Is a company liable for what its AI chatbot tells customers?
Yes, in the case most often cited. In Moffatt v. Air Canada (February 2024), a Canadian tribunal held the airline responsible for its chatbot's wrong refund advice, rejected the idea that the bot was a separate legal entity, and ordered it to pay $812.02.
This page goes deeper on one risk our AI customer service agent page lists among the failure modes: a support agent that tells a customer something untrue. We read the full decision on the tribunal's own site on October 30, 2026, along with the regulator texts linked below, so you can see what was actually decided and what it means for how you build and word an agent.
We are an automation agency, not a law firm. The case is a small-claims decision from one Canadian province, and the outcome in your country may differ. What follows is what the texts say and how we design around them.
What happened in the Air Canada chatbot case?
A customer asked Air Canada's website chatbot about bereavement fares. It said the customer could apply for the discount after travel. Air Canada's policy did not allow that, refused the refund, and argued it was not liable for the chatbot. The tribunal disagreed.
Moffatt v. Air Canada, step by step
The decision, Moffatt v. Air Canada, 2024 BCCRT 149, was issued on February 14, 2024 by the British Columbia Civil Resolution Tribunal. The facts, as the tribunal set them out:
- In November 2022, after a death in the family, Mr. Moffatt used the chatbot on Air Canada's website while researching flights.
- The chatbot "suggested Mr. Moffatt could apply for bereavement fares retroactively".
- The chatbot's reply even linked to Air Canada's own "Bereavement travel" page, which said the policy did not apply after travel was completed. Staff later told the customer retroactive applications were not allowed.
- Air Canada argued it "cannot be held liable for the information provided by the chatbot", and relied on terms in its tariff without filing them.
The tribunal treated the claim as negligent misrepresentation and found each element proven. Its key passage:
"In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission. While a chatbot has an interactive component, it is still just a part of Air Canada's website. 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." Moffatt v. Air Canada, 2024 BCCRT 149, paragraph 27.
| Order (paragraph 44) | Amount (CAD) |
|---|---|
| Damages: fare paid ($1,630.36) minus what the bereavement fare would have been ($979.48) | $650.88 |
| Pre-judgment interest | $36.14 |
| Tribunal fees | $125.00 |
| Total | $812.02 |
The amount is small. The reasoning is what businesses quote, because it answers the question every chatbot owner asks: the bot's words are your words.
Why did the "correct information was elsewhere on the website" defence fail?
Because customers cannot be expected to check one part of a company's website against another. The tribunal found it reasonable to rely on the chatbot and said Air Canada did not explain why its policy page was more trustworthy than its own bot.
In paragraph 28 the tribunal found that Air Canada "did not take reasonable care to ensure its chatbot was accurate", and that it did not explain why customers "should have to double-check information found in one part of its website on another part of its website". In paragraph 29 it added that "there is no reason why Mr. Moffatt should know that one section of Air Canada's webpage is accurate, and another is not." The tariff defence also failed because Air Canada asserted what its contract said without providing it (paragraph 31).
For anyone running a support agent, the practical lesson is that a link to the correct policy does not cancel a wrong answer in the conversation. In this case the chatbot's own reply carried that link, and the airline still lost. The answer itself has to be right, or the agent has to hand off.
What do regulators say about chatbots that give wrong answers?
US regulators have warned that chatbots can break consumer protection law. The CFPB said in 2023 that banks using chatbots "risk violating legal obligations", and the FTC has acted against companies that made untested claims about what their AI could do.
| Regulator or body | Document and date | What it says that matters for support agents |
|---|---|---|
| BC Civil Resolution Tribunal (Canada) | Moffatt v. Air Canada, February 14, 2024 | A business is responsible for all information on its website, including its chatbot |
| Consumer Financial Protection Bureau (US) | Chatbots in consumer finance, June 6, 2023 | Warns that a chatbot "may fail to recognize that a consumer is invoking their federal rights", and that customers stuck in loops "don't have access to a human customer service representative" |
| Federal Trade Commission (US) | Final order against DoNotPay, February 11, 2025 | Said the company "did not test whether its 'AI lawyer' operated to the level of a human lawyer"; ordered $193,000 in monetary relief |
| European Union | AI Act, Article 50, applies from August 2, 2026 | Chat and voice agents must tell people they are AI; this is about disclosure, not accuracy, but it is part of the same build |
None of these says a chatbot must be perfect. Together they point the same way: the company that deploys the agent owns its answers, must not overstate what it can do, and must give customers a way to reach a person. Disclosure is covered on our page about whether you have to tell customers they are talking to AI.
Can a disclaimer protect you from what the chatbot says?
Do not count on it. Air Canada pointed to its own terms and to the correct policy page, and neither saved it. A disclaimer can set expectations; the safer protection is an agent that cannot state policies it was not given.
