Unanswered questions, or what an AI assistant really learns from

Knowledge ages in silence

Anyone who has ever run a company knowledge base knows how it goes. At the start a few dozen carefully written pages appear, the team is pleased, and six months later half of them describe an offer that no longer exists. Nobody decides to neglect the base. Nobody simply knows what is missing from it, so nobody fills anything in.

An AI assistant does not solve this problem on its own, and it can even hide it. A language model asked about something that is not in its material may give a vague answer that sounds reasonable, and the customer leaves feeling they got a reply. The store learns nothing. That is why the first question when choosing an assistant is not how nicely it answers, but what it does when it has nothing to answer from.

A gap is information, not a failure

In our assistant every question it had no material for lands on a list of gaps in the panel. This simple assumption changes how you work with knowledge, because instead of guessing what is worth describing, you look at what customers actually asked and did not get.

The list groups questions by topic rather than by exact wording. Five different ways of asking about returns share one row with a repetition counter instead of scattering into five separate entries. Next to the counter you see that topic’s share of total traffic, so it is clear at once which gap costs the most conversations. The order of work stops being a matter of intuition: you start from the top of the list.

The conversation history completes the picture. Every session is recorded with dates, a message count and a search box, so you can check not only what people ask about but also what the answer that did not satisfy them looked like. You see what customers really ask about, rather than what we assume they should.

One article instead of five answers

Closing a gap does not mean adding an answer to a single question. You select the variants of the same problem and the assistant drafts the content from the knowledge you already have in the base. You edit it your way and save it as a knowledge page, which from then on answers the whole group of questions.

As the base grows, duplicates become the biggest risk, because two pages on the same topic eventually start to contradict each other. That is why saving stops when a similar article already exists, and the form comes back with a list of pages with related content. If the new page should be created anyway, you confirm it deliberately.

There is also a second route, useful when a question concerns a specific item. The question is linked to a product together with the words customers use to look for it. The assistant then finds the right product even when a customer calls it something other than the catalogue description does.

Your article comes first

A page written by your team enters the base as the company’s own material and is treated differently from an uploaded vendor document. When the match is comparable the assistant reaches for it first, because it describes your rules rather than general advice from a manual.

One distinction matters a lot here. The assistant takes prices and availability from live store data, not from the text of the article. A page can therefore describe shipping or return rules, and a price change in the catalogue does not require editing any text in the knowledge base.

A review that reports itself

Writing a page is half the work, because the offer changes and the text does not. Every knowledge page watches the part of your offer it describes and reports itself for review when that material changes or when too much time has passed since the last check. Review stops being a yearly clean-up of the whole base and becomes a small task done exactly when it is needed.

Every change to the content is saved in the version history along with the date and the author. If an edit turns out to be a mistake, you restore the previous version with a single button, so editing the base does not call for the kind of caution that paralyses a team.

Who sees which knowledge

The learning loop also covers internal knowledge. Next to the customer chat runs a separate assistant for employees, isolated from the public one. It answers from internal knowledge, but only within the permissions of the person asking. For every document you decide separately whether a customer on the site can see it or only your team, and which employee groups are entitled to it.

This matters in practice when closing gaps. A complaints procedure for the warehouse and a short note on returns for customers can grow out of the same set of questions, yet reach two different audiences.

Where to start

Do not start by writing a knowledge base in advance. Upload what you already have, such as the terms of sale, shipping rules, product manuals and replies to the most common emails, and let the assistant work for a few weeks. After that, the list of gaps will tell you more about your customers’ needs than any workshop, and the articles written from it will answer questions that are actually being asked.

What this work looks like from the inside is shown in the screenshots of the assistant’s panel, and the full description of the solution, including the choice between the cloud and your own servers, is on the AI assistant page. If you want to see it on your own data, get in touch.