In Depth
Traditional software liability turns on whether a product worked as specified. AI output liability turns on something messier: whether what the system said was wrong enough, relied on enough, and harmful enough to support a claim. The agent didn't crash. It answered. The answer sounded authoritative and was false, and a customer acted on it.
That shift breaks the assumptions most insurance was written under. Tech E&O contemplates a product failing to perform. Cyber contemplates an intruder. Neither anticipated a working system generating its own harmful content with no attacker and no outage. Courts, meanwhile, have been clear about who owns the output: the business that deployed the agent. A company cannot hand a customer an AI and then disclaim what the AI tells them. The output is the company's speech.
Exposure concentrates wherever an agent's words carry weight. A support agent that quotes policy, a financial assistant that states a balance or a term, a legal tool that cites authority, a clinical scribe that records a symptom: each of these turns a single confident error into a chain of reliance and harm. The same wrong sentence is trivial in a brainstorming toy and catastrophic in a system of record. Output liability is what sits underneath hallucination claims, misrepresentation claims, and the negligence theory a plaintiff reaches for when an agent's statement causes a loss.
What It Looks Like
In Moffatt v. Air Canada (2024 BCCRT 149), the airline's support chatbot told a customer he could claim a bereavement discount retroactively. No such policy existed. When the airline refused to honor it, the customer brought the case to a British Columbia tribunal. Air Canada argued the chatbot was a separate entity responsible for its own statements. The tribunal rejected that outright and held the airline liable for what its bot said. The dollar figure was small; the precedent was not. It is now the cleanest statement of the rule that an agent's output is the vendor's output.
Move the same pattern into higher stakes. An AI agent for a lender states the wrong rate lock to an applicant who reorganizes their finances around it. A clinical-documentation agent records a dosage the clinician never ordered, and the next provider treats from the note. In each case the trigger is identical to Air Canada: the agent stated something false and someone relied on it. Only the size of the harm has changed.
Why It Matters For AI Vendors
Output liability is the failure mode your enterprise customers worry about first, because it is the one most likely to reach them. When your agent misstates something inside their workflow, the harmed party names them, and they turn to your contract's indemnity. The vendor becomes the backstop for every confident wrong answer the agent gives at their scale.
The trap is that the Tech E&O policy most vendors carry has been quietly rewritten to exclude exactly this. The AI carve-outs added through 2025–2026 were drafted to keep "the AI said something false and it caused harm" claims out of legacy coverage. A vendor can hold a clean certificate of insurance and still have no policy that answers when output liability lands.