In Moffatt v. Air Canada, the British Columbia Civil Resolution Tribunal ordered the airline to pay damages after its website chatbot described a bereavement fare policy that did not exist — telling a grieving passenger he could claim a retroactive discount within 90 days of travel, when the airline's actual policy, linked on the same page, ruled that out.
The airline's defence is the part worth studying. As the tribunal characterised it, Air Canada's position amounted to suggesting the chatbot was "a separate legal entity that is responsible for its own actions."
It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.
The award was CA$812.02. The precedent is what transfers.
Three questions it answers for agentic systems
Walk the reasoning forward from a system that speaks to a system that acts.
- Who owns the agent's mistakes? You do, entirely. The "the AI did it" argument has now been tested in a real proceeding and lost. There is no reading of this decision in which an autonomous component absorbs its operator's liability.
- What control was actually missing? Not a better model. A policy boundary. The bot contradicted the operator's own published policy — a source of truth that existed, was authoritative, and was sitting on the same page. Nothing evaluated the output against it before it reached a customer.
- What would have survived? An enforcement point between the agent and the commitment: policy checked where the promise is made, not discovered in litigation.
Why agents are the harder case
A chatbot makes a representation. A human then has to act on it, and that human is a second chance to catch the error — in this case, an airline employee who could have honoured or refused the claim.
An agent that moves money, files a claim or adjusts a record removes that second chance. The representation and the action collapse into a single event. By the time anyone reviews it, the effect has already landed in a system of record, and remediation is no longer a customer-service decision — it is a reportable event.
The useful exercise: take your three highest-volume agent workflows and ask, for each, what authoritative source the agent's output should have been checked against, and where in the path that check runs. If the answer is "the model was trained on our policies," you have the Air Canada architecture.
The tribunal's reasoning was not novel or aggressive. It applied an entirely ordinary principle: an operator is responsible for what its systems tell the public. That principle was always going to reach agents. It just reached chatbots first.