When the AI gets it wrong: who pays, and how not to be Air Canada

Yes, your business is liable for what your AI chatbot, voice agent, or automated email tells a customer. The same way it would be liable if an employee said it on the phone. This is not a guess about where the law is heading. A tribunal already tested the question against a real company. In 2024, it ruled plainly: the AI's mistake is still the company's mistake.

A note before any of this: we are not lawyers, and nothing here is legal advice. What follows reflects what the public record shows, the tribunal's published decision and the reporting around it. Your state, your industry, and your specific setup change the analysis. Treat every section below as a starting point for a conversation with your own attorney, not a substitute for one.

What we mean by "the AI gets it wrong"

A chatbot is software on your website or in a text thread that answers customer questions. It runs on pre-written rules, an AI model, or both. An AI agent goes further: it can look things up and take an action, not just answer a question.

"Getting it wrong" covers a few different failures. The AI states a policy that does not exist. It quotes a price or a deadline incorrectly. It says something so far outside its instructions that a customer reasonably treated it as a real offer. The common thread is reliance: a customer acted on what the AI told them, and that reliance cost them something.

The legal term for this comes from the published decision below: negligent misrepresentation. That means giving someone false information you should have gotten right, information they then relied on to their loss. It is an old legal category, not a new one invented for AI. Courts already knew how to handle a wrong answer from an employee. The ruling's point was that a chatbot does not change that analysis.

The case: Moffatt v. Air Canada

In November 2022, a customer named Jake Moffatt asked Air Canada's website chatbot about bereavement fares after his grandmother died. The chatbot told him he could book a full-price ticket. He could claim the bereavement discount later, it said, within 90 days of travel. Air Canada's real policy required the discount to be approved before booking, not claimed after the fact.

Moffatt booked on the chatbot's word. Air Canada later refused the refund and pointed to its actual policy. The dispute reached the British Columbia Civil Resolution Tribunal. In February 2024, the tribunal ruled against Air Canada.

Air Canada's defense was notable. It argued the chatbot was, in effect, its own responsible party, separate from the airline. The tribunal rejected that outright. A company is responsible for all the information on its website, it found. That includes a static page, and it includes an interactive chatbot. Air Canada paid the difference plus damages, a few hundred Canadian dollars in total. The payout was small. The reasoning was the story: "the AI said it" is not a defense.

What mistakes actually happen

Air Canada is the anchor case because a tribunal ruled on it. It is not the only one on record. A UK delivery firm's chatbot was reported in January 2024 to have sworn at a customer. A customer had pushed it to, and it also wrote a poem mocking its own employer. The company switched off the AI feature that same week. A US car dealership's chatbot, tested the same way, agreed in writing to sell a new SUV for one dollar. More than one city government chatbot has been reported giving residents wrong guidance on local rules. Residents had reason to rely on that guidance too.

None of these started as a company trying to mislead anyone. Each traces back to some mix of an outdated policy on file and a customer deliberately probing for a gap. Sometimes it is simpler: a model fills a silence in its instructions with something that sounds plausible and is wrong. None of that reads as an excuse in front of a tribunal. It reads as exactly the failure a business is expected to catch first.

Do disclaimers help

Some, by most legal commentary on this, but do not treat one as a shield. A visible note that the chatbot is AI is cheap to add. Say plainly that its answers do not override your written policies. It can narrow what a customer could reasonably claim to have relied on.

What a disclaimer will not do is excuse a wrong answer after the fact. It also will not replace an actual review step before something customer-facing goes out. Air Canada's chatbot reportedly carried no clear disclaimer of this kind, which likely made its defense harder, not easier. Whether a disclaimer changes the outcome in a similar case today is a real question. Put it to a lawyer who knows your state or country. It is not something to assume from one ruling.

How to limit the damage before it happens

The fix that actually holds up is not a longer disclaimer. It is keeping anything customer-facing or financial behind a human check before it goes out. That way a wrong answer gets caught before a customer acts on it, not after. What needs a person and what does not is its own judgment call. We cover it in what automation cannot do yet.

Some things should always wait for your tap of approval. Others are safe to run on their own. We map out exactly which is which in should an AI ever talk to your customers without a human approving.

Where we come in

We build digital workers, defined plainly here, with this exact failure mode in mind. Anything customer-facing or financial waits behind an approval gate, one tap from you. Every action lands in an append-only record, so you can see what was said and when. None of that makes a business immune to a Moffatt-style claim. It means a wrong answer gets caught by you before a customer acts on it.

None of this is legal advice. Your actual exposure depends on your state, your industry, and your setup, and that conversation belongs with your attorney. What we can tell you plainly is where automation adds this kind of risk and where it does not. The automation profile tool is a two-minute way to see that for your own business. It sorts your tasks into what is safe to hand off and what still needs a person checking the output.