Air Canada Let a Chatbot Make a Promise It Refused to Keep. A Tribunal Said That Was Not Okay.
Air Canada's chatbot gave a grieving passenger incorrect information about bereavement fares. When he asked for a refund the airline argued the chatbot was a separate legal entity not responsible to them. A tribunal disagreed. Here is what the engineering team should have asked before launch.
Jake Moffatt's grandmother died on November 11 2022.
The same day he visited Air Canada's website to book a last minute flight from Vancouver to Toronto. He asked the chatbot about bereavement fares. The chatbot told him he could apply for a discounted rate within 90 days of travel. He booked the full fare tickets trusting that information. He submitted his refund application. He provided his grandmother's death certificate.
Air Canada refused. They told him bereavement fares cannot be claimed after travel has already happened.
He showed them a screenshot of the chatbot's exact words. An Air Canada representative looked at it and admitted the chatbot had provided misleading words.
They still refused to refund him.
In February 2024 the British Columbia Civil Resolution Tribunal ruled against Air Canada.
Air Canada's defence was remarkable. They argued the chatbot was a separate legal entity responsible for its own actions. Not Air Canada's problem.
The tribunal found this argument astonishing. A chatbot is part of your website. Your website is your responsibility. It does not matter whether the information comes from a static page or from a bot. If it is on your site you own it.
Jake Moffatt was awarded 812 dollars. Not a life changing sum. But the ruling set a precedent that every engineering team building customer facing AI should read carefully.
I spent 25 years in software and I can tell you exactly how this happened. Not the legal part. The team part.
There were two governance failures here. Neither of them required advanced technology to prevent. Both of them required someone to ask the right questions before launch.
The first failure happened the moment the chatbot spoke.
When the chatbot told Jake he could apply for a bereavement refund retroactively it was working from outdated policy data. The actual policy said no retroactive claims. The chatbot said the opposite.
That conversation should have been logged. If the chatbot's response contradicted current policy it should have been flagged automatically. A responsible team should have been alerted that a customer just received advice that does not match what the policy page says.
None of that happened. The conversation disappeared into silence. No log. No flag. No alert. No human ever reviewed it.
Jake was not the only person who asked that question. He was just the one who fought long enough to reach a tribunal.
The second failure happened when Air Canada found out.
An Air Canada representative saw the screenshot. They admitted the chatbot had given misleading words. They noted the issue to update the chatbot.
And that was it.
No audit of how many other customers received the same incorrect advice. No investigation into how many people booked full fare tickets trusting information that was wrong. No review of whether other chatbot responses contradicted other policies. Just a note to fix this one thing and move on.
That is not a post mortem. That is burying the problem.
But underneath both of those failures was something more worrying.
There was no automated monitoring of chatbot responses against current policy. There was no exception flagging. There was no escalation path. There was no responsible team to escalate to. There was no audit trigger when a complaint revealed a systemic issue.
The governance structure was not weak. It was absent.
Which means Jake Moffatt's tribunal claim was the only detection mechanism Air Canada had. A grieving customer fighting for 812 dollars was the monitoring system.
Think about that for a moment. A company deployed a customer facing AI tool that was making promises on its behalf. And the only way they would ever know the tool was giving wrong advice was if a customer complained loudly enough and long enough to force the issue.
I have been in hundreds of sprint planning sessions across 25 years. The questions that never get asked are not the hard ones. They are the obvious ones that get crowded out by the pressure to ship.
Who reviews chatbot responses when they contradict current policy.
What is our escalation path when a customer says the AI gave them wrong information.
Who owns the accuracy of this system after it goes live. Not who built it. Who owns it. Every day. Every time the business changes a rule. Every time a new edge case appears.
If those questions do not have named answers before you launch you do not have a governance model. You have a chatbot and a hope.
The Air Canada chatbot did not malfunction. It worked exactly as it was built. The training data was out of date. Nobody had built a process to keep it current. Nobody had built a system to detect when it went wrong. Nobody had defined who was responsible when it did.
That is not a technology failure. It is three governance failures sitting on top of each other.
And the only reason we know about it is because one person refused to give up.
Jake Moffatt got 812 dollars. His grandmother was already gone. He spent months fighting for a refund while grieving.
The amount does not matter. What matters is that a grieving person trusted a company's AI tool on the worst day of his year. And the only thing standing between that trust and accountability was his own persistence.
That is what happens when teams ship without governance. Not frameworks. Not policies. Just three simple questions with named answers before you go live.