AI performance insurance

A company buying an AI model to detect fraud, sort documents or answer customers relies on performance measured in testing. Nothing guarantees it will hold in production: data change, the model drifts, rare cases arise. Classic liability insurance covers this risk poorly, as it results neither from proven fault nor from physical damage. Specialist insurers have developed dedicated covers since the late 2010s. The provider commits contractually to a performance level, for example a maximum false negative rate; the insurer, after assessing the model, its test data and its governance, guarantees this commitment, and pays the gap when measured real-world performance falls below the threshold. Other forms directly cover the user against losses caused by a chatbot's errors. The product makes insurable what was not, and serves as a sales argument for the provider, who can stand behind what it sells. Its central difficulty is measurement: it requires an objective metric, an indisputable test protocol and an observation period, failing which every claim becomes a debate among experts. It also requires the ability to assess a model before insuring it, a rare skill among underwriters.