When Your Chatbot Makes a Promise, Your Business Keeps It

Who is responsible when a website chatbot gets it wrong?

Many owners who add a chatbot to their website think of it as a helpful extra, a friendlier version of the FAQ page. If it gives a wrong answer now and then, that feels like a software problem, not a business problem. Customers do not see it that way. To them, the chatbot is your business talking.

In February 2024, a tribunal in British Columbia did not see it that way either. The case, Moffatt v. Air Canada, involved a few hundred dollars and a grieving customer, and it has been widely reported since. The amount was small. The lesson for an owner-led business is not.

What happened in Moffatt v. Air Canada?

As CBC News reported, Jake Moffatt's grandmother died on Remembrance Day in November 2022. The same day, he went to Air Canada's website to arrange travel to and from Toronto and asked the site's chatbot about bereavement fares.

According to a screenshot he provided, the chatbot told him that if he needed to travel immediately or had already travelled, he could submit his ticket for a reduced bereavement rate within 90 days of the date it was issued by completing a refund application form. The chatbot's answer also linked to Air Canada's bereavement travel page, which said the policy did not cover retroactive requests. He did not open that page. He bought full-price tickets, and when he later sought the reduced rate, Air Canada would not give it.

Before the Civil Resolution Tribunal, Air Canada argued that it could not be held liable for information provided by its chatbot, and that the correct information was available elsewhere on its website. Tribunal member Christopher Rivers summarized the position this way: "In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions." He called it "a remarkable submission" and wrote, "It should be obvious to Air Canada that it is responsible for all the information on its website."

He found that Air Canada "did not take reasonable care to ensure its chatbot was accurate," and added, "There is no reason why Mr. Moffatt should know that one section of Air Canada's webpage is accurate, and another is not." The tribunal decided the claim as negligent misrepresentation and ordered Air Canada to pay $812.02 in total, which included $650.88 in damages.

One caution before the lesson. This was one decision by one tribunal in one Canadian province, and it is not a statement of the law where your business operates. I am not a lawyer, and nothing here is legal advice. The operating lesson, though, does not depend on any jurisdiction.

What did the decision miss before it ever reached a tribunal?

I do not know what conversations happened inside the airline, and I will not guess at anyone's intentions. What the public record shows is a decision about a tool, made long before this customer arrived, that left three questions unanswered.

The first was what the chatbot was allowed to talk about. A question about a bereavement fare is a question about money and about a promise. Those are the topics where a wrong answer costs something real, and they are exactly the topics a customer-facing tool should be given an approved answer for, or kept away from entirely.

The second was whether the chatbot and the written policy agreed. The correct rule existed on a page the chatbot itself linked to. The answer the customer read did not match it. When the same business says two different things in two places, the customer is not the one who should have to reconcile them, and the tribunal said as much.

The third came after the mistake: who owns what the tool says. Arguing that the chatbot was responsible for its own words is the opposite of ownership. Most owners would never say that about an employee at the front desk, and a customer hears it the same way.

What is the better way to think about a customer-facing chatbot?

Treat it as a member of your team who speaks for you, not as a piece of software you installed. You would not let a new hire quote refund terms on day one without training, a script for the sensitive questions, and someone checking their work. A chatbot deserves the same discipline, and arguably more, because it talks to every visitor at once and never tells you when it is unsure.

This is where the Four A's help. Greeting a visitor and answering where you are located is work AI can Automate. Explaining your service can be Augmented with approved content. But anything that commits money or a promise, such as refunds, cancellations, discounts, guarantees, and exceptions to policy, belongs in Advise or Avoid: the tool can point to the written policy or hand the customer to a person, and a person makes the commitment.

A disclaimer does not change that picture. It may set expectations, but it does not make a wrong answer right, and it does nothing for the customer who acted on the answer. There is an old instruction to let your yes be yes. A chatbot that says yes on your behalf is still your yes.

Telling the chatbot to be careful is not enough either. An instruction makes a mistake less likely; it does not make it impossible. I learned that with my own AI phone agent and wrote about it in A Prompt Is Not a Control. When something must never happen, remove the ability or put a person in the path.

How do you check your own chatbot this week?

You can run this check in under an hour, and you do not need to buy anything to do it. Open your website as a customer would and work through these steps.

1. 01Ask your chatbot the questions that commit you: refunds, cancellations, prices and discounts, warranties or guarantees, and how fast you will respond. Copy each answer word for word.
2. 02Compare each answer with your written policy. Every mismatch is a promise your business may end up keeping whether you meant to make it or not.
3. 03For each of those topics, choose one of two paths: give the chatbot a single approved answer to use word for word, or have it hand the customer to a named person.
4. 04Check what the tool can do beyond talking. If it can issue refunds, book or cancel appointments, or change orders, decide whether it should, and remove what it should not.
5. 05Keep each policy in one place, and make updating the chatbot part of the same task as updating the policy, on the same day.
6. 06Name one person who owns the chatbot, reads a sample of real conversations each month, and corrects what they find.

If you find more than one surprise, write the rules into your AI policy so the next tool starts from them. A Practical AI Policy for a 20-Person Company has a one-page outline that includes human review and named owners.

If you would rather have someone review your customer-facing AI with you, an AI audit catalogs what each tool says and does, and AI governance consulting sets the owners, approved answers, and review steps that keep it accurate. For phone agents, chat, and follow up in a service business, AI consulting for service businesses covers how that work is set up.