Air Canada's Chatbot Invented a Refund Policy — and a Tribunal Made the Airline Pay
Summary
Air Canada's website chatbot told a grieving customer he could apply for a discounted bereavement fare retroactively, within 90 days of booking. That policy did not exist — the airline's actual policy prohibited retroactive bereavement claims. The customer booked at full fare relying on the chatbot, was refused the refund, and took Air Canada to the British Columbia Civil Resolution Tribunal. The tribunal rejected the airline's argument that the chatbot was a separate entity and held Air Canada liable for negligent misrepresentation.
Severity
SEV-1 — A tribunal held the company legally liable and awarded damages; the failure produced a binding legal and financial consequence, not just reputational harm.
What happened
- Customer's grandmother dies; he visits Air Canada's site to book funeral travel.
- He asks the chatbot about bereavement fares.
- The chatbot states he can claim the discount retroactively within 90 days of ticket issue — contradicting the airline's real policy, which was linked on a separate page.
- Relying on this, he books two last-minute flights at full fare.
- He submits a refund request within the stated 90 days; Air Canada refuses, citing its actual (no-retroactive) policy.
- He files with the BC Civil Resolution Tribunal.
- The tribunal finds negligent misrepresentation and orders Air Canada to pay C$812.02 in damages and fees.
Root cause
Primary: A generative chatbot was allowed to author authoritative-sounding policy statements to customers with no grounding constraint tying its answers to the airline's actual, verified policy text.
Contributing factors:
- The correct policy existed only on a separate linked page; the chatbot was free to paraphrase (and contradict) it rather than being forced to quote it.
- No human review or citation requirement on customer-facing policy claims that carry financial consequences.
- The company assumed the chatbot's output was informational rather than a representation it could be held to.
Governance analysis
The control that would have prevented it
Ground all policy answers in retrieval: the agent may only state policy by quoting or directly citing the canonical policy document, never by paraphrasing from memory.
Would have: The chatbot would have surfaced the real bereavement policy text (no retroactive claims) instead of generating a plausible but false alternative.
Constrain the agent to link-and-quote for any answer that involves money, eligibility, or a deadline — no free-text policy generation on financially consequential topics.
Would have: The specific class of answer that caused the harm (a fabricated refund deadline) would have been structurally impossible to produce.
Log every policy claim the agent makes and sample-review them against source policy.
Would have: The contradiction between the chatbot's answer and the real policy would have surfaced in review long before it reached a tribunal.
Why it wasn't caught
There was no monitoring signal comparing what the chatbot told customers against the airline's actual policy. The divergence was only discovered when a harmed customer escalated to a tribunal — the most expensive possible detection path.
Recurrence risk
HIGH for any organization running a generative support agent that can paraphrase policy from memory rather than being forced to quote a grounded source.
Sources
Check your own agent against this failure pattern before it ships: