Public-sector casework with AI under transparency rules

Public administration in Norway operates under a particular obligation to be auditable. Casework has to be traceable, decisions have to be explainable, and freedom of information is the default position. When AI agents enter the flow, the question isn't whether they help. It's how to capture the help without undermining what's the foundation of public service.

The first thing to understand is that transparency law doesn't stand in the way of AI. It sets the frame, and frames are something we can build for. What actually stands in the way is loosely thought-out automation where agents take decisions nobody can explain. That's a problem with or without AI. It's just easier to make that mistake with AI if the flow lacks discipline.

Three principles

Three principles hold the structure up.

The first is that the agent does explicit prep work, not decisions. Classifying incoming applications, pulling relevant law and regulation, surfacing prior similar cases, drafting decisions. All of that is prep work the caseworker has traditionally done manually. The decision itself stays with the caseworker or the delegated authority, taken with proper authority and documented reasoning.

The second is that every agent action has to be traceable. Which source did the agent read? Which legal basis did it cite? Which prior case did it identify as relevant? Which draft did it generate? If a transparency request comes in, we should be able to surface this just as reliably as we surface the caseworker's own notes. This isn't an IT function. It's an archival function, which means the agent's work has to be journaled with the same seriousness as human casework.

The third is that the citizen should be able to understand that AI was involved, where in the flow, and what that means. This sits under administrative law generally and becomes more concrete under the EU AI Act. If a draft decision is agent-generated, that should be marked clearly. If a caseworker has overridden, that should be marked clearly. The citizen has the right to know that human judgement has taken place, and where.

Where the gains land

When these principles are in place, real gains open up.

Application classification is the most immediate. An agent can read incoming applications, identify the case type, the applicable regulation, and the missing attachments. It can send the citizen a precise message about what's missing within minutes, not weeks. That falls within good administrative practice, and it improves the citizen's experience meaningfully.

Decision drafts are more demanding, but more valuable. On standardised case types with clear regulation, the agent can draft decisions including facts, legal basis, assessment and conclusion. The caseworker reads the draft, checks against their own assessment, corrects where necessary, and signs. On similar cases where 80 percent of the work is known, cycle time drops by a factor of four or five. Caseworkers gain time to concentrate on the cases where the assessment is genuinely difficult.

Citizen inbox triage is undervalued. State agencies and municipalities receive thousands of inquiries. Many could be answered faster if they were just classified correctly from the start and routed to the right caseworker. An agent can read each inquiry, classify it, identify which other cases from the same citizen are active, and propose which unit should handle it. If the inquiry is a follow-up on an active case, it attaches automatically.

Equal treatment

Equal treatment is traditionally the hardest principle to hold in public casework. Humans aren't consistent. Agents can be consistent in a way that fulfils the equal-treatment principle better than current practice, but only if the policy that drives them is open and controlled. An agent classifying applications based on a vague prompt isn't equal treatment. An agent classifying based on a published, versioned policy is.

Privacy and accountability

Privacy doesn't require new legislation. It requires strict discipline on data residency and sharing. Case data should never go to general model services with unclear retention. Storage inside the EU/EEA is not nice-to-have. Logging on every lookup, inference and decision is part of journalling. Retention regimes follow the archival law, not the vendor's defaults.

What about accountability? Administrative law places responsibility with the deciding authority. The agent's draft is a working tool, not a decision document. The caseworker owns the assessment, and the line manager owns the delegation. None of this is new. It's just that the flow has to make clear which hat the caseworker had on at which point.

Practical entry

Practical recommendation for a municipality or state agency considering this in 2026. Pick a case type with high volume, clear regulation and low conflict risk. Common examples are parking exemptions, building applications under a certain size, social benefit renewals with standardised assessment. Build the flow with journaled prep work, human decision and an open log. Measure cycle time and rights-of-citizen indicators for six months before extending.

AI in public casework shouldn't change administration's foundational position. It should reinforce it. If the flow is built right, it delivers citizen-facing speed without compromising equal treatment, traceability or transparency. That's a concrete project for 2026, not a vision.