AI customer service fails the way a bad hand-off does: it answers the question and misses the customer. The fix is not a smarter bot. It is a bot that knows when to stop talking and hand the customer to a person.
I made a similar point about a famous hotel soap story in an earlier post on how to listen to your customers. Four staff members each answered a guest’s one-line request with a correct policy, and he ended up with 54 bars of soap. Now put a chatbot in that chain and let it answer at 2am with total confidence.
What did the Air Canada chatbot case actually show?
In 2022, Jake Moffatt asked Air Canada’s website chatbot about bereavement fares after his grandmother died. The bot told him he could book at full price and claim the reduced fare afterward. That was wrong. The airline refused the refund.
The airline’s defence is the part worth remembering. It argued the chatbot was a separate legal entity, responsible for its own actions. In February 2024 a British Columbia tribunal disagreed and found negligent misrepresentation. The American Bar Association’s summary puts it plainly: the company bears responsibility for everything on its site, whether a static page or a chatbot wrote it. The award was C$812.
The money was small. The lesson was not. The bot answered the question it was asked, fluently, and nobody owned the answer.
Why is a wrong answer worse than no answer?
A customer who gets no answer keeps looking. A customer who gets a confident answer acts on it, then finds out the answer was wrong when it costs them something.
That is the soap letters again. Every reply in that story was polite, on-policy and useless. AI makes it faster, cheaper and available around the clock. It does not make the reply more useful by default.
Bad news for the roadmap: the tooling is getting better, and customers are getting less patient with the failures that remain.
What do customers actually want from AI customer service?
A way out. That is not a guess. Gartner surveyed 3,566 customers in February and March 2026, and 87% said it is essential to have an option to reach a human agent when a company uses generative AI. At the same time, 50% said their interactions were easier when companies used it.
Read those two numbers together. Customers will use the bot. They will not tolerate being trapped by it.
When Gartner asked people who were unwilling to engage with AI what might change their minds, the most common answer was being able to switch to a human. The escape hatch is what makes the bot usable.
How do you design AI support that listens?
Start from the hand-off, not the model. Four checks, all of them practical.
First, restate before you resolve. Have the bot repeat the request in plain words and ask “is that right?” before it acts. It is the same ten-second habit I want from agents, and it catches the Air Canada type of error before the customer relies on it.
Second, make the human path one step. Do not bury it under three menus and a “Did that answer your question?” loop. If a customer asks for a person twice, they get one.
Third, pass the whole conversation across. Nothing burns trust faster than a customer explaining their issue to a bot and then explaining it again to an agent. A clean transfer carries the request, what the bot said, and what the customer tried. It is the same rule as an escalation matrix: decide the route before the problem arrives.
Fourth, own the answers. Anything the bot tells a customer is something your company told them. Review its answers on refunds, eligibility and policy the way you would review a new agent’s first week.
How do you know whether your AI is really helping?
Do not lean on containment rate alone. A bot that “contains” a customer who gave up is a bad result that looks like a good one.
Track what happens after the bot conversation. How many customers contact you again within a few days about the same issue? How many ask for a human in the first two messages? How many escalations arrive with the customer already angry? Those numbers tell you whether the bot resolved the problem or just closed the session.
I looked at where AI is worth buying and where it isn’t in AI in Your Support Stack, and at the agentic side in a practitioner’s reality check. The common thread is the same: the tool is only as good as the hand-off around it.
Frequently asked questions
Is AI customer service worth it?
Yes, for routine, well-defined requests where a fast answer is the whole job. It stops paying off when it blocks access to a person for anything unusual. Customers accept the bot when they can leave it.
Are companies responsible for what their chatbot says?
In the Air Canada case, a British Columbia tribunal held the airline liable for its chatbot’s incorrect bereavement fare advice. Treat every bot answer as your company’s own statement, and review it that way.
How should an AI chatbot hand off to a human agent?
Quickly and with the whole conversation attached. The customer should never have to repeat their request. Set a firm rule, such as a human after two requests, so the customer is not stuck in a loop.
What metrics show whether AI support is working?
Repeat contacts on the same issue, requests for a human in the first few messages, and customer effort after the bot conversation. Containment rate on its own can hide customers who simply gave up.
Hutch Morzaria is a Director-level CX and Support Leadership professional with 19 years of experience building global support organizations across SaaS, Fintech, and enterprise technology. He has hired dozens of support and CX leaders across his career and holds ITIL Expert certification across V3 and V4.
