Signal
The mistakes AI makes are becoming a priced risk, written out of ordinary policies and sold back as cover, failure by failure. In the United States the standard liability forms have excluded generative AI since January, and one insurer's exclusion names promises made by a chatbot. Under many contracts, lawyers warn, the customer rather than the supplier bears the risk of what an AI agent does. Into that gap a new market is moving. Lloyd's insurers write policies that list hallucinations and model drift as covered events. Munich Re's performance guarantees, now sold from Bermuda to South Africa, pay out when a model crosses a measured error threshold. A voice-agent company insures what its agents say to customers, and China's PICC covers model providers whose products miss their performance standards. Cover like this does for AI what the warranty did for machines: it turns a vague fear into a list of named failures with a price on each.
So what?
AI now comes with a list of the mistakes someone else will pay for. The list is becoming the product.
Evidence
Chaucer, 22 April 2025 (launch): Chaucer, a Lloyd's insurer, and Armilla launched a standalone liability policy for AI systems that underperform. It covers hallucinations, meaning false or misleading outputs, and model drift, meaning performance that degrades over time, and it is sold to policyholders in the United States with cover that applies worldwide.
ElevenLabs, 12 February 2026 (launch): ElevenLabs became the first company to activate insurance on its AI voice agents under the AIUC-1 standard, after the agents passed 5,835 tests across 14 risk categories. An agent giving a customer wrong information is now an insured event. The insurer behind the policy is not named.
Beijing Business Today, 28 September 2026 (launch): PICC Property and Casualty launched an AI product application insurance and signed the first policy in China. It covers model providers, development platforms and AI application vendors for the cost of adjusting, repairing, replacing or refunding an AI product that fails its performance standards, plus inspection and appraisal costs after an incident.
Worth asking
If a tool you use at work came with a list of the mistakes it is insured against, which mistake would you check for first?
