Your AI Bot Just Made a Promise Your Company Can't Keep. Here's Who's on the Hook.

A man's father died. He needed to fly home for the funeral, so he asked his airline's chatbot about a bereavement fare. The bot told him yes, absolutely, book the trip, then send in the documentation within 90 days for a reimbursement. He did exactly that. The airline later told him the bot was wrong and refused to pay. He took them to court. He won. The airline had to pay.

This is a story from 2024 that I recently chatted about again with Julie Cole, co-founder and Senior Director of Mabel's Labels, on a recent episode of my podcast (AI Literacy for Entrepreneurs). Her point: "Everybody knows that you are responsible for what your bots tell people." That's now settled fact, and most companies deploying AI-driven customer service still haven't absorbed it.

The Ruling Very Few Leaders Have Read

The case is Moffatt v. Air Canada, decided by Canada's Civil Resolution Tribunal in 2024. Air Canada's defense, in essence, was that the chatbot was "a separate legal entity responsible for its own actions". The tribunal rejected that outright. A company is responsible for every representation made through its channels, whether that representation came from a human employee or a chatbot running on an LLM.

That ruling should have reshaped how every company thinks about deploying AI in customer-facing roles. For a lot of organizations, it hasn't, because the story got filed away as "an airline thing" instead of what it actually is: a preview of the liability every company takes on the moment a bot starts making commitments to customers on the company's behalf.

If your AI tool tells a customer something false about a refund, a policy, a deadline, or an entitlement, your company owns that statement exactly as if your best employee had said it in an email. The bot doesn't absorb the liability. You do.

When Efficiency Becomes the Product's Enemy

I ran into a smaller, less life and death version of this pattern myself, on the very platform I use to produce this podcast.  For two years, I never once needed to contact support. The interface wasn't fancy, but it worked: a straightforward process, a few steps, and I got what I needed. Then the platform replaced the whole thing with an AI interface. Now, to produce an episode, I have to chat with a bot instead of just doing it myself.

In theory, I understand the logic: tell the bot what you want, and it handles the rest. In practice, I ran into seven separate points where the bot didn't understand what I needed, and a task that used to take half an hour turned into roughly 24 hours of back-and-forth with support, the first time I'd needed support in two years. When I finally asked if I could just go back to doing it the old way myself, the answer was: sure, that's a feature available to our enterprise here's the enterprise pricing page.

That's an epic failure in AI use. The redesign wasn't built around what made the product actually work for me. It was built around a vision of simplicity that, for an existing user with an existing workflow, made everything slower, then tried to upsell me out of the friction it created.

Two Different Failures, Same Root Cause

The Air Canada case and my own platform frustration look like different problems. One is a legal liability story. The other is a product-experience gripe. But they come from the same place: a company let AI make commitments or handle interactions without asking whether the AI understood the actual customer's actual situation.

Air Canada's bot didn't know that "your father just died and you need to get home" is a moment that calls for a human, not a policy summary generated on the fly. My platform's new bot didn't know that a two-year user with a working process doesn't need to be onboarded into a conversational flow designed for someone starting from zero.

Julie's read on this, from her own experience running a company built on being "customer obsessed" is: "AI could help to a certain point, but when you really need somebody... that's when you need that human touch, and that's where it isn't helping for that customer experience."

Where to Draw the Line

The fix is in being deliberate about where the line sits for customer-facing AI roles, and building that line before a bad outcome forces the question.

Anything that constitutes a commitment, a promise, an exception, a policy interpretation, needs a human checkpoint, or needs the bot constrained to only repeat verified, pre-approved answers. Air Canada's bot generated a plausible-sounding answer about a bereavement policy that didn't actually match the airline's real policy. That's exactly the failure mode generative AI produces when it isn't constrained: a confident, well-formed answer that happens to be wrong.

Anything involving grief, urgency, or a customer already in a difficult moment needs to route to a person, fast. This is the test Julie's own business runs on: would you want a bot handling this if it were your situation. A bereavement fare question is not a shipping-status question.

Test the redesign against your existing customers before rolling it out, not just new ones. My platform's AI interface may genuinely be easier for a brand-new user with no existing workflow. It was measurably worse for me. If a company only tests a new AI-driven flow against a blank-slate user, it will miss exactly the failure mode that cost me a day and cost Air Canada a lawsuit: existing, functioning relationships getting worse in the name of a smoother average experience.

What This Means for Your AI Rollout

If your organization is deploying AI anywhere it talks to customers, someone needs to be able to answer a specific question before launch: what happens when this bot is confidently wrong, and who is checking? If the honest answer is "we haven't thought about that yet" or if it lands on the desk of some junior with an average script, that's a governance gap worth closing before a customer, or a court, closes it for you. (Further reading: What Is the Difference Between AI Governance and AI Compliance?) 

A simple policy answering what data goes in, how output gets reviewed, and who has authority to make exceptions covers most of this before it becomes a legal problem. (Read: How to Write a One-Page AI Policy)

The Short Version

Air Canada lost in court because its chatbot made a promise the company didn't intend to keep, and the tribunal ruled a company is responsible for what its bots tell customers, no exceptions. That single ruling should inform every AI customer-service deployment happening right now. Pair it with a simpler truth: AI redesigns built around theoretical simplicity often make things worse for the customers who already had a working process, not better. Test the line between where AI helps and where a human needs to step in before a customer, or a courtroom, tests it for you.

You can hear the full conversation with Julie Cole, co-founder of Mabel's Labels, including her take on founders proving real expertise in an AI-saturated market, on episode 277 of AI Literacy for Entrepreneurs.

If you want a structured way to check where your own AI deployments might be creating this kind of exposure, the NorthLight AI Readiness Audit gives you a structured baseline across all five areas in about 15 minutes. Run the audit now.