How to Stop a Chatbot From Giving Wrong Information: A 10-Point Checklist
Chatbot quoting old prices or inventing answers? 10 controls before and after launch that keep your AI agent accurate and honest when it doesn't know.
To stop a chatbot giving wrong information: give it one up-to-date source of information and tell it to answer from that source only; have it say "I don't know" and hand over to a staff member when it can't find the answer; set off-limits topics; require confirmation of important numbers; and review conversations weekly, updating the information with every change to prices and policies.
01Why do chatbots get things wrong?
- Outdated information: prices changed and nobody updated them.
- Missing information: it was asked about something not covered, so it guessed.
- Vague instructions: it was never told it may admit it doesn't know.
- Too much authority: it promises discounts or exceptions you don't offer.
02The 10-point checklist
- One source of information: an organised, up-to-date knowledge file.
- Answer from the information only: no guessing.
- "I don't know" is allowed: with a handover to a staff member.
- Off-limits topics: special prices, legal promises, medical advice.
- Confirm the numbers: order number, amount and appointment time.
- No promises outside policy: discounts and refunds need staff approval.
- Last-updated date: for every price and offer.
- Test before launch: 30 real questions, including some tricky ones.
- Weekly review: of handed-over and negatively rated conversations.
- Immediate updates: with any change in prices or policies.
03Example: the right reply when it doesn't know
This reply is far better than "Yes, we have a branch" and then the customer finding out there isn't one. Customers forgive "I don't know", but they don't forgive wrong information.
04Test your agent with these questions
The right replies: admitting the service isn't available, politely declining the non-existent discount with a handover, and a friendly apology for the off-topic question.
Key takeaways
- Most mistakes come from outdated or missing information.
- "I don't know" with a handover beats a wrong answer.
- Set off-limits topics and promises the agent never makes.
- Test before launch, and review weekly after it.
Frequently asked questions
Is it possible to reach zero mistakes?
No system is error-free, but guardrails cut mistakes significantly and route them to a staff member instead of reaching the customer.
Who should review the agent's conversations?
One person on the service team, at least once a week.
Can I connect the agent to live prices?
Connecting to live data reduces pricing mistakes. Read the MCP article and ask the Modeer 360 team about integrations.
- #AI mistakes
- #Chatbot hallucinations
- #AI agent guardrails
- #Bot reply quality