Will an AI support bot make things up to my customers?
An AI support bot will make things up if it is allowed to answer from general knowledge. It will not if it can only answer from your published articles, says so when nothing matches, and hands the question to a person. That is how DeskCrew is built: grounded answers or a handoff, and ticket replies drafted for your approval.
Last updated: October 2026 · View as Markdown
Why does this keep happening?
Language models are built to produce a plausible answer. Asked about your refund policy with nothing to read, one will produce a plausible refund policy. It will sound like yours and it will be wrong, and the customer will hold you to it.
That is not a theory. In February 2024 a Canadian tribunal ordered Air Canada to pay a passenger after its chatbot invented a bereavement refund rule. The airline argued the chatbot was responsible for its own words. The tribunal said it makes no difference whether information comes from a static page or a chatbot.
So the question is not whether a model can hallucinate. It is whether the product lets it. The safe design is simple: the bot answers only from what you wrote, and when it has nothing, it says so.
How do I fix it, step by step?
Give it only your material
Publish your answers as knowledge base articles. The DeskCrew AI answers from those articles and nothing else. There is no setting that lets it answer from the open internet.
Make "I do not know" a real answer
When no article covers the question, the widget says so and offers to pass it on. A handoff costs a few minutes. A confident wrong answer costs a customer, or a tribunal.
Keep a human on ticket replies
Replies to tickets are drafts in an approval queue by default. You read, edit and send. You can let a workspace send drafts on its own later, once the articles have earned it.
Read the unanswered questions weekly
Every handoff is a question your articles did not cover. Some become articles. Some should stay human, like anything with money or an account.
Write the policy in plain words
A vague article produces a vague answer. "Refunds within 30 days of delivery, to the original payment method, within five business days" leaves nothing for a model to fill in.
What is true about this today?
| Fact | Value | Source |
|---|---|---|
| Air Canada, after its chatbot invented a bereavement refund rule | Ordered to pay the passenger; "it makes no difference whether the information comes from a static page or a chatbot" | B.C. Civil Resolution Tribunal, reported by CBC News, February 2024 |
| Consumers who switch to a competitor after one bad experience | More than half, and 73% after several | Zendesk Benchmark data, updated January 2026 |
| Customers who want to resolve as many issues as possible on their own | 69% | Zendesk, self-service research, July 2021 |
| DeskCrew AI answers | From published knowledge base articles only; ticket replies are drafts for approval by default | DeskCrew AI policy, October 2026 |
What changes, and what does not?
Grounded means narrower
A bot that only quotes your articles cannot help with what you have not written. That is the trade. The first month you will write articles from the handoffs, and the bot gets more useful without getting less honest.
A wrong article makes a wrong answer
Grounding stops invention; it does not check your facts. If the article says the wrong return window, the bot says it confidently. Keep the articles current and the answers follow.
Approval slows day one
Reading every draft is work. It is also how you learn whether the articles are good enough to send on their own. Most workspaces keep approval on for money and account questions permanently.
What else do people ask?
How is "grounded" different from a normal chatbot?
A normal chatbot answers from its training, which includes the whole internet and none of your policies. A grounded one is only allowed to answer from the documents you gave it, and must say when they do not cover the question. The difference is the handoff: a normal bot never admits it has nothing, a grounded one does by design.
What does the customer see when the bot does not know?
A short message saying this is not covered yet and offering to pass it to a person, with a field for their email if they have not given one. The conversation becomes a ticket with the full transcript, so your reply picks up where the bot stopped and the customer never has to repeat the question.
Can it still be wrong?
Yes, in one way: if your article is wrong or out of date, the bot repeats it. It cannot invent a policy, but it can quote an old one. The fix is to keep the articles current, and the weekly review of handoffs and drafts is where you catch the stale ones before customers do.
Should I let the AI send replies without approval?
Not on day one. Read the drafts for a few weeks. When they are consistently right for a type of question, let that workspace send on its own, and keep approval for anything involving money, accounts or a complaint. The default is approval, and switching it off is a deliberate choice you make per workspace.
Who is responsible if the bot gives a wrong answer?
You are, the same as for any page on your site, which is exactly what the Air Canada tribunal said. That is why the design matters: a bot that quotes your published policy puts you in the same position as your policy page, and a bot that improvises puts you somewhere you have never read.
AI support widget · AI policy · Security · Stop answering the same five questions · All how-to guides