Air Canada and the Chatbot That Made a Promise the Company Wouldn't Keep

How a customer-service bot invented a refund policy, and a tribunal made the airline pay for it.

In November 2022, Jake Moffatt's grandmother died, and he went to Air Canada's website to book flights to the funeral. Like a lot of travelers, he interacted with the airline's support chatbot before buying. He asked about bereavement fares, the discounted rates airlines sometimes offer the recently grieving. The chatbot told him he could book his flights at full price and then apply for the bereavement discount retroactively, within ninety days, by submitting a claim.

So that is what he did. He booked. He applied. And Air Canada refused, because the airline's actual policy was that bereavement discounts cannot be claimed after the fact. The chatbot had been flatly, confidently wrong. It had described a policy that did not exist.

Here is the part every business owner should sit with. When Moffatt took it to a Canadian tribunal, Air Canada's defense was, essentially, that the chatbot was a separate legal entity responsible for its own actions. The airline argued it should not be held to what its own bot said. The tribunal rejected that flatly. In February 2024 it ruled that Air Canada was responsible for all the information on its website, whether it came from a static page or a chatbot, and the British Columbia Civil Resolution Tribunal ordered Air Canada to pay him $650.88 in damages (about $812 with interest and fees). The company later quietly took the chatbot down.

The root failure was not really a software bug. It was a worldview. Air Canada deployed a system that could generate plausible answers about its policies without those answers being anchored to its actual policies, and then treated the output as someone else's problem. The bot was built to sound authoritative. It was not built to be correct. Those are completely different design goals, and the gap between them is exactly where the company got hurt.

A customer-facing AI is not a novelty bolted onto the side of your business. Legally and practically, it is your business, speaking. Whatever it says, you said. The tribunal made that principle explicit, but it was always true. Your customers never thought of the bot as a separate entity. Why would the law.

And notice the specific way it failed. It did not crash. It did not say "I don't know." It made something up and delivered it with total confidence. That is the single most dangerous failure mode an AI can have in front of a customer, because a confident wrong answer is worse than no answer at all. No answer sends the customer looking. A wrong answer sends them acting.

The architecture that makes this specific failure impossible

Here is the encouraging part: the Air Canada failure is not an inherent property of AI. It is a property of that design, and it can be engineered out.

The fix is to stop letting the model speak from its general impression of how a business works, and force it to answer only from the business's own verified material. A grounded assistant retrieves the relevant policy and answers from it; when nothing in the source covers the question, it is required to disclose that and can hand off to a human, rather than invent a plausible-sounding answer. That single rule, say what you can source, admit what you cannot, is the difference between Air Canada's bot and a safe one.

That is the core of how the instantAIguru Guru is built. Its Hybrid RAG pipeline anchors every answer in your indexed content, validated at 97%+ accuracy across more than 100,000 monthly interactions in production, and by architectural rule it never fabricates a policy, a price, or a promise that is not in your source. When it has no grounded answer, it says so and routes to a person, the exact behavior the airline's bot lacked.

And there is a second layer the Air Canada bot never had, for the cases where there is no acceptable error rate at all. Not every answer is general knowledge to be paraphrased. Some are exact facts that must be read live from a system of record, a real fare, today's schedule, a seat's availability, an account balance, available credit, and some are actions, applying a fare adjustment, issuing a refund, filing a claim. Neither is left to the AI's discretion. Both run through the JavaScript Flow Engine, conventional scripted code that reads the true value or commits the real action, with the AI removed from the path. So the protection works at three levels: the grounded conversational layer will not invent a policy; the deterministic lookup returns the actual number rather than a 97%-confident guess; and the action layer will not commit something you never authorized. Across more than 200,000 flows, including over 100,000 phone calls in production, those completed exactly as scripted, not 97 percent of the time, but deterministically.

The "whatever it says, you said" principle is not a reason to fear customer-facing AI. It is a reason to insist on AI that cannot make a promise you never made. For the engineering behind that guarantee, see Agentic AI Can Be Hallucination-Proof.