AI Customer Support Without the Air Canada Mistake

Here is the question I want you to answer before you read another word about AI support: if your chatbot tells a customer something wrong, whose words are they?

Not the software vendor's. Not the model's. Yours.

That is the plain lesson from a small tribunal case in British Columbia. And the other half of the story, from a fintech company, is that the upside is real. An AI assistant can carry a huge share of your support volume. Both facts are true at once, and the businesses that win with AI support will be the ones who hold both.

What happened with Air Canada's chatbot?

The facts are simple. A customer, Mr. Moffatt, was booking a flight to attend his grandmother's funeral and asked the airline's chatbot about bereavement fares. The chatbot told him he could book a regular ticket and request a partial refund within 90 days under the bereavement policy. Air Canada employees later told him retroactive applications were not permitted.¹

He took it to the Civil Resolution Tribunal of British Columbia, in the case Moffatt v. Air Canada. According to the Manatt analysis, the tribunal found that Air Canada owed him a duty of care, that it did not take reasonable care to make sure its chatbot was accurate, and that this amounted to negligent misrepresentation.¹ The tribunal also rejected Air Canada's argument that the chatbot was a separate legal entity responsible for its own actions.¹ The award was a partial refund of CAD $650.88, plus damages covering interest and tribunal fees.¹

Why this small case matters

The dollar amount is tiny. The principle is not. A customer asked a question on your website, got a confident answer, and acted on it. The tribunal's reasoning treated that answer as something the company was responsible for, and questioned why customers should have to cross-check one part of the airline's website against another.¹

I am not a lawyer, and this is one tribunal's ruling in one jurisdiction. It does not set a rule for every country. But if you are building customer support on AI, you should assume a judge, a regulator or an angry customer will treat what your bot says as something your company said.

The bot was wrong in a very human way

Notice what the chatbot got wrong. It did not fabricate a fact about the weather. It stated a policy, confidently, that did not match the official one. That is the highest risk category for an AI support agent: money, refunds, deadlines, guarantees and eligibility. The places where a wrong answer costs real dollars.

What did Klarna show about how much AI can carry?

Now the other half. On February 27, 2024, Klarna announced results for the first month of its AI assistant. It reported 2.3 million conversations, about two thirds of its customer service chats, equivalent to the work of 700 full-time agents.²

Klarna also reported that resolution time fell from 11 minutes to under 2 minutes, repeat inquiries dropped 25%, and the assistant worked in more than 35 languages across 23 markets, around the clock. It said customer satisfaction was on par with human agents.²

How to read a company's own numbers

These are Klarna's figures from its own press release, not an independent audit. A company announcing a launch is telling the best version of the story. I still take the order of magnitude seriously. Fast, repetitive, well-defined questions, like where is my order or what does this status mean, are exactly where AI helps most.

The smart reading is not "replace your support team." It is "find the repetitive share of the work, say 60% as an example, so your people can spend their time on the part that needs judgment."

Why do both cases point at the same lesson?

Because the dividing line is not AI versus human. It is information versus commitment.

An AI assistant can be excellent at the left column, as long as it draws its answers from approved, current material. The right column creates obligations. Treat it like signing a contract: the machine can gather the facts, but a person with authority makes the promise.

What does "no AI exemption" mean for small operators?

In September 2024 the FTC announced Operation AI Comply, a set of enforcement actions against companies using AI claims or tools in deceptive ways, naming cases involving DoNotPay, Ascend Ecom, Ecommerce Empire Builders, Rytr and FBA Machine.³

That sentence is the one to remember. Truth in advertising, consumer protection and refund promises apply whether a person typed the message or a model generated it. If you would be responsible for it in an email, you are responsible for it in a chat. The same logic sits behind the FTC's fake review rule.

The point is not to scare you off AI. It is to treat your assistant like a new hire who speaks for the company: trained, limited in what it can promise and supervised at first.

How do you let AI answer customers safely?

Here is the build order I would use for a lean business.

Keep the bot on script

In January 2024 the delivery company DPD disabled part of its AI chatbot after a customer got it to swear and criticize the company. DPD said an error had occurred after a system update the previous day.⁴ You do not need adversarial customers to have a bad day. A system update, a changed prompt or a new document can shift behavior overnight. That is why weekly review and a fast off switch are not optional.

How do you decide which questions to automate first?

Sort every common question by two things: how often it comes up, and how expensive a wrong answer would be. That gives you four boxes.

High volume and low risk is where you start: order status, hours, how-to steps, where to find a link. High volume and high risk, like refund requests, is where the assistant gathers details and hands off. Low volume and low risk can wait. Low volume and high risk should stay with a person, full stop.

Count the cost of being wrong, not just the cost of the ticket

It is tempting to measure support automation by tickets deflected. That number flatters you. A better ledger includes the cost of the wrong answers that slip through: refunds you honor because the bot promised them, customers who leave after a confident mistake, and the hours your team spends cleaning up. If the bot saves ten hours a week and creates one costly dispute a month, you have to do the math honestly.

Klarna's own release is careful to describe the assistant's results against both speed and satisfaction, not speed alone.² Copy that habit. Track resolution time, repeat contacts and customer satisfaction together, and track escalations to a human as a healthy signal, not a failure.

What should AI never answer on its own?

My short list for any small business:

• Anything about refunds, cancellations, chargebacks or guarantees.
• Pricing exceptions and custom quotes.
• Legal, medical, tax or financial questions that touch a person's specific situation.
• Complaints about a person on your team.
• Anything where the customer sounds upset, confused or vulnerable.

For those, the right job for the assistant is to collect the details, tell the customer a person will respond, and put the case in front of someone with the authority to decide.

The Manatt checklist, in my words

The Manatt analysis offered practical advice for companies after the Air Canada ruling: regularly review and update the information your automated systems rely on so it matches your posted policies, work with your vendors to define what the system can and cannot say about policies and contractual terms, give staff a procedure for resolving conflicts between automated answers and official policy, and assume you will likely be bound by what the bot says, even when a third-party vendor supplied it.¹ That last point deserves underlining. Outsourcing the tool does not outsource the responsibility.

How should you start this week?

Pick the five questions your customers ask most that have a factual, stable answer. Write the answers once, in plain language. Put them in front of an assistant, with a human fallback, and read every conversation for the first two weeks.

I have worked as a one-person operation since 2024, and I use AI heavily, and the lesson I keep relearning is the same: speed without supervision is not efficiency, it is a loan against your reputation. The wins come from small, checked systems that run every day.

If you are weighing where AI belongs in your customer path, see the AI-assisted systems I build for clients.

Keep building.Ekai