When the chatbot's answer became the airline's problem

What happened

On November 11, 2022, Jake Moffatt used the Air Canada website to ask about bereavement fares after the death of a grandmother. According to the tribunal's decision as reported, the chatbot told Moffatt that he could apply for the bereavement rate after travelling, and linked to a separate page about the policy. That page said the opposite: bereavement fares could not be applied for retroactively.

Moffatt booked at the regular fare and later applied for a partial refund. Air Canada refused, citing its policy. Moffatt took the dispute to the British Columbia Civil Resolution Tribunal.

Air Canada's defense, as described in reporting on the decision, included arguing that it could not be held liable for information from a chatbot, which it characterized as a separate entity, and that the correct policy was available elsewhere on its site. The tribunal rejected those arguments. It found that Air Canada owed a duty of care, that it "did not take reasonable care to ensure its chatbot was accurate," and that the chatbot's error amounted to negligent misrepresentation. It also questioned why a customer should have to double-check one part of the website against another.

Moffatt was awarded CAD $650.88 in damages, the difference between the fare paid and the bereavement fare, plus pre-judgment interest and tribunal fees.

Why it happened: a design reading

We don't know how Air Canada's chatbot was built, so we can't say exactly what failed technically. We can read the design from the outside:

• No single source of truth reached the user. Two parts of one website disagreed, and the assistant gave the wrong one.
• No grounding or citation check. The answer wasn't verified against the authoritative policy page, even though the chatbot linked to it.
• A high-stakes moment treated like routine Q&A. A customer asking about bereavement fares is often grieving, cost-sensitive and time-pressured. A refund policy with exceptions is exactly where invention is costly.
• No safe path for exceptions. Fare exceptions and refunds are areas where an assistant should show the exact policy and offer a person, not paraphrase.

Design lessons

1. Answer policy questions from approved text, with citations. The reply should quote or closely track the policy and link to it, so the user can verify (module 6).
2. Contradiction checks. Where the assistant's answer and the policy page can diverge, test for it continuously (module 14).
3. Slow down at moments that matter. For refunds, exceptions and bereavement, narrow the assistant's freedom: show the policy, state the conditions plainly, offer a human (module 3).
4. Don't count on disclaimers. "The chatbot may be wrong" doesn't help a customer who followed its advice, and the tribunal didn't accept the argument that users should cross-check.
5. You own what it says (module 18).

How to build it better

A grounded design for this question might work like this:

The key moves: retrieve the current policy, answer only from it, cite it, and offer a person for a sensitive, exception-prone case.

Try it

Write five test cases for a refund-policy assistant (module 14 shows the format). Include one where the user asks about an exception the policy doesn't allow, one where two pages disagree, and one where the user is upset. For each, write the expected behavior rather than the exact words.