A Badly Configured Chatbot on Your Site Can Drive Clients Away: When to Use One
By Anis Hammouche·November 6, 2026
A badly configured chatbot drives clients away because it answers with confidence, but gets the answer wrong. It invents a price, promises a refund that doesn't exist, or states a rule you never wrote. The client believes it, and the business is held to it. A chatbot stays useful when its questions are closed and its answers come from a source you have validated. Outside that scope, a form or a person does the job better.
What one invented sentence cost
In November 2022, a passenger from British Columbia, Jake Moffatt, lost his grandmother. He used the chatbot on the Air Canada website to find out whether the bereavement fare could be claimed after the flight. The bot said yes, and that the request could be made retroactively. That was not the airline's policy.
The British Columbia tribunal ruled in February 2024 (decision 2024 BCCRT 149). Air Canada argued that the chatbot was a separate entity, responsible for its own statements. The tribunal rejected that argument: the bot is part of the airline's website, and the airline answers for what it displays. The award covered the difference between the full fare and the bereavement fare, 812.02 Canadian dollars in total, interest and fees included, according to published reports (Mondaq analysis).
The amount is modest, but the lesson reaches far: an automated answer on your site is a statement from your business. If it is wrong, the client does not have to prove you lied. They only have to prove they relied on your page.
Why a badly configured bot drives clients away
The mechanism is simple. A language model always answers, even when it doesn't know. Without a clear instruction, it fills the gaps with a plausible reply. To a client, a plausible answer and a correct one look alike. They notice the error when it costs them: a non-refundable ticket, a wrong quote, a missed appointment.
The client does not forget the answer, they forget the business. A single false piece of information is enough to make them doubt the rest of the site.
When to use one, and when to skip it
A chatbot is justified when questions are repetitive and answers are stable and sourced. It is not justified when the answer depends on a human decision or a personal situation.
| Situation | Suitable chatbot? | Why |
|---|---|---|
| Opening hours, address, delivery times | Yes | Fixed answer, easy to verify |
| Frequently asked questions about a product | Yes | Answers drawn from your validated FAQ |
| Booking or order tracking | Yes, with a link to the real tool | The bot should point to the data, not guess it |
| Refunds, disputes, complaints | No | A human decision creates your liability |
| Negotiated prices, special terms | No | The bot doesn't know the agreement made with that client |
| Legal, medical or tax advice | No | Refer to a professional, with official sources |
The practical rule: if a wrong answer costs money or creates a dispute, the bot does not decide alone.
How to set it up so it promises nothing
The setup matters more than the choice of tool. Four elements make the difference.
First, anchor answers in your validated documents. The bot only answers from your FAQ, your terms and your published prices. If the answer isn't there, it says so. An instruction like this limits drift:
Answer only from the documents provided (FAQ, terms of sale, price list).
If the answer is not there, say so clearly and suggest contacting
[address or form]. Never infer a rule, a price or an exception.
Never promise a refund, a deadline or a commercial gesture.
Always state which document your answer comes from.
Second, say it is an AI. Since 2 August 2026, the European AI Act requires people to be told they are interacting with an AI system, at the latest at the first exchange (Article 50, EUR-Lex). The notice must not be buried in terms of use. Check your case with a lawyer, but do not delay the notice.
Third, provide a route to a human. A "talk to someone" button should be on the first screen, and it should work. On the topics listed in the table, the bot points to that button without pushing.
Finally, review the conversations. Keep a log, and each week read the questions the bot answered without a source. That is where invented answers hide.
Test before you switch it on
Before going live, prepare twenty real questions, taken from your emails and calls. Put them to the bot and record each answer as exact, incomplete or false. An incomplete answer is fixed with the FAQ. A false answer blocks the launch until the instruction or the documents are corrected.
Also test the trick questions: "Can I get a refund if I change my mind?", "Do you make an exception for me?", "Is my case covered?". A good bot says it doesn't know, and points to a human.
Frequently asked questions
Does a chatbot improve conversion? It can cut waiting time on simple questions. Measuring its effect means tracking the conversations and the requests that arrive afterwards by email or phone, not just the number of exchanges.
Do I need a specific tool? The framework matters more than the tool. A well-configured bot on a simple tool is better than a badly configured bot on a sophisticated platform.
Can a chatbot replace customer service? Not on matters that commit the business. It absorbs repetitive questions and leaves cases that require judgment to people.
What if the bot has already given false information? Correct it immediately, contact the client concerned, and update the instruction and the documents. Keep a record of the incident. For legal consequences, consult a professional.
For a broader framework on the costs of an AI tool, also read the real cost of AI and its three hidden costs.
