What should you do when your AI assistant gives a customer wrong information?
Act in this order. If the error is serious, switch the assistant off or to message-only. Find the conversation and the customer, and correct it yourself in your own words. Check your agreement and policy before promising anything. Then find the cause, fix it, add the customer's question to the questions you test the assistant with, and log it. Fix the cause, not only the one answer, because the same gap can give the next customer the same wrong answer.
Right away
- Contain it. If the error is about a price, a promise, safety, health or money, or involves another customer's details, switch the assistant off or to message-only (it takes details and says a person will reply) until you know the cause. Decide in advance who may do this; the related answer on what to ask before hiring someone to build AI covers the off switch.
- Find the conversation and the customer. Pull the exact words, the time and the contact details. Then search the assistant's saved conversations, if the tool keeps them, for the same question; other customers may have been given the same answer.
- Correct it yourself, in your own words. A person should call or write, say what is true and say sorry for the confusion. Put the correction in writing, as the related answer on what to say when a customer repeats something wrong an AI told them describes: by email, or by text only with the person's consent.
- Check before you promise. Read what the customer signed or was sent, and your written policy, before offering to honor a price, a date or a refund. Whether you are bound by what an assistant said is a legal question that depends on the facts and is not answered here; ask an attorney if money, safety or a signed agreement is involved.
Then find the cause
Check which row fits. The related answers on stopping an AI assistant from making things up and on what an AI assistant should never say cover how to write the facts, the limits and the hand-off.
| What you find | What it often means | The fix |
|---|---|---|
| It had no fact on the topic, or an old one | The facts it works from are missing or out of date | Add or update the fact and remove the old one |
| It answered something it should have passed on | No written limit or hand-off for that topic | Add the limit and the plain line it says instead |
| It answered a different question than the one asked | Your wording is unclear, or two items conflict | Rewrite so only one reading is left |
| It should have alerted someone, and no one saw it | The alert went to the wrong person or to no one | Fix the alert and name a backup |
Make it stick
- Add the question, in the customer's own words, to the real-question set described in the related answer on stopping an AI assistant from making things up, and run the whole set again after the fix. A change that corrects one answer can alter another.
- Log it: the date, the customer, what the assistant said, what is true, the cause, the fix, who made it and when the assistant went back on.
- Switch it back on only after the test set passes, then read its next conversations yourself for a few days. If the same kind of error comes back, narrow its job or keep a person on that topic.
Which answer is which
This page is about your own assistant. The related answer on what to say when a customer repeats something wrong an AI told them is about public AI answers such as ChatGPT, and the related answer on whether a person should approve what AI sends to customers covers keeping a log of what was sent.
Related answers
- How do you stop an AI assistant from making things up?
- What should you say when a customer repeats something wrong an AI told them?
- What should an AI assistant never say about your business?
- What should you ask before hiring someone to build AI for your business?
- Should a person approve what an AI system sends to customers?
AI assistants and chatbots
A good assistant answers the question a customer asks at nine at night, in your words, from your information, and brings in a person when it should. We build that, and we test it on your real questions before a customer sees it.