The tribunal did not rule on a chatbot disclaimer, because Air Canada did not rely on one. Disclosing that the agent is an AI, which our AI chat agents do in their first message, is a separate duty and does not make wrong answers acceptable. It did rule that terms a company asserts but does not produce are no defence, and that a correct page elsewhere does not excuse a wrong answer. A line such as "AI answers may contain mistakes" may help set expectations, but we treat it as a courtesy, not as protection. The protection is in the build.
How do you reduce the risk of a chatbot giving a binding wrong answer?
Limit the agent to approved content, make it hand off when unsure, keep money and policy exceptions with people, log every answer, and test against real tricky questions before launch. The Air Canada answer would have failed each of those checks.
Where a risky question should go
| Risk | How it happened in Moffatt | What we build instead |
|---|---|---|
| Agent states a policy it was not given | Chatbot described a retroactive refund the policy did not allow | Answers only from approved knowledge-base articles; a policy question with no matching article goes to a person |
| Two sources disagree | Chatbot and the "Bereavement travel" page said different things | One source of truth: the agent reads the same policy page customers see, with an owner and a review date |
| Money decided by the agent | Refund eligibility given as a chat answer | Refunds, credits and exceptions: AI gathers facts, a person approves |
| No record of what was said | The customer's screenshot of the chatbot reply was the evidence of what it said | Every conversation logged on the ticket, so you can check what the agent said |
| No test before customers see it | Not stated in the decision | A test set of real past questions, including edge cases like bereavement or hardship, run before launch and after every change |
Our page on how to test an AI agent shows how we build the test set, and the money page sets out which ticket types we automate and which stay with people. When an agent does give a wrong answer after launch, we trace it to the article or rule that caused it and fix it; fixes to anything we built are free for 6 months, at the time of writing, and optional support after that covers keeping answers current.
For the kinds of mistakes agents make and why they compound over several steps, see do AI agents make mistakes?
More questions about chatbot liability
Is the Air Canada chatbot case binding outside Canada?
No. It is a small-claims decision by a tribunal in British Columbia, and courts elsewhere are not bound by it. It is widely cited because the reasoning is simple and the facts are common: a company's chatbot gave a customer wrong information and the company tried to disown it. Expect lawyers and regulators in other countries to make the same argument the tribunal accepted.
Who is liable when an AI chatbot gives wrong information, the company or the AI vendor?
To the customer, the business that put the chatbot on its website or phone line, based on the Moffatt reasoning. Whether you can recover anything from your chatbot or model vendor depends on your contract with them, and most vendor terms limit their liability. Read the liability and indemnity clauses before you launch, and keep your own logs.
Can a chatbot's answer be legally binding?
It can have legal effect. In Moffatt the tribunal did not say a contract was formed; it found negligent misrepresentation and awarded damages equal to the discount the chatbot described. The practical result was the same: Air Canada paid the difference. Design your agent on the assumption that what it says about prices, refunds and policies will be held against you.
Does the AI need to be told it cannot make promises?
Yes, and that instruction alone is not enough. We write rules that stop the agent from quoting prices, refunds or exceptions it cannot find in approved content, and that send those questions to a person. Then we test those rules with questions designed to tempt the agent into a promise, because instructions that look clear on paper still fail on unusual wording.
Should we keep transcripts of chatbot conversations?
Yes. In Moffatt the customer filed a screenshot of the chatbot's reply as evidence of what it said. Your own log of every conversation lets you check what the agent said, fix the cause, and respond to a complaint with facts. Keep transcripts on the ticket in your helpdesk, in line with your privacy policy and any retention rules that apply to your industry.
Does this apply to voice agents too?
The reasoning does. The tribunal said it made no difference whether information came from a static page or a chatbot, and nothing in the decision suggests a voice agent would be treated differently. Voice adds one difficulty: customers rarely have a record of the call. Record or transcribe calls where the law allows it, and keep the same handoff rules as chat.
What should we do first if our chatbot already gave a wrong answer?
Find the conversation, find the cause (a missing article, an outdated policy, a rule the agent ignored), and fix the content the same day. Then decide with your own adviser how to treat the customer. Add the question to your test set so the same mistake is caught before any future change goes live.
Sources and further reading
- BC Civil Resolution Tribunal: Moffatt v. Air Canada, 2024 BCCRT 149, issued February 14, 2024, checked October 30, 2026
- CFPB: Chatbots in consumer finance, June 6, 2023, checked October 30, 2026
- FTC: FTC finalizes order with DoNotPay, February 11, 2025, checked October 30, 2026
- European Commission AI Act Service Desk: Article 50, transparency obligations, checked October 30, 2026
