When Your Chatbot Talks, Your Company Is on the Record

In early 2024, a Canadian tribunal ruled against Air Canada after one of its website chatbots gave a grieving customer incorrect information about the airline's bereavement fare policy. Air Canada's defence was, in effect, that the chatbot was a separate entity responsible for its own words. The tribunal rejected that outright: the chatbot was part of the airline's website, the airline was responsible for everything on it, and it made no difference that a link to the correct policy existed elsewhere on the page. That case has since been echoed rather than reversed. Search for how courts are handling AI-generated customer statements in 2026 and the pattern holds — a German regional court ruled this spring that a company was liable for misleading statements its chatbot made to a customer, on the same underlying logic: an AI system speaking on a company's behalf is the company speaking.

It is worth sitting with why this feels surprising to a lot of business owners, because the instinct runs the other way. A chatbot answer feels provisional — closer to a suggestion than a commitment, something a reasonable customer should double-check before relying on. Regulators and courts are not extending that grace. In the United States, the FTC has been treating chatbot output as a company representation for purposes of deceptive-practices enforcement, meaning an inaccurate or unsubstantiated claim from a bot carries the same exposure as the same claim in an ad or on a pricing page. The legal system's working assumption is now the plain-language one: if your company deployed the tool and the tool told a customer something, your company told the customer that.

The practical risk for most small and mid-sized businesses is not a chatbot behaving maliciously. It is a chatbot behaving exactly as designed — fluent, confident, and occasionally wrong. A modern conversational AI does not hedge the way a careful employee would when unsure; it answers in the same assured tone whether it is citing your actual return policy or quietly inventing a plausible-sounding one. A customer has no way to tell the difference from the outside, because a wrong answer reads identically to a right one. That is a different failure mode than a human employee giving bad information, where tone, hesitation, or a promise to check often signal the uncertainty. A chatbot rarely signals it at all.

The first real safeguard is scope, not supervision after the fact. A chatbot that answers from a narrow, curated set of the business's own current policies — pricing, returns, appointment availability, service coverage — is far safer than one given open-ended license to reason from general knowledge about how businesses like yours typically operate. The difference matters because the second kind will, sooner or later, fill a gap in its actual knowledge with a confident guess that sounds like policy. Keeping the bot's answers grounded in a specific, business-maintained source of truth, and having it decline or hand off rather than improvise when a question falls outside that source, closes off the failure mode that produced the Air Canada ruling in the first place.

The second safeguard is an ordinary management practice most businesses already apply to new hires and simply have not extended to software: someone reviews a sample of what it actually said. Chat transcripts are cheap to log and rarely get read unless something has already gone wrong. A short weekly review of a random sample — not just the flagged or escalated conversations — catches the wrong answer before it becomes the customer's evidence in a complaint, rather than after. Pair that with a fast path to update the bot's source material the same day an error is found, the same way a manager would correct a front-line employee's misunderstanding as soon as it surfaces rather than at the next quarterly review.

The third is recognizing what a disclaimer can and cannot do. A line of small print noting that the chatbot may make mistakes has become a standard inclusion, and it is worth keeping, but the Air Canada and Hamm rulings both establish that it is not a shield a court will treat as decisive when a customer reasonably relied on what the bot actually told them. The disclaimer manages expectations; it does not transfer legal responsibility for the specific answer given. The protection that actually holds up lives upstream of the disclaimer, in what the bot was allowed to say and how closely anyone was watching.

None of this argues against deploying a customer-facing chatbot — the productivity and availability gains are real, and pulling back from the technology solves nothing the underlying business need. It argues for treating the chatbot the way a well-run business already treats any other channel that speaks to customers on its behalf: with a defined scope, a named person accountable for what it says, and a habit of actually checking. A chatbot is often a business's highest-volume and least-supervised customer-facing voice. Given what courts are now willing to hold a company to, that combination is worth closing before a customer, not a regulator, is the one who closes it for you.