AI on a website should answer from your material, not from memory.
A language model will always come up with some answer. The difference between a useful chat and a dangerous one is where that answer comes from.
Put a chat built on a language model on a company website, let it answer from memory, and you get a tool that never says “I don’t know”. It always answers something — fluently, confidently, and occasionally nowhere near the truth. For a company selling cupcakes that is embarrassing. For a company selling services with prices, deadlines and terms, it is a problem.
A model is not a database
A language model cannot look a fact up. It can continue a sentence with whatever most plausibly comes next — and it does that even when it does not know the answer. So it invents a price that looks like your prices, opening hours that look like ordinary opening hours, and an e-mail address that looks like yours.
You cannot fix this by writing “do not make things up” into the instructions. You fix it only by handing the model the answer.
How I build it
An answer is made in two steps:
- Find what the company’s own material says about this. The material is the price list, the service descriptions, the FAQ, the terms — things somebody in the company wrote and approved.
- Answer only from what was found. If nothing was found, say you will not answer that one and hand the question to a person.
The second step is the important half. A chat that can say “a colleague had better confirm this for you, I am sending him your question” is usable. A chat that has an answer for everything is a liability.
What that means in practice
AI does not handle exceptions. Contracts, discounts, complaints — anything where a judgement is made. Those belong to a person, and the chat has to recognise them and pass them on. Not because it could not technically answer, but because a wrong price is paid by you, not by the model.
The material is work nobody will do for you. If the price list exists in three versions and two of them are out of date, even flawless technology gives you two out-of-date answers. In practice, preparing the material takes longer than deploying the chat itself.
You have to be able to see what it answered. Without a record of questions and answers you will not notice it is answering badly — you will hear it from a customer, usually too late.
The chat is not the goal. The goal is that the enquiry reaches the right person with what is already known. The best outcome is often not an answer at all, but an enquiry that arrived complete and went to the right colleague.
When it is not worth it
If you get ten enquiries a month, do not deploy AI — you will handle them yourself, cheaper. It starts to make sense when the same questions keep repeating, when they arrive outside working hours, or when answering them means assembling facts scattered across the company.
Want a website that answers from your own material and passes the enquiry on? See how it works.