Who is responsible when your chatbot gets it wrong.

What happened

On 11 November 2022, the day their grandmother died, Jake Moffatt went to Air Canada's website to book a flight from Vancouver to Toronto at the airline's bereavement rate, and asked its chatbot about bereavement fares. The chatbot said that anyone who needed to travel straight away, or had already travelled, could submit their ticket for the reduced rate within 90 days of it being issued, using a refund form.

The words "bereavement fares" in that reply linked to Air Canada's bereavement travel page, which said the policy did not apply to requests made after travel. Relying on the chatbot, Moffatt booked at the full fare, flew, and applied for the partial refund well inside the 90 days. Air Canada refused. In an email, one of its representatives admitted the chatbot had used "misleading words" and said the airline would update it.

Moffatt took the claim to British Columbia's Civil Resolution Tribunal, which hears small claims. The decision, Moffatt v. Air Canada, 2024 BCCRT 149, came on 14 February 2024.

What the tribunal decided

The tribunal treated the claim as negligent misrepresentation. That needs a duty of care, a statement that was untrue or misleading, carelessness in making it, reasonable reliance on it and a loss as a result. It found all five.

Air Canada's main defence was that it could not be held liable for information from its agents or representatives, chatbot included. The tribunal called this "a remarkable submission". Its reasoning, in short:

• A chatbot is still part of the company's website, and the company is responsible for all the information on it, whether it comes from a static page or a chatbot.
• Air Canada did not take reasonable care to make its chatbot accurate.
• It did not explain why its bereavement page should be trusted more than its chatbot, or why customers should check one part of its website against another.
• Relying on the chatbot was reasonable. A customer has no way to know that one part of a website is accurate and another is not.

A defence based on the airline's tariff, its contract terms, failed because Air Canada never produced the relevant part of it. The tribunal ordered it to pay the difference between what Moffatt paid and the bereavement fare, with interest and the tribunal fees.

A company is responsible for all the information on its website, whether it comes from a static page or a chatbot. That is the line every business with a chatbot should remember.

What it did not decide

Read the case for what it is. It is a small claims decision from one Canadian province. It does not bind courts in the US, the UK or the EU, and it says nothing new about AI law. The tribunal applied ordinary rules on careless statements to a company's website, and the chatbot happened to be the part of the website that was wrong.

It also did not decide whether a warning in the chat window would have changed the result, because Air Canada did not argue that. It argued that the chatbot was responsible for itself, and that the right answer was on another page. Both arguments failed.

The point that travels is the simple one. Customers treat what your chatbot says as what your business says, and so did the tribunal.

What it means for a dentist or a plumber

Air Canada's chatbot got a policy question wrong, and policy questions are what small business chatbots answer all day. Picture the same mistake in your own trade:

• A dental practice's chatbot says a new patient check-up includes X-rays, when the fee list says they are extra.
• A plumber's chatbot says there is no call-out charge at weekends, when there is.
• A salon's chatbot says cancellations are free up to an hour before, when the policy says 24 hours.
• A hotel's chatbot says late checkout is free, when it is free only for members.
• A restaurant's chatbot says a dish is nut free, when the kitchen has never confirmed it.

Each one is a customer acting on what your business told them. The first four cost money and goodwill. The last can hurt someone. None needs an unusual failure, only a chatbot answering from something other than your current, approved information, or improvising when that information is silent.

Where wrong answers come from

Wrong answers rarely come from nowhere. When you trace one back, it usually starts in one of these places:

• An old page. Last year's price list, a holiday notice nobody removed or a PDF from before the policy changed, still among the sources the chatbot reads.
• Two sources that disagree. Your website says 24 hours' notice for cancellations, your booking confirmation email says 48, and the chatbot picks one.
• General knowledge. Asked something your information does not cover, a loosely set up chatbot fills the gap with what businesses like yours usually do. That is how a chatbot invents a policy.
• The exception left out. Air Canada's policy had a condition, no requests after travel, and the answer dropped it. Policies with conditions need answers that keep them.
• A question about one person dressed as a general one. "Can I still get the discount?" depends on who is asking, and only a person can check.

Each cause has a fix you can make before launch: retire old sources, settle conflicts, forbid answers from general knowledge, write the conditions into the approved answer and hand over anything that depends on the person asking.

