Organisations that deploy AI agents can hand over tasks, but AI accountability stays with them. Because when a system promises a customer something incorrect, a person ultimately has to explain how it happened. Few cases illustrate this as clearly as a dispute between a grieving grandson and Air Canada.
A Bereavement and a Wrong Answer
In November 2022, Jake Moffatt from Vancouver loses his grandmother and needs to fly to Toronto at short notice. On Air Canada’s website, he asks the chatbot whether there is a discounted bereavement fare. The answer is yes, and he can apply for the discount up to 90 days after purchasing his ticket.
So he books at full price and then submits his application. Air Canada, however, refuses. The chatbot had made a mistake, the airline says, and the correct rule was on a different page of the website.
The Argument That Travelled Around the World
Before the Civil Resolution Tribunal of British Columbia, an online tribunal for smaller disputes, the airline then put forward an astonishing argument. The chatbot, it claimed, was a separate legal entity responsible for its own actions. In 30 years of crisis communication, I had never heard anything like it.
The tribunal called it a “remarkable submission”. In February 2024, Air Canada had to pay the customer around 800 Canadian dollars.
Let us do the maths for a moment. For a claim worth 800 dollars, a headline travelled around the world in which a national airline declared that it was not responsible for its own website. Simply refunding the man the difference would have been considerably cheaper.
If the AI chatbot makes false promises, the responsibility lies with the company—and a human must take the blame. Photo: ChatGPT
AI Accountability: Delegate, Yes. Deflect, No.
With AI agents, situations like this will become more frequent. After all, these systems answer enquiries, prepare quotes and commit to terms.
I am currently building such agents myself, namely for our CRM, for quotes and for travel expense reporting. The first decisions along the way, however, were not technical ones. For us, security is the top rule.
That is why our AI has no access to our emails. Instead, it receives data only as a copy on a separate server. There, it may prepare, calculate and make suggestions. Sending anything, on the other hand, is not allowed, and neither is deciding.
Yes, that is slower than it could be. Yet that is exactly how I want it.
Tasks can be delegated, but responsibility remains with the organisation. Who is ultimately liable is a legal question to be settled case by case. Who has to explain it, by contrast, is not a question at all.
An angry customer, after all, does not want to know which language model was in use. They want to know why it happened, who allowed it and how you will prevent it from happening again. Anyone without a clear answer turns a technical error into a crisis of trust. Air Canada demonstrated exactly that.
Rules on Paper Are Not Enough
In September, EY published a survey of 202 AI leaders at companies with revenues of more than one billion dollars. According to the results, 98 percent have formal governance rules. However, 47 percent say these rules have already been set aside for urgent projects. And 36 percent have already experienced an AI incident with noticeably negative consequences.
Almost every company, therefore, has rules. Yet one in two suspends them as soon as things get urgent.
I know this pattern well from crisis communication. The crisis manual sits in the cupboard. Whether anyone actually knows what to do, though, only becomes clear at 7:40 on the morning after.
Three Questions Before Your Next AI Project
- Who really decides? What may the system do on its own, where must a person review it, and who is responsible for which process? Settle this beforehand, not afterwards.
- Can you explain it in three sentences? Specifically, to a customer who has been harmed by an automated decision. If you cannot, you have a communication problem.
- Who says what when things go wrong? Prepare your messages, run through scenarios and designate spokespeople. The G.A.M.E. Code helps here: Goal, Analysis of the audience, Message and Extras.
No Reason to Avoid AI Agents
To avoid any misunderstanding: I see no point in refusing to use AI agents out of fear of mistakes. People in customer service make mistakes too. The difference lies elsewhere. When an employee makes a mistake, the company apologises. Air Canada, by contrast, tried to offload the mistake onto the tool.
An AI can draft an apology. Delivering it, however, takes a human being.
If your system promises a customer something incorrect tomorrow, who in your organisation will call them?
If no name comes to mind immediately, you have found the real gap.
Dr. Nikolai A. Behr CSP® ist Keynote Speaker, Kommunikationsexperte und Medientrainer für Führung, Vertrauen und empathische Kommunikation in Zeiten von Wandel und KI.