In 2022, Jake Moffatt lost his grandmother and needed to book a last-minute flight for the funeral. On Air Canada's website, he asked the company's AI assistant about the bereavement fare. The answer: he could book at full price and request the fare difference back within 90 days. That was wrong — the airline's actual policy said the opposite — but he followed the advice, bought the ticket, and asked for the refund afterward. Air Canada refused.
The case ended up in a British Columbia small claims tribunal. Air Canada's defense, according to the ruling itself, tried an unusual argument: the AI assistant was "a separate legal entity," responsible for its own statements. In February 2024, the tribunal rejected that argument outright — a company is responsible for everything published on its own website, AI or not — and ordered the airline to pay the fare difference plus costs.
A year later, on the opposite end of automated customer service, Swedish fintech Klarna faced the reverse problem. It had replaced much of its support with an AI assistant that, by the company's own account, did the work of 700 agents. In May 2025, CEO Sebastian Siemiatkowski announced the return of human agents. The line that sums up why: "we went too far."
The price of a wrong answer
The Moffatt case looks small — the fare difference involved a few hundred Canadian dollars. But the precedent it left is large: anything an AI assistant states on an official company channel counts as a statement from the company itself, with the same consequences as an employee saying the same thing over the phone. There's no category of "that was just the bot talking" that shields a company from responsibility.
The most telling detail in the ruling isn't the damages amount. It's the tribunal's reaction to the defense argument. Calling the assistant a "separate entity" was described as a claim that should be too obvious to need explaining — of course a company is responsible for what it publishes on its own website. The precedent became an international reference precisely because it made explicit something many companies still treat as a gray area: text that comes out of AI isn't a draft, it's a statement.
That matters because most companies putting an AI assistant in front of customers still don't have a ready answer to the most basic question: when the assistant states something about policy, deadlines, or price, where did that information come from, and did anyone approve it being used that way? Without that answer, every conversation is a fresh gamble.
Why cutting costs on quality gets expensive
The Klarna case shows the problem from the opposite angle: not wrong information, but degraded experience in the name of cost. In February 2024, the company launched an AI assistant that, in its first month, handled 2.3 million conversations — volume equivalent to 700 full-time agents. From an efficiency standpoint, the result was impressive.
But perceived quality dropped. According to reporting from the period, customers complained about impersonal interactions and about how hard it was to reach a person when a case fell outside the common script — disputes, complex refunds, financial hardship situations. The CEO himself summed up the diagnosis: cost weighed too heavily in the decision, and the result was lower quality. In May 2025, the company announced it was hiring human agents back, in a hybrid model: AI for volume and speed, a person for cases that require judgment and empathy.
The pattern shared between the two cases, despite looking opposite on the surface, is the same one: neither company had clearly defined where what AI can resolve on its own ends, and where a person needs to step in begins. Air Canada let its AI state policy without checking. Klarna let its AI cover situations too delicate for an automated script. Both mistakes come from the same place: treating AI-powered support as one single thing, instead of a set of decisions that need a boundary.
The lesson applies just as much, if not more, to companies across Latin America, where much of customer service already runs through automated assistants on channels like WhatsApp. The legal principle behind the Moffatt case isn't a Canadian quirk: in practically any market with consumer-protection law, a company is responsible for what its official channel states, regardless of who — or what — wrote the message.
What has to be in place
Avoiding both problems at once — wrong information and degraded experience — requires four specific mechanisms, not a generic "use AI responsibly" policy.
Approved knowledge with a source, not loose text. The answer about policy, deadlines, or price needs to come from a document with an owner, a current version, and a defined scope — never from AI's own generalization. When the refund policy changes, the document changes once, and every future answer reflects the change immediately.
Human approval before a sensitive action. When the conversation moves outside the common script — an exception, a complaint, an out-of-range amount — the right flow is to pause and get confirmation from a person before promising anything to the customer, right inside the conversation. This is exactly the kind of situation where Klarna saw the experience break down.
Escalation rules defined in advance, not discovered on the fly. Instead of leaving each AI agent to decide on its own when to bring in a person, the criteria are set beforehand: which types of requests always go to a human, regardless of how confident the AI is that it can answer.
An audit trail of what was said. Every relevant answer stays on record — what was stated, based on which document, at what moment. If a customer later says "the AI guaranteed me this," the company can check exactly what happened, instead of arguing from memory.
This is how Skyller was designed: answers grounded in approved corporate knowledge, human approval before sensitive actions, and an audit trail by default — not an advanced setting someone has to remember to turn on.
From risky answers to defensible service
What the two cases teach, read together, is that the risk in AI-powered service isn't concentrated in one place. It sits at both ends: in what AI states without checking, and in what it tries to resolve on its own when it should escalate. A well-designed limits policy addresses both.
The practical gain from getting this right isn't abstract. It's the difference between a company that can show, in a dispute or a complaint, exactly which document backed an answer and who approved that content — and a company that only has the customer's word against the memory of an automated conversation. It's also the difference between a customer who feels there's a way out when a case is delicate, and a customer who feels trapped in a script with no escape.
Neither case happened because AI "didn't work." Both happened because the company hadn't decided, ahead of time, where AI stops and a person steps in.
Three questions for the next meeting
Before expanding AI-powered service to more channels, it's worth answering these questions with legal and operations:
- Does every answer about policy, deadlines, or price come from an approved document, with a defined owner and version? If the answer comes from AI's own generalization, the risk is exactly the one from the Air Canada case.
- Is there a clear list of situations that always escalate to a person, defined in advance — not decided by the AI on the fly? Disputes, exceptions, and sensitive financial cases should be on that list by default.
- Can the company reconstruct, months later, what a conversation stated and on what basis? Without that trail, any complaint turns into one word against another.
- Does the customer know, at any point in the conversation, that they can ask to speak with a person? Keeping that door visible is exactly what Klarna admitted having closed too tightly.
- Who reviews the content before it becomes an official AI answer? If the answer is "no one, formally," that's the first boundary to close.
