What happens when it cannot answer something?
It says so, offers a callback, and sends you the enquiry with the whole conversation attached. Then that question goes on the list we write an answer for, so the next person who asks it gets a real reply.
Saying "I do not know" is a feature, not a failure mode. It is the behaviour that makes every other answer trustworthy, because an assistant that never admits ignorance is one you cannot believe when it sounds confident.
In practice the visitor gets a plain admission and an offer: leave a name and a number and somebody will come back to them. If they do, you get an email with their details, what they were asking about, and the conversation that led up to it — so you are not phoning a stranger cold.
The part that compounds is what happens afterwards. Every week we read the questions that got that response. The recurring ones become written answers, and the assistant stops needing to hand them back. A month in, it is answering things it could not answer on day one.
That loop is the actual product. The software is the delivery mechanism.