An assistant that only answers from what you actually wrote.
The one on our homepage is the same build we ship to clients. Here is how it works, what stops it inventing answers, and what you can see in the log the next morning.
See your website with an AI on it.
Type your address. Twenty seconds later it is on your site, answering your customers in your name.
Most business chatbots fail in one of two directions. Either they are a decision tree wearing a chat window, four buttons and a dead end. Or they are a raw language model that will confidently tell a customer you offer something you have never offered.
The version worth having sits between the two. It answers in natural language, but every answer is grounded in a specific piece of content you approved, and it is built to say 'I do not know, here is a human' rather than guess.
How it is built
Your content becomes the source
Service pages, FAQs, policies, past quotes, product specs: whatever you already have. It is split into passages and indexed. Nothing else is in scope.
Retrieval before generation
A question first finds the passages that could answer it. Only those passages are handed to the model, along with an instruction to answer from them and nothing else.
A refusal path that gets used
When retrieval comes back weak, the assistant says so and offers the enquiry form or a phone number. Refusing is a normal outcome, not a failure state.
Handoff carries context
When it hands over, the conversation goes with it. The customer does not retype what they just spent four messages explaining.
It runs on your site, in your name
Not a third-party widget with someone else's branding in the corner and your data on their servers under their terms.
What stops it making things up
- It can only quote from indexed passages. There is no general-knowledge fallback.
- Low retrieval confidence triggers a handoff instead of a guess
- Prices, availability and commitments are treated as human territory by default
- Answers can cite which page they came from, so you can check one in ten seconds
- Change the page and the answer changes. There is no second copy to keep in sync.
What you see the next morning
Every question, in the customer's own words
Not categories. The actual sentences people typed, which is the most honest customer research you will get for free.
The list of things it could not answer
This is the useful one. Each item is either a page you should write or a service you should stop being vague about.
Where conversations turned into enquiries
And where they stopped one step before. That step is usually a missing sentence, not a missing feature.
What it is not
- Not a replacement for your people. It is the thing that stops them answering the same question for the ninth time today
- Not a sales agent. It does not negotiate, discount or promise
- Not a project that ends. The unanswered-question list is a monthly job, and a small one
Questions we get asked
What happens when it does not know?
It says so and offers a person: form, email or phone, whichever you prefer. The question lands on your unanswered list so the gap gets closed properly.
Where does our content go?
Into an index we run for you. What gets included is your decision, and it can be removed. We will tell you plainly which model provider processes the text.
Can it read our prices or stock?
Technically yes, by connecting it to the system that holds them. Whether it should answer with them is a separate decision, and for most businesses the answer is no.
What if it says something wrong?
Answers trace back to a source passage, so wrong output means a wrong or missing page, fixable at the source. Anything where being wrong is expensive stays with a person.
Tell us what is not working.
A 30-minute call, then a one-page diagnosis in writing within two working days: what we would do first, and whether you need us at all.