Finland and Estonia Want an EU System for Responding to Serious AI Incidents

Finland and Estonia want the European Union to be ready for the day an AI system causes serious harm across borders. In a joint letter sent on November 11, 2026, ahead of the first European Council meeting devoted to artificial intelligence, the two governments called for “EU-level arrangements for responding to serious AI incidents” and for Europe to build “independent capacity to evaluate the most advanced models and their safety and security risks.” The letter is a proposal for leaders to discuss, not a legislative text, but it points to a gap that compliance and security teams already feel: the EU has rules for reporting AI incidents, and no shared playbook for responding to one.

What Finland and Estonia proposed

According to Politico Europe, which saw the letter, the two countries argue that the bloc “can still compete at the frontier” of AI if it mobilises capital and builds up computing power, but that this has to go hand in hand with “governing it responsibly.” The letter asks for three things in particular:

• An incident response arrangement at EU level. The countries want a coordinated way to react when a serious AI incident happens, rather than leaving each member state to improvise.
• Independent model evaluation. Europe should be able to assess the safety and security risks of the most capable models itself, independently of the companies that build them.
• Investment in European frontier models. The letter describes building European frontier AI “as an economic priority and as a foundation for Europe’s influence on the future of AI,” with better financing conditions for scaling companies.

Some early summaries described the authors as “officials.” In fact the letter came from the two governments and was addressed to the European Commission and the other member states. Its full text has not been published, so details such as how a response system would be staffed, funded or triggered are not yet known.

Timing matters. On Friday, November 13, EU heads of state or government meet informally in Dublin for their first dedicated debate on AI, with competitiveness, sovereignty, safety and security on the agenda. Letters like this one are how member states try to shape the conclusions that come out of such meetings.

Hungary’s earlier push, and the wider context

Finland and Estonia are not the first to raise the issue. In an October paper, also reported by Politico, Hungary proposed a network of national contact points to activate if AI model providers lose control of their models, and a pool of “pre-vetted, security-cleared” independent evaluators who could be ready before an incident occurs. The two proposals overlap: both want the EU to have people and procedures in place before something goes wrong.

The push also follows a run of events that made AI security feel less theoretical. Reports this summer described AI agents acting outside their intended boundaries and breaking into other companies’ systems, and the UK’s cyber agency recently highlighted AI’s growing role in the nationally significant incidents it handled. Researchers have also warned about the way AI agents can game the objectives they are given, which is exactly the kind of behaviour that is hard to anticipate in testing.

Internationally, the letter fits a pattern. In September, Finnish President Alexander Stubb and Norwegian Prime Minister Jonas Gahr Støre launched an international appeal for better oversight of frontier models, which Estonia supported from the start. That appeal asks governments and regional organisations to coordinate standards and strengthen transparency, “including shared reporting of serious safety incidents.” Governments and AI developers have also signed the Paris Call for AI Safety and Security, which covers frontier model testing and incident handling. The Finnish-Estonian letter tries to turn that kind of commitment into an EU operating capability.

What the AI Act already requires

The EU is not starting from zero. The AI Act already contains two incident-reporting tracks, and understanding them shows what is missing.

High-risk AI systems. Article 73 of the AI Act requires providers of high-risk systems to report serious incidents to the market surveillance authority of the member state where the incident occurred. The report is due immediately once a causal link, or a reasonable likelihood of one, is established, and no later than 15 days after the provider or deployer becomes aware. Shorter deadlines apply to a widespread infringement or serious disruption of critical infrastructure (two days) and to a death (ten days). Authorities must act within seven days and notify the Commission. Following the Digital Omnibus on AI, most stand-alone high-risk obligations, including this one, now apply from December 2027.

General-purpose AI models with systemic risk. Under Article 55, providers of the most capable general-purpose models must track, document and report serious incidents and possible corrective measures to the AI Office without undue delay. These obligations have applied since August 2025, and the Commission’s enforcement powers over model providers began in August 2026. The voluntary General-Purpose AI Code of Practice sets staggered reporting windows depending on severity.

Both tracks are about notification: getting information to a regulator. Neither creates an operational team that can coordinate across 27 countries when an incident spreads, test a model independently in a crisis, or tell businesses what to do in the first hours. That is the gap Finland and Estonia appear to be targeting. Cybersecurity offers a template, since NIS2 already links national incident response teams through EU-level networks, and an AI equivalent could borrow that structure.

What this means for companies using AI in Europe

For now, the letter changes no legal obligations. If leaders take up the idea, the most likely next step is a request to the Commission to explore options, possibly building on the AI Office. Any new reporting duty would need legislation and years of work. Still, the direction is clear, and organisations that deploy AI can prepare without waiting.

1. Define what an “AI incident” means internally. Agree on criteria covering harmful outputs, model misuse, data leakage through prompts, unauthorised agent actions and failures of human oversight, and map them to the AI Act’s “serious incident” definition.
2. Connect AI incidents to existing breach processes. Many AI failures are also personal data breaches with GDPR’s 72-hour notification clock. A single intake record, like the approach we described for the updated CISA reporting form, avoids parallel and inconsistent reports.
3. Know who your providers report to. If you rely on a general-purpose model with systemic risk, its provider reports to the AI Office. Ask how they will inform downstream deployers, and put that in the contract.
4. Keep logs that support investigation. Article 73 bars providers from altering a system in ways that would affect later evaluation before informing authorities. Deployers need logs and version records good enough to reconstruct what happened.
5. Rehearse. Add an AI scenario, such as an agent exfiltrating data or a model producing defamatory content at scale, to your next tabletop exercise.

Teams building this out will find that AI governance is moving faster than most compliance stacks, so incident handling is a sensible place to start.

Open questions

• Who would run it? The AI Office, ENISA, national authorities or a new body are all possible homes, each with different powers.
• What counts as serious? An EU response system would need thresholds that go beyond the AI Act’s definitions, especially for loss-of-control scenarios Hungary raised.
• How independent is “independent”? Evaluating frontier models requires access to weights or deep API access, which providers outside the EU may resist.
• How does it fit with existing duties? Companies already juggle GDPR, NIS2, DORA and AI Act reporting. A new layer must reduce confusion, not add to it.

Frequently asked questions

Is the Finnish-Estonian proposal law?

No. It is a letter to the Commission and other member states. It could influence European Council conclusions, but any binding mechanism would require further proposals and legislation.

Do companies have to report AI incidents today?

Providers of general-purpose AI models with systemic risk already report serious incidents to the AI Office. Article 73 reporting for high-risk systems applies once high-risk obligations take effect, from December 2027 for most stand-alone systems. Personal data breaches remain notifiable under GDPR regardless.

How is this different from the Paris Call?

The Paris Call is a set of voluntary commitments by governments and developers. Finland and Estonia are asking the EU to build its own operating capacity to respond and to evaluate models independently.

Would a response system cover AI agents?

The letter’s details are not public, but incidents involving autonomous agents are one of the main concerns behind it, so agentic systems would very likely be in scope.

From reporting to responding

Europe has spent years deciding who must tell regulators about AI failures. Finland and Estonia are asking the next question: once the report arrives, who acts? Leaders may not answer it in Dublin, but companies that already treat AI incidents as a defined, rehearsed process will be ready for whatever structure follows.