"The chatbot said it, not us." Two courts have now said: yes, you.

In 2024 a Canadian tribunal ordered Air Canada to honour a bereavement discount its chatbot had invented. The airline's argument was that the bot was 'a separate legal entity responsible for its own actions'. The tribunal called that remarkable, and ruled against it. In May 2026 a German appeals court reached the same conclusion in a different dispute: the company publishes the answer, so the company owns the answer.

That's the legal reality small teams are now deploying into. It doesn't mean don't use AI on your site. It means the question 'what is this thing allowed to say?' is now a business question, not a prompt-engineering one.

The failure is structural, not rare

Independent measurements of live support bots put the wrong-answer rate somewhere between one in seven and one in four. That isn't a bad vendor. It's what a general-purpose language model does when asked something it wasn't given: it produces the most plausible shape of an answer, in your brand's voice, with total confidence. Your returns window becomes 60 days because that's a common number. Your plan includes a feature because the sentence flowed.

A bot that's wrong 15 percent of the time on policy questions is a bot writing you 15 percent of your policy. You just don't find out which parts until a customer holds you to one.

Three controls that actually change the risk

One: ground every answer in your own pages. The agent retrieves from your live site at question time and can only assert what it found. If retrieval comes back empty, the right output is 'I don't know, here's how to reach us', not a guess. That single rule removes the entire category of invented policy.

Two: show the receipt. A citation on each grounded answer lets the customer check, and lets you audit. When something looks off, the citation points to the page that produced it. Nine times out of ten, the page was ambiguous and needs a rewrite. The fix is on your website, not in the model.

Three: define what the agent must not decide. Refunds over a threshold, cancellations, anything with a legal edge: these get an escalation rule that pulls a human in, notifies them wherever they are, and lets them take over the live chat. The agent explains, the human commits. That split is what the courts are effectively asking for.

The upside of being careful

Teams sometimes read all this as 'AI is a liability, keep it away from customers'. The opposite is true. A grounded agent that cites sources and declines gracefully is more trustworthy than the average human reply, because it can't drift from the written policy and it never gets tired at 11pm. It also produces a permanent, searchable record of exactly what was said and why.

Put a general chatbot on your site and you've hired someone who makes things up. Put a grounded one on and you've published your policies in a form people will actually read.