A grieving customer asked an airline chatbot a simple question. The bot invented a policy that did not exist. A tribunal made the airline honor it anyway, and set the precedent that companies own every word their AI says.
In late 2022, Jake Moffatt's grandmother died. He booked an Air Canada flight to attend her funeral and asked the airline's website chatbot whether bereavement fares could be claimed after the fact. The chatbot said yes: apply within 90 days of the ticket purchase and you'll get a partial refund. The policy did not exist. Air Canada's real rule is the opposite: bereavement fares must be requested before travel. When Moffatt filed the refund claim and was denied, he took the airline to a tribunal. In February 2024 he won. The chatbot's invented policy became, in effect, the policy that Air Canada had to honor.
The British Columbia Civil Resolution Tribunal heard the case as Moffatt v. Air Canada. Air Canada's defense was striking: it argued the chatbot was "a separate legal entity that is responsible for its own actions." The tribunal disagreed. Tribunal member Christopher Rivers ruled the airline was liable for negligent misrepresentation, and awarded Moffatt CAN$812.02 in damages and fees. The chatbot, Air Canada confirmed afterward, was quietly retired.
The amount is small. The precedent is not. The ruling established that companies are responsible for what their AI says, even when the AI says things that aren't true. And the chatbot's invented policy wasn't a wild fabrication. It was a plausible-sounding refund rule that lots of airlines actually have. It looked exactly like the kind of answer a competent customer-service agent would give.
That is the core of the AI hallucination problem: the wrong answers and the right ones look identical.
Air Canada's chatbot was built on a large language model trained to generate fluent, plausible text. The model has no concept of "what Air Canada's actual refund policy is." It has a concept of "what a refund policy for an airline sounds like," and it will compose one on demand, with the same confident tone whether the underlying facts are correct or completely invented.
Most AI products that touch business data have this exact architecture under the hood. The model gets a question, generates an answer that sounds right, and ships it to the user. There is no step in the middle that asks: Is each specific claim in this answer actually supported by a source in our system? Without that step, the model is free to interpolate, filling in details that feel reasonable but aren't grounded in anything verifiable.
The same failure mode is rife in AI analytics tools. An LLM asked "What was our revenue last quarter?" will happily produce a confident, specific number. Whether that number came from your warehouse or from the model's statistical guess about what "a SaaS company's quarterly revenue" usually looks like is, without proper architecture, anyone's guess.
The fix isn't to make the model "smarter" or "less likely to hallucinate." It's to make hallucination easy to catch by requiring every claim in the output to be traceable back to a specific, real data point, and flagging the output if it isn't.
Askalytics' AI Insights pipeline is built around this constraint. The AI queries the customer's own data warehouse directly and runs the queries it needs, but every number it returns traces back to the exact query that produced it. Each figure carries a receipt you can open, the Advanced Breakdown, showing the query, the rows, and the source that produced it. On top of that, every AI Insight is re-checked by a second AI on a different provider, a Confidence Check, so a confident claim the data does not actually support gets flagged before the customer ever sees the insight.
Askalytics grounds every number in your real warehouse data: no hallucinated policies, no unchecked revenue figures.
The same idea applies, in a different shape, to chatbots that take customer-facing actions. If Air Canada had required every policy statement to cite a specific section of an internal policy document, and rejected statements that couldn't cite one, the tribunal case wouldn't exist. The chatbot would have answered "I can't find a policy that confirms this, so let me transfer you to an agent," which is the right answer when the model genuinely doesn't know.
This is the practical difference between AI that's "trained on your data" and AI that's grounded in your data. Training shapes tendencies. Grounding enforces facts.
Air Canada's case is the legal system catching up to a technical reality that engineers have known for years: a language model on its own is a confident, fluent, and wrong narrator, and putting one in front of customers or in front of business data without a grounding layer is a liability waiting to crystallize. The architectural fix is older than the problem (cite your sources, validate the citations, reject the output if it can't), but it's only just becoming table stakes for AI tools used in serious decisions.
If you want to dig deeper into how Askalytics structurally prevents the Air Canada failure mode in analytics, see our take on how AI Insights are grounded and the Confidence Check that backstops every insight.