AI Chatbot Gave a Customer Wrong Information: What Now
A tribunal has already ruled that a company owns what its chatbot says. So the question is not whether you are responsible, it is what you do in the next hour.
Your AI chatbot gave a customer wrong information, they acted on it, and now someone is asking what you are going to do. The legal question is more settled than most people assume, so skip it and go straight to the operational one.
In February 2024 Canada's Civil Resolution Tribunal ordered Air Canada to pay a passenger who had been told by its chatbot that a bereavement discount could be claimed after booking, which contradicted the airline's actual policy. Air Canada argued it was not responsible for what the chatbot said. As CBC reported, the tribunal member called that submission remarkable and found that a chatbot is part of the company's website like anything else on it.
Assume you own what your bot says. Now here is the sequence.
The first hour
Preserve the conversation before anything else. Export the full exchange, with timestamps, the model version, and whatever context or retrieved documents were in play. Do this before touching the prompt or the knowledge base, because the moment you edit either, the evidence of what the customer was actually told starts degrading. If you cannot reconstruct a conversation from three weeks ago, that is a finding in itself, and it is what an audit trail of AI use exists to prevent.
Find out how many others got the same answer. One customer reporting it means the underlying error has been sitting there for however long, answering everyone who asked something similar. Search your logs for the same question shape rather than the same exact wording. This number changes everything downstream, because a single case is a customer service decision and forty cases is an incident.
Contain before you fix. If the topic is high-stakes, route it to humans immediately. A hard rule that sends any question mentioning refunds, cancellations, medical topics, or pricing to a person is crude and effective, and it buys you the time to fix things properly rather than under pressure.
Then talk to the customer. Do not lead with an explanation of how language models work. Acknowledge what they were told, confirm you are looking at it, and give a time you will come back. The technical account is not what they want.
Honour it or correct it
This is the decision people agonise over, and it comes down to four factors.
Reliance. Did they act on it? Someone who booked a flight, cancelled a service, or made a purchase based on the answer is in a different position from someone who read it and asked a human.
Plausibility. Was the wrong answer credible? A discount that plausibly could exist is different from a chatbot promising something obviously impossible.
Cost. What does honouring it cost against the cost of the argument, including the reputational component if the exchange is posted publicly.
Precedent. Would honouring it commit you to a rule you cannot sustain at volume?
For most small businesses, most of the time: honour it when the customer relied on it and the cost is bounded. The Air Canada figure was $812 in a case that has been cited in coverage worldwide ever since. Contesting a small commitment your own system made is rarely the cheaper path once everything is counted.
Where you genuinely cannot honour it, say plainly what went wrong, what the correct position is, and what you are doing instead. People forgive a system error explained honestly. They do not forgive being told they misread something they did not misread.
Then find the actual cause
"The AI hallucinated" is not a root cause, it is a shrug. There are four real ones, and they need different fixes.
Cause | What it looks like | Fix |
|---|---|---|
Stale source | The bot repeated a policy that changed in March | Sync the knowledge base to whatever holds the real policy |
Missing source | The answer was not in the sources at all | Add the content, and make abstention possible |
Retrieval miss | The content exists but was not fetched | Fix search, chunking, or phrasing coverage |
Ungrounded generation | The bot answered from general knowledge | Constrain it to sources and require evidence |
The last two are the common ones and both are addressable. A model with no support for an answer should be able to return a non-answer, which is a design property rather than a hope. That is the whole argument in making AI say I don't know instead of guessing.
Stop the repeat
Three changes, in order of how much they buy you.
Restrict what the bot can assert. Commitments about money, timing, eligibility, and policy are a different class from general information. Route them to a human or to a fixed, pre-approved answer that nobody generates on the fly. This is the core of setting guardrails for a customer-facing chatbot.
Make the knowledge base the same object as the real policy. Most of these incidents are documentation drift wearing an AI costume. If the policy lives in one place and the bot reads from a copy, the copy will be stale eventually, and the failure will surface through the bot.
Log everything, and read the logs weekly. Not to catch every error, but because the third time a similar question produces a shaky answer, you want to see it before a customer does.
The uncomfortable part
Somebody in your organisation has to own this before it happens, not after. When the answer to "who decides whether we honour this" is discovered during the incident, the response is slow and inconsistent, and that inconsistency is what turns a support issue into a complaint.
Name the person, write down the honour threshold in advance, and decide what topics the bot is not allowed to speak on. It is a twenty-minute conversation that changes the shape of every future incident. The wider framing sits in who is responsible when AI makes a mistake, and the operational scaffolding in an AI incident response plan.
FAQ
Are we legally bound by what our chatbot says?
The Air Canada ruling found the company responsible for information its chatbot presented, rejecting the argument that the bot was somehow separate. It is one tribunal in one jurisdiction rather than universal law, but it is the direction of travel, and planning around any other assumption is optimistic.
Does a disclaimer protect us?
Less than people hope. A footer saying answers may be inaccurate sits awkwardly against a bot that gave a specific, confident, actionable answer. It may help at the margin. It is not a substitute for not saying the wrong thing.
Should we tell other customers who got the same wrong answer?
If they may have acted on it, yes, proactively. Finding out later that you knew and stayed quiet is considerably worse than the original error.
How do we know it is not still happening?
Ask the bot the same question and its near variants, from a clean session, after the fix. Then add those questions to a small test set you re-run on every prompt or model change. If you cannot demonstrate it is fixed, it is not fixed.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
