a municipal management team meeting, the town clerk asks whether AI literacy has been dealt with. The CISO points to the lunchtime talk in the spring: 140 attendees, a presentation on Copilot and the rule that no personal data goes into free tools. Then the room goes quiet, because nobody is sure that covers a youth care consultant whose conversations are written up by a transcription tool.
The answer depends on the role. Article 4 of the AI Act asks a municipality to take measures that bring everyone who uses AI on its behalf to a sufficient level of AI literacy. What is sufficient for the front desk is not sufficient for social services, and one talk for the whole town hall does not bridge that gap. What Article 4 asks in general is covered in our overview of AI literacy in 2026.
When does Article 4 apply to a municipality?
As soon as employees use an AI system for work. The regulation is addressed to providers and deployers, and a municipality where people use Copilot, a chatbot or a transcription tool is a deployer. Whether the municipality procured those tools itself makes no difference. The text is in Regulation (EU) 2024/1689, Article 4, and AI literacy is defined in Article 3.
"On their behalf" reaches beyond the payroll. A municipality works with contractors, with a joint arrangement (a body run together with other municipalities) for social services, with an external communications agency. Anyone who uses an AI system on the municipality's behalf is covered by the measures, whether employed by it or not.
Article 4 mentions no training, no certificate and no number of hours. The European Commission confirmed in its questions and answers on AI literacy that a certificate is not required and that an organisation may record its own measures. Since August 2026 enforcement has been in the hands of national authorities; how that is divided in the Netherlands has not yet been settled.
Why does it differ by role in a municipality?
Because the work differs, and so does the data in it. A front-desk employee makes an email easier to read, a policy adviser gets AI to research a topic with sources, a youth care consultant works with notes about families and the town clerk explains the decision on each tool to the council.
| Role | What sufficient means | What you record |
|---|---|---|
| Front desk, contact centre | Which tools are allowed; that citizen service numbers and case file contents never go into a prompt; that an answer is checked before it goes to a resident | The explanation this group received and the decision on each tool |
| Policy adviser, communications | That an answer with sources still means opening those sources yourself, and that a resident's case is not material for an example in a prompt | The same, plus the condition on research tools |
| Youth care, social support, debt counselling | That case work never goes into an unapproved tool, even without a name, and which approved channel to use instead | The instruction for social services and who was briefed on it |
| IT, CISO, data protection officer (DPO) | Which tools are in use, what their status is, how to assess a tool and how the decision reaches employees | The register, the assessment questions and the decisions |
| Municipal executive, town clerk, management team | That the municipality decides, not the vendor; which applications may be high risk; who decided what | The policy, the decisions and their dates |
Most employees fall into the first two rows, where the explanation is short. What a municipal employee should and should not put in a prompt, task by task, is covered in our article on ChatGPT and personal data in municipalities.
When does it become more than an explanation?
When AI is used in a decision about a resident. Annex III of the regulation lists, among other things, AI systems that public authorities use to assess whether someone is eligible for benefits and public services. A chatbot that rewrites a letter does not fall under that. A system that pre-sorts applications may. Whether a specific application is high risk is assessed system by system and use by use, and that assessment is for the municipality to make, not the vendor. How to make it is explained in our article on when an AI application counts as high risk.
Anyone who works with such a system needs to understand how it reaches an outcome, what it misses and when to depart from it. This is also where the Dutch government's Algorithm Register comes in: public bodies are asked to describe their high-impact algorithms there in public, and what employees are told should match that description. Everyday AI use with chatbots and transcription tools does not go in it, and needs an internal register of its own.
What does that look like in a municipality?
After the meeting, the municipality in the opening paragraph looks into what is in use. Alongside Copilot it finds Perplexity, ChatGPT and three specialist tools, among them a transcription tool that two youth care teams have been using for months in conversations with families. No decision was ever taken on that tool, and the lunchtime talk did not cover it, because in the spring nobody knew it existed.
That does not prove the talk failed. It explained what was known in May; the work carried on with tools that nobody at the time recognised as AI. What it does show is what a talk cannot do: be there when a consultant starts a recording.
The municipality then does three things. It decides on the transcription tool, after looking at the vendor, the processing location and the terms. It writes a shorter instruction for social services. And it makes sure that anyone who opens an AI tool sees the decision there and then, with the approved alternative next to it.
How do you demonstrate it?
With three kinds of evidence. The explanation for each role, and who received it, shows that you told people what they need to know. The register and the decisions show what was decided. The figures on what happens while people work show whether that decision made it into the working day. Which of the three an auditor gives most weight, and what you should honestly label as not substantiated, is covered in our article on demonstrating AI literacy.
That third kind of evidence is what an attendance list cannot give you. A notice that appears when someone opens a tool can be counted together with what the person did next: opened the alternative or carried on anyway. A highlight on a citizen service number in a prompt can be counted together with its outcome: masked, removed or sent anyway. Only counts, never who.
