Echo — Air Canada Chatbot (2024)

What happened

In November 2022, Jake Moffatt was booking a last-minute flight to attend his grandmother's funeral, and asked Air Canada's website chatbot about the airline's bereavement fares. The chatbot told him he could book at the regular price and then apply for the bereavement discount retroactively, within 90 days of the ticket being issued. He booked on that basis. When he later applied for the difference, Air Canada refused — the airline's actual bereavement policy, on a page the chatbot had itself linked to, said the reduced fare could not be claimed after travel was completed. Moffatt took the dispute to British Columbia's Civil Resolution Tribunal, and in February 2024 (Moffatt v. Air Canada, 2024 BCCRT 149) the tribunal ruled in his favour, awarding him $650.88 in damages for negligent misrepresentation.

The mechanism

The chatbot invented a refund window that does not exist. There was no retrieval step binding its answer to Air Canada's canonical policy: the model composed the most plausible-sounding continuation of the conversation and stated it in the same confident register it uses for facts. A retroactive bereavement discount is exactly the kind of policy that would plausibly exist — which is precisely why a next-token predictor produced one. Nothing errored, because nothing malfunctioned; generating fluent, policy-shaped text is the model's entire job. The contradiction sat in plain sight: the correct policy was one link away, on a page the bot itself surfaced, yet the composed sentence carried the same authority as the retrieved page. Air Canada then argued in the tribunal that the chatbot was "a separate legal entity that is responsible for its own actions." The tribunal rejected that outright — the bot is part of Air Canada's website, and the company is responsible for all the information on it, whether the words come from a static page or a chatbot.

Blast radius

In dollars the case was trivial: $650.88, the difference between the fare paid and the bereavement fare, plus tribunal fees and interest. The real blast radius was precedent and reputation. The ruling became a widely-cited reference point establishing that a company owns what its customer-facing bot says — the "separate legal entity" defence does not work, and apparent authority extends to automated channels. One composed sentence turned into a legal finding of negligent misrepresentation, global press coverage, and a warning that every organisation deploying a customer-facing bot now has to answer to.

The same-class lesson

Where a bot's words create obligations — prices, fees, eligibility, refunds, policy terms — the model must retrieve and quote from a canonical, owned source, or decline and route to a human. Cite or decline; there is no third path where it composes the answer itself. "Don't make up policies" in the system prompt is not a fix: hallucination is not disobedience, and the model carries no internal flag distinguishing a retrieved fact from a fluent invention. The guardrail has to be structural — a canonical policy store queried by retrieval, quote-with-citation on any obligation-shaped reply, and a commitment classifier that blocks or escalates sentences like "you're eligible…" or "fees are waived…" before they are ever sent.

flowchart TD
    A["Customer asks about<br/>bereavement fares"] --> B["Retrieve from the<br/>canonical policy store"]
    B -->|"policy found"| C["Quote it with a<br/>citation"]
    B -->|"nothing found"| D["Decline and route<br/>to a human"]
    C --> E["Commitment check on<br/>obligation-shaped output"]
    E -->|"informational"| F["Send the reply"]
    E -->|"creates an obligation"| G["Human review before<br/>anything is promised"]

Rhymes with

• Drill 17 — The Policy That Never Existed. A customer-facing assistant confidently states a fee-waiver policy that does not exist; legal's view is that the bot speaks for the company. Air Canada is the real, adjudicated instance of that exact shape — a composed sentence about an obligation, treated by a tribunal as the company's own word.

Named concepts: grounding/RAG with citation, output constraints, apparent authority / legal blast radius, graceful refusal.