Blog
Tech E&O

When your AI hallucinates, who pays?

Truly Insurance July 9, 2026 7 min read
When your AI hallucinates, who pays?

The Air Canada chatbot case is now precedent Ontario tribunals cite. If you ship an LLM feature, your Tech E&O needs an AI endorsement - not a hope.

Why the Air Canada case still matters

A Canadian tribunal held Air Canada liable for commitments its AI assistant invented - and rejected the 'the chatbot is a separate entity' defence outright. Ontario courts have since cited the reasoning. The rule for founders: your model output is your output.

That makes hallucination a financial-loss claim dressed as a customer-service problem. The customer didn't get the price the bot promised and now sues for the difference plus damages - that is precisely the economic-loss lane Tech E&O exists to defend.

The gap in a standard Tech E&O policy

An unendorsed Tech E&O policy was built around 'your software failed to perform as contracted.' Many will argue that a model producing a wrong answer is not a software 'failure' - the model performed as designed, it was just wrong. Without an AI endorsement, the carrier has a coverage position to deny the hallucination claim.

Training-data IP is the second gap. If your model was trained on copyrighted text or images, the infringement claim may sit outside your E&O's intellectual-property carve-back entirely.

Talk to an Advisor

Reading is the start. A short conversation turns it into a plan.

No sales pressure. We'll look at your stage, your contracts, and your AI exposure, and tell you - plainly - where your current coverage won't hold up.

Newsletter

Risk notes, once a month.

One short note on Ontario tech risk - breaches, regulation, and what your policy should actually cover.

What an AI E&O endorsement does

A proper AI endorsement affirms coverage for losses caused by model output, extends to training-data IP infringement, and names hallucination as a covered cause of loss. It also clarifies that 'professional services' includes the design, training, and deployment of AI - closing the 'it just worked as designed' denial.

Underwriters price this on whether you've shipped, or only experimented. Get it bound before the LLM feature goes live, not after the first complaint.