How to build a knowledge base an AI can actually answer from

Teams usually start a knowledge base the wrong way round: they plan a structure, assign sections to people, and three weeks later have an empty outline nobody has written into. The version that works starts from the questions you already answer twenty times a week.

Start from the last 100 conversations

Open your inbox, read the last hundred customer messages, and group them. Almost every team finds the same shape: five or six questions account for well over half the volume, and a long tail of everything else.

Write those five or six first. Not as a policy document — as the answer, in the words a manager would actually send.

If a document cannot be pasted into a chat and sent as-is, it is not a support document yet. It is a legal text with an answer buried in it.

What belongs in it

That fourth row matters more than it looks. A knowledge base full of what you offer and silent on what you refuse will produce confident answers about services you do not provide.

How to phrase an entry so it can be found

Retrieval works by matching a customer's question to a passage. The phrasing of the passage decides whether it is ever found.

• Use the customer's words, not the internal ones. People write "коли приїде" and "нова пошта", not "expected transit interval". Include both if you must, but lead with theirs.
• One entry, one question. A 4,000-word page covering delivery, returns and payment will be retrieved for everything and answer nothing precisely.
• Put the answer in the first sentence. Context after, conditions after that.
• Write the conditions explicitly. "Returns accepted within 30 days" and "after 14 days the refund is store credit" are two facts; leave one out and the AI will smooth over the gap.
• Date anything seasonal. Prices, promotions and holiday schedules need a validity period written into the text, or they outlive themselves.

What to keep out

• Drafts and internal debate. "We are thinking about changing this" indexed next to the policy produces answers about a policy that does not exist.
• Screenshots as the only source. An image is invisible to retrieval. If a rule lives only in a picture, type the rule out.
• Two versions of the same rule. The most common cause of a wrong AI answer is not the model — it is an old policy that was never deleted.
• Anything with personal data. Customer names, phone numbers and order contents belong in the CRM, not in indexed documents.

Let the gaps write the next entries

The fastest-growing knowledge bases are not written in planning sessions. They grow from the questions that had no answer.

A working loop looks like this:

1. A question arrives that the knowledge base does not cover.
2. The system records the gap and routes the question to a person.
3. The manager answers it properly, as they always would.
4. That approved answer is saved as an entry, with an owner.
5. The next customer who asks gets it immediately.

After a month of this you have a knowledge base shaped by demand rather than by imagination — and a measurable share of questions that no longer need a person at all.

Keeping it honest

Review the entries with the highest usage once a month: they are the ones where a stale sentence does the most damage. Anything that has not been used in three months is either obsolete or badly phrased — check which before deleting it.

Two numbers tell you whether the base is healthy: the share of answers that carried a source, and the share of drafts a manager sent without editing. Both are in the metrics worth counting, and both move quickly once the first five entries are real. What happens after that is described in AI answers with sources.