The Air Canada chatbot case: who's liable when your AI gets it wrong

In 2024 a Canadian tribunal answered a question that had been purely theoretical until then: when a company's chatbot gives a customer the wrong information, who pays? The answer was unambiguous — and it applies to every business that has put a chatbot or AI agent in front of its customers.

If you use AI to talk to customers, this case is the most useful thing you'll read this year. Not because it's complicated, but because it dismantles a costly misconception.

What happened, briefly

A passenger asked Air Canada's chatbot about bereavement fares. The chatbot assured him he could book at full price and claim the discount retroactively, within 90 days. That policy did not exist. The chatbot invented it — what we now call a large language model "hallucination."

When the passenger requested the refund, Air Canada refused: its actual policy required the claim to be made before travel. The dispute went to British Columbia's Civil Resolution Tribunal.

The argument that failed

Air Canada's defence was remarkable in its nerve: it argued, in essence, that the chatbot was a "separate legal entity responsible for its own actions." In other words — "don't blame us, blame the bot."

The tribunal rejected the argument as absurd: the chatbot is part of the company's website, and the company is responsible for all the information it provides — whether it comes from a static page or a chatbot.

Air Canada was held liable and ordered to compensate the passenger. The amount was small; the precedent it set was not.

Why it matters, even if you're not an airline

The principle travels everywhere: a business is responsible for what its AI tells a customer, even when the AI "made it up." It doesn't matter whether you use an off-the-shelf chatbot platform, whether a contractor built it, or whether you "didn't know" what it would say.

In practice this means a chatbot that promises a discount that doesn't apply, quotes the wrong price, or "confirms" a returns policy that doesn't exist can bind you legally. The customer reasonably relied on what you told them — through your tool.

How it connects to the EU AI Act

The Air Canada case is about civil liability, not the AI Act. But it illustrates exactly the risk the EU AI Act sets out to contain. Article 50 of the regulation requires that users know they're talking to AI — and that obligation takes effect on 2 August 2026. Transparency isn't bureaucracy; it's your first line of defence. When a customer knows they're talking to a machine, the way they weigh the answer changes — and so does your exposure.

What to do this week

• Disclose that it's AI. Every chatbot or voice agent should open by making clear it isn't a human.
• Constrain what it can "promise." Prices, discounts, legal or binding statements shouldn't be left to free text generation — route them through a verified, signed-off FAQ.
• Assign human oversight. Someone accountable for what the system says, with a clear path to escalate to a person.
• See where you stand. Run the free AI Act test to find which risk category your system falls into, and use the AI policy generator for a first internal framework.

The lesson of Air Canada isn't "don't use a chatbot." It's: deploy it with limits, with transparency, and with someone accountable behind it. That's what separates you from the business that learns the same lesson the hard way.