Dutch public sector IT administrators need AI literacy to connect agents to municipal and national systems with narrow access, keep the decision record complete for the algorithm register and the AI Act, and read refusals correctly. Privileged access and incident calls stay with them. Articles 4 and 26 of the EU AI Act, GDPR and NIS2 frame the duties.
What do IT administrators in Dutch public bodies spend time on?
Municipalities carry a heavy load. They run systems for the population register, social support and benefits under national schemes, youth care, permits, taxes and finance, mostly delivered by a small number of specialist suppliers and connected to national registries. Administrators manage accounts, privileges, integrations with those registries, endpoint security and supplier access. Water boards, provinces and national agencies have similar duties at different scale.
Security is governed by a shared government information security baseline, and incidents at public bodies have made phishing, ransomware and supplier compromise standing concerns. The Netherlands has also learned, at real cost, what happens when risk scoring of citizens runs without proper oversight, and that history shapes how any new automated system in government is received.
Where do agents help, and at what depth?
Four routine tasks show how agents fit, each held to a depth earned on the body's own cases:
| Use | Depth | What stays with the administrator |
|---|---|---|
| Privilege reviews against current roles | 2 | Approving each removal |
| Supplier session monitoring | 1 | Acting on flagged sessions |
| Helpdesk triage and draft answers in Dutch | 2 | Sending or escalating |
| Registry interface error grouping | 1 | Fixing the interface |
Administrators also connect business agents, for example one that drafts replies to residents about permit applications, and decide what each may read. Our overview of what AI agent training organisations in the Netherlands need places IT within the wider programme.
What must IT administrators learn, and what stays with them?
The first skill is keeping people in the decision. An agent that drafts a reply about a social support application must not be connected in a way that lets it decide eligibility. Every action meets five bounds, role scope, certificate, sources, the body's own limits and the platform guardrails, and the lowest wins. IT sets role scope; that is where an administrator prevents a drafting agent from quietly becoming a deciding one.
The second is the record. Each verdict is filed in an append-only record with the bound that decided and the sources used. Administrators keep that record complete and retrievable, because it is what the body uses to describe the system in the public algorithm register, to answer a resident's question, and to show human oversight to a supervisor.
The third is reading refusals. A refusal is a control working. If an agent is refused because its certificate lapsed (agent certificates lapse after 90 days) or a failed Attest run dropped it to depth 1, the response is to route work to people and inform the owner, not to widen access.
What stays with the administrator: privileged access, approval of new connections to national registries, supplier access, incident declaration under NIS2, and any change widening an agent's reach into residents' data. Pryme Intelligence Academy's Foundations track gives the shared vocabulary for these conversations, free.
Which Dutch rules shape oversight and literacy?
The EU AI Act applies directly. Article 4 requires deployers to ensure AI literacy among staff who operate AI systems, and our explainer on Article 4 in the public sector covers what that means. For high-risk uses, including some public uses such as decisions on access to public benefits, Article 26 sets deployer duties such as assigning oversight to competent people and keeping logs; our piece on Article 26 deployer duties in the public sector explains them. Public bodies must also complete the Article 27 fundamental rights impact assessment before deploying a high-risk system.
GDPR governs residents' data: Article 22 on solely automated decisions, Article 30 records and Article 35 impact assessments. The Dutch data protection authority supervises and also has a coordinating role in algorithm oversight. NIS2 applies cybersecurity risk management and incident reporting to essential and important entities, including public administration. Government bodies are also expected to publish information about their algorithms in the national register. De Nederlandsche Bank and the AFM supervise financial services and do not govern a municipality's IT.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
A 90-day training plan
Days 1 to 30. All IT staff working on agents cover depth, the minimum chain, provenance and verdicts, with completion recorded for Article 4. The IT manager inventories agents and their register entries.
Days 31 to 60. Two administrators pilot privilege reviews at depth 2 and rehearse an incident, including the NIS2 reporting decision.
Days 61 to 90. Check that no drafting agent can decide, confirm the record supports each register entry, and set certificate expiry alerts.
Questions people ask
Does every agent need an entry in the algorithm register?
Government guidance expects publication of algorithms with impact on residents. Check the criteria with your register coordinator for each agent.
Is IT responsible for the fundamental rights impact assessment?
The deploying body is. IT contributes facts about access, logging and connections.
Why does IT matter so much for public sector agents?
Because scope and record-keeping are set in IT. As we argue, the public sector will not scale AI agents without trained reviewers, and the guide to AI agent skills by role shows each role's part, alongside which roles in the public sector need AI agent training.