Why a disclaimer is not the fix

The reflex after a story like this is to add a line: "Answers may be inaccurate. Please check our policies." That does not change what the chatbot says. In March 2024 The Markup reported that New York City's chatbot for business owners told them they could take a cut of workers' tips, that landlords did not have to accept housing vouchers and that shops did not have to take cash. All three answers were wrong, and the page already carried a note that the chatbot could sometimes produce incorrect or harmful content.

A disclaimer sets expectations, and some trades use one as standard. It does not make a wrong answer right, and the Air Canada tribunal saw no reason a customer should have to check one part of a website against another. Our guide on what a chatbot should hand over answers the disclaimer question for advice-giving trades. The fix comes earlier: control what the chatbot may say.

Answers from approved information only

This is why our AI chatbots answer only from information the business has approved, and say so when the answer is not there instead of guessing. In practice:

• One source for each fact. If your website says one thing and your price list another, fix the conflict before launch, because the chatbot repeats whichever it reads.
• A short answers document for questions your pages do not cover, each answer approved by the person who owns it.
• No answering from general knowledge. "Do you charge a call-out fee at weekends?" gets your policy, or "I don't have that information, so I'll pass your question to the team."
• A person behind it. Anything that needs a decision, an exception or a judgement goes to someone, with the conversation attached.
• Every conversation readable afterwards, so a wrong answer can be traced to the page or document it came from and fixed there.

Air Canada's chatbot linked to the right page and still gave the wrong answer. Linking to a policy is not the same as answering from it.

Things it should never decide

Some answers commit you, so keep them out of the chatbot's hands entirely:

• Prices beyond your published list, and any quote for work nobody has seen.
• Discounts, refunds, waivers and exceptions to a policy. A visitor who says a member of staff promised them a discount gets a handover, not a yes.
• Whether a repair or a product is covered by a warranty.
• Confirmed bookings, unless it can see and write to your real calendar. Otherwise it takes a request and says so: "This is a request, not a booking. Reception will confirm a time with you."
• Eligibility for anything, from a special fare to a free consultation.
• Advice about someone's own health, legal case, money or safety.

A customer service chatbot built this way still answers the repeated questions about opening hours, services, prices from your list and how things work. It stops where a member of staff would have to check with you first. A booking chatbot follows the same rule for appointments: it takes the request and says who confirms it.

Test it the way Moffatt used it

Moffatt asked a reasonable question at a hard moment and acted on the answer. Test your chatbot the same way before launch, and again after every policy change:

• For each policy, such as cancellations, refunds, deposits, call-out fees and warranties, write the question three ways, including one from someone upset or in a hurry.
• Add the edge cases: after the deadline, already used the service, a family emergency, a promise from staff.
• Check every answer against the policy page, and every number exactly.
• Click every link the chatbot gives, and check that the page says the same thing as the answer.
• Ask something your information does not cover, and check that it says so instead of guessing.

Keep the questions. When you change a policy, run them all again, because a change that fixes one answer can break another.

When it does get something wrong

It will happen, as it does with staff. What matters is what you do next:

• Find the conversation, and the source the answer came from.
• Contact the customer if they left details, before they contact you.
• Decide what to honour. For a small amount, honouring the answer often costs less than the argument. For anything large, take advice before you reply.
• Fix the source as well as the answer, and run your test questions again.
• Note what happened and what you changed, with the date.

Air Canada spent months emailing a bereaved customer and then defended a small claim. The goodwill it lost was worth more than the fare.

Keep the records that protect you

Air Canada lost its contract defence because it did not produce the contract. Records cut both ways, so keep the ones that show what you said and when:

• Dated versions of your policies and price list.
• The answers document behind the chatbot, with a log of changes.
• Chat transcripts, kept for a set period that you have agreed and told visitors about.
• Your test questions and their results from before each change.

Balance this with privacy. Keep transcripts as long as you need them for follow-up and disputes, and no longer than your privacy notice says.

The same rule covers AI your staff use

The tribunal's point was about a website, but the reasoning fits any channel. If a receptionist pastes a question into an AI tool and sends back an answer with the wrong refund terms, the customer got that answer from your business. Anything AI helps write needs a person who checks the facts against your own documents before it goes out. An AI policy for staff puts that rule on one page.

Check your contracts and your insurance

Two questions worth asking:

• Your chatbot vendor: who pays if the tool states something your information does not say, and which logs will they hand over if a customer disputes an answer?
• Your insurer or broker: does your policy cover a claim that starts with something your website or chatbot told a customer?

If you are choosing a chatbot now, how to add an AI chatbot to your website covers testing and launch in order, and our prices page shows what a chatbot from us costs to set up.