Who Pays When Your AI Agent Signs the Wrong Check?

A Canadian man asked an airline's chatbot about bereavement fares. The chatbot gave him a wrong answer. He won.

That's the whole case. Jake Moffatt asked Air Canada's website chatbot in 2022 about a bereavement discount. He had just booked a ticket to his grandmother's funeral. The chatbot told him he could apply anytime, even after travel. The airline's real policy said the opposite: apply before you travel. Air Canada refused to honor the chatbot's answer. It told a tribunal the chatbot was "a separate legal entity." Responsible for its own words, not the airline's. The tribunal did not buy it. In February 2024, it ordered Air Canada to pay Moffatt $812.02 CAD. Its line deserves a spot above every general counsel's desk: the chatbot is "still just a part of Air Canada's website." The airline owns everything it says there.

Every company now deploying an agent to book travel, approve a refund, or move money is one confused customer away from making Air Canada's argument. And losing it the same way.

Here's what's different this time. A chatbot misleading one customer is an $812 problem. An agent with a corporate card and standing permission to act is a bigger animal. In November 2025, Amazon sued Perplexity in federal court. The complaint: Perplexity's Comet browser agent logged into customer accounts and completed purchases on its own. It disguised itself as an ordinary Chrome browser. It kept transacting after Amazon tried to block it. By March 2026, Amazon had won a court order stopping Comet-enabled purchases on its site. Perplexity called the suit a bully tactic. It said Comet only followed user instructions. A court is deciding, for the first time, how far an agent can act for a person before that action becomes someone else's legal problem.

The insurance industry reached the tribunal's conclusion from the other side. Gallagher Re, working with MIT and Testudo, published a report in March 2026: "Smart Systems, Blind Spots: Rethinking Insurance for the AI Era." Its finding is plain. Old policies were never built for AI failures like made-up answers, bad automated calls, and models that drift off course over time. Courts and regulators, the report notes, are putting the bill on the company that deployed the AI. Not the vendor who built it. The vendors, for their part, cap their own risk. A standard vendor contract caps damages at 12 months of fees. Buy a $100,000 annual license, and $100,000 is the most you will ever get back. That's true no matter what the agent actually cost you. Insurers added new exclusions for AI claims to their 2026 policy forms. The safety net companies assumed was there has holes right where agents do their work.

Legal and procurement teams are already rewriting the paper. The new clauses reveal what they fear. Buyers now ask vendors for kill switches: the right to shut an agent down fast, in under 5 minutes for routine use, under 1 minute for anything touching money. They ask for audit trails that log every step an agent took, and which model version took it. A dispute should never come down to a vendor's word against a customer's. They ask for 30 to 60 days' notice before a vendor swaps the underlying model. An agent that passed testing in March can behave differently in June, on a model nobody approved for that job. None of this sat in a standard software agreement three years ago. It exists now because someone already paid for an agent's mistake, and decided never again.

The law settled this question before, for a different kind of worker. A telegraph clerk who quoted the wrong price. A switchboard operator who routed the wrong call. A clerk who filled out the wrong form for a stranger. Courts spent decades on cases like these, and built a simple rule: an employer owns its employee's mistakes. Businesses kept trying to hide behind whoever, or whatever, said the wrong thing. The answer held every time. If you hired the mouth, you own the words. Air Canada's tribunal simply applied an old rule to a new mouth. The chatbot's own lawyers lost the argument they were making.

Zoom out, and this stops being a legal problem. It becomes a boardroom one. Gartner expects more than 40% of agentic AI projects to get cancelled by 2027. Forrester puts a number on the risk piling up at companies with no rules in place: over $10 billion. That gap is a plain gap in ownership. Companies rolled out agents faster than they built the human-plus-agents stack, the layer that decides who watches an agent, who can stop it, and who answers when it's wrong. Skip that layer, and a company borrows speed against a bill. The bill comes due later, with interest, in front of a judge or a jury.

An agent does not get sued. The company that deployed it does.

Speed was never the risk here. Silence was. Somebody in your organization needs to be able to say, by name, who owns what an agent just did. A person, a title, a line in a contract that survives a bad afternoon. Not "the vendor." Not "the model." If you cannot name that person today for the agents already live in your business, schedule that meeting before next quarter, not after the next incident. I'm easy to find if you want to talk through how.

Sources: BC Civil Resolution Tribunal, Moffatt v. Air Canada, 2024 BCCRT 149 (Feb. 19, 2024) · CBC News, "How can I mislead you? Air Canada found liable for chatbot's bad advice on bereavement rates," Feb. 2024 · eWeek, "Amazon Sues Perplexity Over Comet, Its AI Shopping Agent," Nov. 5, 2025 · CNBC, "Amazon wins court order to block Perplexity's AI shopping agent," Mar. 10, 2026 · Gallagher Re with MIT and Testudo, "Smart Systems, Blind Spots: Rethinking Insurance for the AI Era," Mar. 24, 2026 · Honigman, "The AI Insurance Gap and What It Means for Technology Contracts," 2026 · Zylos Research, "Buyer-Side Governance: What Enterprise Customers Now Demand From AI Agent Vendors," Jul. 2, 2026.