Guide
Your AI assistant answers customers wrongly? Five checks to keep it on track
An assistant that invents prices, promises or policies costs you trust. Five practical checks: a narrow scope, approved sources, a human for anything that matters, a clear handoff and a test list you re-run.
The short answer
An AI assistant that talks to your customers should only answer from material you approved, and hand over to a person for everything else. Check five things. First, give it a narrow job, such as opening hours, delivery status or the first reply to a form message. Second, let it answer only from approved sources, and say so when it does not know. Third, keep a person in the loop for anything about prices, contracts, refunds or health. Fourth, make the handoff to a human easy and visible. Fifth, write a list of real customer questions and re-run it after every change. Wrong answers usually come from a vague job and no approved source, not from the model alone.
In this guide
Who this is for
You already use an AI assistant, or are about to, to answer customer questions: a chat widget, a first reply to contact-form messages, or a draft for your support team. Sometimes it sounds sure and is wrong: a price you never quoted, a policy that does not exist, a promise nobody can keep. This guide lists five checks that keep an assistant useful without letting it speak for your business where it should not.
You do not need technical knowledge to run the checks. Changing what the assistant can read or do may need whoever set it up.
The five checks
Give it a narrow job. Write the job in one sentence: “answer questions about opening hours and delivery times”, “draft a first reply to contact-form messages for a person to approve”. A job like “help customers with anything” invites invented answers. List what it must not do: quote prices, accept orders, give legal, medical or financial advice, change bookings.
Answer only from approved sources. Decide which pages, documents or records the assistant may use, and keep them current. Tell it to say “I do not know, I will pass this to a colleague” when the answer is not in those sources, instead of guessing. Test this on purpose with a question you know is not covered.
Keep a person in the loop for anything that matters. Anything about prices, contracts, refunds, deadlines, health or personal data is checked by a person before it reaches the customer. A simple pattern is that the assistant writes a draft, saved as a private note on the conversation, and a person sends it. Start there, and widen only what you have seen work.
Make the handoff to a human easy. Show a clear way to reach a person in every conversation, and make sure the assistant offers it when the customer is upset, asks twice, or the topic is outside its job. Check that the handoff arrives: send one yourself and see who receives it and how fast.
Test with real questions, and re-test after every change. Collect twenty real customer questions, including awkward ones, and write down what a good answer looks like. Run them after any change to the assistant, its sources or its instructions. Keep the list: it is how you notice a regression before a customer does.
If customers talk to AI on your site, tell them. Our guide Does your website need an AI notice? covers that.
A fifteen-minute self-check
- Write the assistant’s job in one sentence. Can you?
- Ask it something outside that job: does it hand over instead of guessing?
- Ask it for a price: does it avoid inventing one?
- Ask for a person: how long until someone answers?
- Do you have a list of real questions, and when did you last run it?
When to get help
Ask someone experienced if:
- the assistant has already given a customer a wrong price, promise or policy
- you want it to read company data such as orders or bookings, and need access rules
- you want drafts reviewed by a person inside your existing tools, such as your helpdesk
- you cannot tell which source an answer came from
We go through your setup with you in a free 30-minute video call: book a call. The practical side of this work is described on our AI page.