Property damage
When the agent has a body.
On our autonomy scale, level 5 denotes a system that acts physically. Two cases in the dataset reach it. They are also the two most serious.
The first: a collision between three warehouse robots in the United Kingdom, in 2021, which started a fire. Property damage and business interruption. No injuries reported, no public litigation on record.
The second: a driver assistance system, in France and across Europe, whose speed behaviour gave rise to a regulatory proceeding. Here the human stays in the loop — the driver supervises — but the machine acts.
These two files are a minority in the dataset: two cases out of seventeen. Yet they are the only ones where the damage is physical, immediate and visible. All the others produce intangible loss: a reputation, a right, a piece of data, a piece of content.
This asymmetry explains why insurance lags behind. The market knows how to cover a warehouse fire: it has been doing so for two centuries. It does not yet know how to cover a news article fabricated from nothing by a machine, or a dismissal notice sent in error.
Agent risk is therefore not one more industrial risk. It is intangible risk at scale, whose insurance is still to be built — and it begins with a measurement.
Source: European Agentic AI Loss Database, compiled by Garenzia from the public OECD AIM screening and the AI Incident Database.