Market study
The risk is documented. The price is not.
European courts and regulators have already ruled on disputes caused by AI systems. Almost none has made public what it cost.
Since 2018, across nine European countries, companies have been sentenced, fined or sued over a decision taken or a piece of content produced by their AI system. Professional liability, data protection, employment law, property damage: these are not textbook scenarios, they are decided cases.
The heaviest penalty we have on record stands at EUR 30.5 million. It was not handed down by a civil court, but by a data protection authority. The first creditor of an agent gone wrong is not always the customer who suffered.
For most of the other matters, the amount has never been made public. Settlements close without a press release, proceedings are still running, and indirect losses — lost traffic, reputation, management time — are quantified nowhere.
That asymmetry is exactly what leaves you without cover. Your insurer can observe frequency; it cannot establish severity. Faced with a risk whose scale escapes it, its only prudent answer is exclusion: which is what happens, renewal after renewal.
Getting out of it does not mean waiting ten years for case law. It means measuring exposure upfront, agent by agent: what it can commit to, what it can decide on its own, what it can publish unreviewed. A measured risk becomes a priceable risk again.
Source: European Agentic AI Loss Database — Garenzia's compilation of public European claims 2018-2026, from the OECD AIM screening and the AI Incident Database.