An airline's website chatbot invented a ninety-day refund window. The policy page said the opposite. In February 2024 a tribunal awarded the customer C$812.02, called the "separate legal entity" defence remarkable, and found no reason a customer should know which part of a website to trust. The side that lost was the side with no evidence. Five checks.
The checks
Everything starts ticked. Untick what you will not do; the choice is saved on this device only, and Print keeps only what you kept.
What happened. The company could not show what its bot had been told or why its page should outrank the bot. A dispute is won by whoever can replay the conversation.
The check.
- Store, per answer: an id, the timestamp, the exact text shown, the sources used (with their version), the prompt version, the model id and the retrieval set.
- Keep it for the whole dispute window of your market, hashed as it is written, where the bot cannot change it.
- Replay from the log, never from memory, when a customer disputes an answer.
{"answer":"a81f","ts":"2026-10-16T09:12Z","shown":"...","sources":["policy/bereavement#v12"],
"prompt_v":"14","model":"pinned-2026-09","retrieved":["doc-2231","doc-0917"]}Source: legal-news report of the decision, 2024 BCCRT 149, 16 Feb 2024
What happened. The ninety-day window came from nowhere. A figure, a date or a rule that the bot cannot point to is a promise nobody authorised.
The check.
- Every answer that carries a number, a date or a policy rule quotes the page it came from, with a link.
- No source found means the bot says so and hands over; it never fills the gap from training data.
- Count answers with an empty sources list every day; the target is zero.
grep -c '"sources": \[\]' answers.jsonl # answers with no citation: want 0
Source: legal-news report of the decision, 2024 BCCRT 149, 16 Feb 2024
What happened. The tribunal found no reason a customer should know that one section of the website is accurate and another is not. Precedence has to be yours, written down, before a dispute tests it.
The check.
- The policy page is canonical; the bot's answer is derived from it and says so in its footer.
- When they disagree, the page wins and the bot is corrected inside a set time; log every correction.
- Put the precedence statement in the terms the customer accepts, and in the bot's own first message.
Source: legal-news report of the decision, 2024 BCCRT 149, 16 Feb 2024
What happened. Disputed answers that sit in a queue become claims. A person, a replay and a reply inside a fixed number of days turn them back into tickets.
The check.
- Route every disputed answer to a named owner with the replay attached; the deadline is in days, not 'soon'.
- Track the count and the time to reply as a release metric, alongside accuracy.
Source: AI Incident Database, incident 639 (the chatbot's answer, Nov 2022)
What happened. A process that has never been run is a process that does not exist. The airline found out in front of a tribunal.
The check.
- Take one real wrong answer from last month, run the replay, the source check and the reply, and time it end to end.
- If any step needs a person to search a chat channel for a screenshot, fix the log first.
Source: legal-news report of the decision, 2024 BCCRT 149, 16 Feb 2024
The decision is Moffatt v. the airline, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal, 14 February 2024): damages C$650.88, C$812.02 with interest and tribunal fees. The tribunal's site does not admit automated readers, so the link above is a legal-news report that quotes the decision; search the neutral citation to read the full text.
Take this list with youThe PDF is the handout: this list on one or two pages. The Markdown is for your own docs, your wiki, or an assistant you want to hand the list to. Print gives you this list only, without the navigation.
Something in your production AI broke in a way that is not on this list? Tell me what broke.