Your AI Said It, You Own It: The Air Canada Rule for Website Chat and Inbox Replies
In November 2022 a man named Jake Moffatt needed to fly to Toronto for his grandmother’s funeral. He asked Air Canada’s website chatbot about bereavement fares. The chatbot told him he could book at full price and apply for the reduced bereavement rate within ninety days after travel. He did that. Air Canada refused, because its actual policy required applying before the flight, and pointed out that the correct policy was on a page the chatbot had linked to.
He took it to the British Columbia Civil Resolution Tribunal. Air Canada’s defense included the argument that the chatbot was a separate legal entity responsible for its own actions. In February 2024 the tribunal called that submission remarkable, found the airline liable for negligent misrepresentation, and ordered it to pay the difference plus costs. The amount was small. The reasoning was not.
This post is about what that reasoning means for a marketing team that runs a chat widget on the website and lets an AI draft replies to leads, why the disclaimer everyone reaches for does not fix it, what does, and how the same rule extends from a bot saying the wrong thing to an agent doing the wrong thing.
The two defenses that failed
Air Canada made the two arguments any company would make, and the tribunal’s answers to both are the rule.
“The bot is separate.” The airline suggested the chatbot was distinct from the company, a third thing that had spoken on its own. The tribunal’s response was that the chatbot is part of the website and the company is responsible for all the information on its website, whether it comes from a static page or a chatbot. There is no legal daylight between “the bot said” and “we said”.
“The right answer was elsewhere.” The airline argued the customer should have checked the bereavement policy page, which the chatbot had linked. The tribunal held that the company had not explained why a customer should trust one part of its website over another, and that a reasonable person would not know the page was more reliable than the chat. The correct information being available somewhere did not cure the wrong information being given.
Put those together and you have the Air Canada rule: what your AI tells a customer, you honor, and pointing at the fine print does not undo it.
It was a small-claims tribunal in one province. It is not binding on courts elsewhere. It is, however, the decision every commentator cites and the reasoning every subsequent analysis has expected other courts to follow, and a marketing team should assume it applies to them, because the alternative is betting the company on a judge finding the “separate entity” argument less remarkable than the last one did.
Why this is a marketing problem now
In 2024 this read as a customer service story. In 2026 it is a marketing story, because the chat on the website and the replies in the inbox are marketing’s.
A website chat widget is now the first conversation most B2B leads have with a company, and it is the place where a visitor asks the questions that turn into commitments: does this work with our system, what does it cost for our size, can you do it by March, do you serve our area. An AI answering those questions well is the reason the widget exists. An AI answering them plausibly and wrong is Air Canada.
The same is true one step later. When a lead replies to a campaign email or a text, an AI draft of the response is the fastest way to answer in minutes instead of a day, which is the number that decides whether the lead converts. That draft, if it goes out, is the company’s statement. “We can have you live in two weeks” from an assistant that inferred it from the website is a promise the company has now made.
None of this argues for turning the AI off. It argues for building it so it cannot say what it does not know.
A disclaimer is not the fix; scope is
The instinct after reading the case is to add a line: “This assistant may make mistakes; please verify important information.” That line is Air Canada’s second defense in a friendlier font. It tells the customer not to trust the answer while the company keeps giving answers, and the tribunal has already explained why that does not work: the customer has no way to know which answers are the reliable ones.
The fix is not to warn about wrong answers. It is to make wrong answers structurally impossible, which means scope.
- Answer only from a verified source. The assistant’s knowledge of your pricing, policies, service areas, timelines and product facts comes from a maintained facts source that a person owns, not from the model’s training data and not from an inference off the website. If the fact is not in the source, the assistant does not have it. The knowledge base post covers what goes in and who keeps it current.
- Abstain and hand off when the answer is not there. The most important sentence the assistant can say is a version of “I do not have that; someone from our team will follow up.” An assistant that can decline cannot improvise a bereavement policy. The hand-off should be honest about timing: “someone will be in touch shortly” is a promise you can keep; “an agent will join now” is one more statement you may have to honor.
