Why your website chatbot should say "I don't know"
22 October 2026 · by Polycarp
A chatbot that always has an answer sounds like the goal. On a business website it is the problem.
The policy you never wrote
Ask an ordinary AI chatbot whether a wool blanket is machine washable and it will tell you, confidently, from whatever it learned about wool in general. Ask it whether you take returns after thirty days and it will produce a returns policy that reads well and may be nothing like yours. Your customer reads it as yours. Now you honour it, or you argue with a customer who has a screenshot.
That is the failure that matters for a small business. Not a wrong answer about the weather; a wrong answer about your prices, your hours, your terms, delivered in your name.
What "grounded" has to mean
The fix is not a cleverer model. It is a rule about where answers may come from.
Polycarp answers only from what your business has given it: your website's pages, the documents you upload, and the answers your team has approved. Every answer shows the page it came from, so the customer can check it and so can you. When none of that covers a question, the agent says so:
Our pages don't say, so I won't guess. I've asked the team, and their answer will appear here in this chat.
That line is not a fallback for when the model is unsure. It is what happens whenever the retrieved pages do not settle the question, decided before the model writes a word. The model is never the judge of whether it may answer.
"I don't know" is not a dead end
An old rule-based chatbot also says it doesn't understand, which is why people hate them. The difference is what happens next.
- The question goes to your team by email, with the conversation so far, translated if the customer wrote in another language.
- Your team's reply appears in the customer's chat, in their language. Nobody starts again.
- The reply is drafted into your knowledge base for you to approve. Approve it, and the agent answers that question itself next time, citing it as coming from your team.
So the agent gets better in exactly one way: your team answered something, and said yes. Nothing a visitor types ever becomes knowledge on its own.
The same rule for actions
The rule extends to doing things. When you let the agent act in your Shopify store or your own system, the model proposes the step, but your rules are checked in code and your amount is a hard line. Allow refunds up to €50 and it refunds €38; a €51 refund waits for a person. The agent never picks its own figure.
How to check this on any product
Before you put an AI chatbot on your website, ask it three things it cannot know from your pages: a price you don't publish, a policy you don't have, a product you don't sell. A good one says it doesn't know and offers a person, three times. Try it on any of our sample businesses: they are fictional, so you can ask anything.
Then ask the vendor where the answer came from. If the chat window cannot show you, you are trusting the model to be honest about its sources, which is the thing it is worst at.
What this costs you
Nothing extra. It is how every Polycarp plan works, from Starter at €29 a month. The first month is free and needs no card: give it your web address and it reads your site while you watch.