Agentic AI in a country with no AI law

Norwegian companies are building AI agents in a regulatory gap.

The EU AI Act entered into force on 1 August 2024 and is being phased in through 2028. Since 2 August 2026, its transparency rules have applied across the EU: people should be able to tell when they are talking to a machine, and AI-generated content must be marked in a machine-readable way.

In Norway, none of that applies yet. The regulation has been assessed as EEA-relevant, but as of September 2026 it has still not been incorporated into the EEA Agreement. The government circulated a draft Norwegian AI act for consultation, originally aiming for entry into force in late summer 2026. That plan has slipped. Amendments following the EU's simplification package go out for consultation in autumn 2026, with an ambition to present a bill in spring 2027 — and progress depends on EEA adaptation talks that nobody has put a date on.

It is worth being precise about what kind of gap this is. It is not a free pass. It is a lag.

Why the lag doesn't help you

Three reasons Norwegian companies should build as though the rules already applied.

The market decides, not your address. The AI Act follows the product. If you place an AI system on the EU market, or the output of your system is used in the EU, you are covered regardless of where the company is registered. For Norwegian companies — where scaling in practice means internationalising — this is not an edge case. It is the normal case.

Your customers aren't waiting. Large Nordic buyers, and public sector bodies in particular, have started writing AI requirements into procurement ahead of the law. A tender does not ask about the EEA process. It spends one row in the requirements matrix, and you either have the documentation or you don't.

Retrofitting is expensive. Logging what an agent did, which data it saw and who approved what is cheap to build in and costly to reconstruct. Existing law — GDPR, copyright law, criminal law, and public administration law for government bodies — applies today regardless. The Norwegian AI act will arrive on top of that, not instead of it.

What actually makes a system agentic

An ordinary language model takes text and gives text back. An agentic system gets two more things: access to tools, and permission to decide the order itself.

The second is the real shift. The model stops being a function someone calls and becomes a process that runs until it believes the task is done: it reads the situation, picks an action, carries it out, reads the result and decides the next step. The loop ends when the goal is met, when the attempts run out, or when a human intervenes.

Most of your risk management lives in that third stopping condition.

Norwegian is a small language, and agents notice

Something international writing on agentic AI rarely mentions: models are noticeably weaker in Norwegian than in English, and the gap widens the further into the tail you go.

Bokmål is mostly fine. Nynorsk is thinner. Spoken dialects are worse still — and voice agents are exactly what many Norwegian companies consider first, because the phone is still a primary channel into customer service here. An agent that understands Oslo bokmål beautifully and Sunnmøre dialect poorly delivers a service that is unevenly good in a way users experience as unfair.

The same applies to domain vocabulary. Terms from Altinn, the labour and welfare administration, the population register, healthcare and Norwegian accounting are not uniformly badly covered — they are unevenly covered, which is worse, because the failures arrive without warning.

The practical consequence: build a test set in your own Norwegian domain language before choosing a model. Twenty real customer enquiries with known correct answers tell you more about what will work for you than any international benchmark.

The advantage we tend to overlook

At the same time, Norway has something most markets lack: infrastructure an agent can actually act in.

BankID gives strong authentication that nearly everyone has. Altinn provides an established API layer into government services. Vipps covers payments across the population. The population register and the Brønnøysund business registers are structured and available. Digital maturity is high enough that few users need persuading to use a digital channel.

This lowers the bar for building agents that do something rather than merely answer something. But the same property makes mistakes more expensive: an agent with write access to real registries and payment systems has a shorter path from bad judgement to consequence than one that can only suggest text.

That is why the boundary between read and write operations is the simplest and most underrated safety mechanism available. Let the agent read freely. Require approval for anything that sends, pays, deletes or changes data outside the system.

The economics are different here

American advice on agentic AI is written for a market with very large volumes and relatively lower labour costs in the roles being automated. The Norwegian context inverts both.

Wage levels mean that even modest time savings on routine work pay back faster than in most other countries. But volumes are smaller. A mid-sized Norwegian company might handle two hundred enquiries a week, not two hundred thousand. Solutions that demand heavy investment before any value appears rarely justify themselves — and off-the-shelf models via API almost always beat training your own.

It also means human review is cheaper than people assume. At two hundred cases a week, draft-and-approve is a perfectly reasonable end state, not a compromise on the way to full autonomy.

Start where failure is cheap

The most common mistake is starting with the most impressive task. Pick something where an error costs little and surfaces quickly instead.

Good first candidates share three traits. The task repeats often enough for automation to pay. The result is easy to verify. And a human sees the outcome before it has consequences.

In operation, agents also fail differently from ordinary software. They rarely crash — they do something plausible that is wrong, and do it with confidence. Uptime and response time won't catch that. You need traces of every step the agent took, which tools it called and with which arguments, and a way to replay a conversation that went wrong.

That log is not just for debugging. It is precisely the documentation both the AI Act and a public buyer's requirements form will ask for. Building it now is the cheapest available answer to a question that is coming anyway.

In short

The technology is mature enough for real workflows in bounded areas with clear limits. It is not mature enough to be let loose on an entire domain unsupervised.

Norway's regulatory lag buys breathing room, not an exemption. Companies that spend it on logging, clear boundaries and Norwegian-language test sets will meet the AI act as a formality. Companies that spend it postponing will meet it as a project.