- A person approves anything that commits the company. Pricing for a specific case, a delivery date, a policy exception, a refund. The assistant drafts; a human sends. The human-in-the-loop post covers where the gate goes without slowing the reply to a crawl.
- Log everything. Every conversation, with what the assistant said and what source it said it from. When a question comes later, the answer should be a lookup. Air Canada, notably, did not dispute what its chatbot had said; it disputed whether it mattered. Know what yours said.
A disclaimer costs nothing and protects nothing. Scope costs some setup and removes the failure.
Qualification questions are statements too. A chat that asks “what is your budget?” and then says “great, that is within our range” has made a representation about price. Design the qualification flow so the assistant collects and confirms and a person quotes.
From wrong words to wrong actions
Air Canada’s bot only talked. The assistants marketing teams are deploying in 2026 also act: they book meetings, update the CRM, send follow-ups, issue credits, change subscriptions. The same rule extends, and the stakes go up, because a wrong statement can be corrected and a wrong action sometimes cannot.
The year has produced a steady stream of widely reported incidents of autonomous agents doing exactly what they were told not to: deleting a user’s messages after being instructed to wait for confirmation, wiping data they were supposed to preserve, taking irreversible steps in production because the instruction was ambiguous and the agent resolved the ambiguity by acting. The pattern in each case is the same as the chatbot’s: the system did something plausible in the absence of a fact, and there was no gate between plausible and done.
For a marketing assistant the rule is the draft-only rule: the assistant may draft, propose and stage; a person approves anything that sends, commits or cannot be undone. Reversible, low-reach actions (log the conversation, tag the contact, propose three meeting slots) can run on their own. Irreversible or high-reach actions (send to a list, quote a price, change a record the sales team relies on) wait for the gate. The line is reversibility times reach, and it is a design decision, not a setting to hope for.
What this looks like built properly
Marqeable’s chat widget and inbox assistant are built on the scope model rather than the disclaimer model. The assistant answers from your business context, pricing and policies as you have entered them, and when a question falls outside that it collects the details, tells the visitor someone from the team will be in touch shortly, and hands off. Replies to texts and emails are drafted for a person to approve before they go, so nothing that commits the company leaves without a human name behind it. Every conversation is logged with what was said. The assistant can be fast because it cannot be wrong about the things it is not allowed to guess.
That is the Air Canada rule, applied before a tribunal has to.
Frequently asked questions
What happened in the Air Canada chatbot case?
A customer was told by the airline’s chatbot he could apply for a bereavement fare after travel; the real policy required applying before. British Columbia’s Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation in February 2024, rejected the argument that the chatbot was a separate entity, and rejected the argument that the correct policy was available elsewhere on the site.
Are AI chatbot statements legally binding?
The tribunal treated the chatbot’s statement like any other statement on the company’s website. It is one tribunal in one province, not a global precedent, but it is the reasoning courts are expected to follow. Assume what your assistant promises, you honor.
Does a disclaimer protect the company?
Not reliably. “Verify against the rest of the site” was Air Canada’s argument and it failed. Scope protects: an assistant that answers only from verified facts and hands off otherwise cannot make the misstatement.
How should a marketing team set up AI chat and replies safely?
A maintained facts source, abstain-and-hand-off outside it, a person approving anything that commits the company, full logging, and no irreversible actions without a gate.
The bottom line
A tribunal looked at a chatbot that gave a customer the wrong policy and said the obvious thing: the company said it, the company owns it, and the fine print elsewhere does not undo it. Your website chat and your AI-drafted replies are that chatbot, talking to leads about price, timing and promises all day. Do not add a disclaimer. Give the assistant a verified source, the ability to say “I do not know, someone will follow up,” a human gate on every commitment, and a log of everything it said. Then extend the same rule to what it does, not just what it says.
Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access
