In 2022, a Canadian man named Jake Moffatt asked Air Canada's website chatbot about bereavement fares after his grandmother died. The bot told him he could apply for the discount retroactively, after booking. That was wrong, and Air Canada spent two years arguing before a small-claims tribunal that it should not be held responsible for what its own chatbot had told a customer. In February 2024 the tribunal disagreed and ordered the airline to pay. The ruling has since become the reference point support organizations cite when they argue for the harder, more expensive project of connecting a bot's answers to something that can actually be checked.
For a stretch of years the industry's proof point was Klarna's early-2024 announcement that its OpenAI-built assistant had handled roughly 2.3 million conversations in its first month, work the company said was equivalent to about 700 full-time agents. That framing, a contact avoided is a cost avoided, spread fast, and containment rate became the number vendors sold and support leaders reported upward. The Moffatt ruling complicates that framing in a way a containment dashboard cannot: a contact resolved by a bot is not resolved at all if the bot was wrong, and now there is a legal record saying so.
The metric shift shows up in how contracts get written. Intercom now sells its Fin agent on a per-resolution basis rather than per seat, which forces a definition of resolution precise enough to bill against. Zendesk and Salesforce's Agentforce have moved toward similar outcome-based framing. A contact the bot handled but that resurfaces as a new ticket within a week no longer counts as a win under these models. It counts as a failure that took two tries to notice.
Systems built purely to deflect, narrow intent classifiers routing to canned answers, struggle under this accounting, because they were optimized to end a conversation, not to fix an account, a shipment, or a fare policy. What is opening up instead is demand for support AI with write access to billing systems, order management, and reservation records, with the authority to issue a refund or waive a fee rather than only explain the policy governing one.
That authority is the harder engineering problem, and it is exactly what Air Canada discovered the expensive way. Giving a model write access to a customer's account multiplies the cost of a mistake, so support organizations are now building the same approval chains and escalation triggers around AI agents that they once built around junior human staff. The interesting work has shifted from language understanding to permissioning: what can this system do without anyone watching, and what should always require a person.
Moffatt never wanted a lawsuit. He wanted a bereavement discount and got a chatbot's confident, wrong answer instead. Every support team deploying an agent with a refund button now is making a bet about how many Moffatts it can absorb before a tribunal, somewhere, makes the decision for it.