Getting expert knowledge out of people's heads with AI interviews

Every company has a few people everyone goes to. The one who knows why that customer gets different payment terms. The one who can tell from a photo whether a claim is worth checking. They didn’t learn it from a manual, and it isn’t in one.

When one of them retires, changes jobs or is simply on holiday, you find out how much of the company ran through their head.

Why documentation projects stall

The usual answer is “let’s write it down”. Someone gets the task of documenting the process, there’s a template, a deadline, maybe a wiki. Months later you have a few long pages that describe the normal case well and skip the parts that matter.

That isn’t laziness. Experts are bad at listing what they know, because most of it doesn’t feel like knowledge to them. It feels like common sense. Ask “how do you handle returns?” and you get the five steps everyone already knows. The exceptions only come out when someone asks about a specific case.

Writing also costs the expert time they don’t have, for a reader they can’t picture. So the document stays half done.

What an interview does better

A good interviewer doesn’t ask for the process. They ask for cases. “Tell me about the last return that didn’t go the normal way.” “What would you do if the customer is a wholesaler?” “You said always, are there times you don’t?”

That’s how the exceptions surface: one concrete example at a time, with follow-up questions on whatever sounds like a rule.

Until recently that took a person with time and some skill, usually a consultant. Now an AI can run that kind of interview. It’s patient, it’s available when the expert is, and it keeps track of what it heard.

How an AI interview works in practice

We run them by voice, because talking is faster than typing and people explain more when they talk. A typical round looks like this:

1. Preparation. The AI gets what already exists on the topic: old wiki pages, a checklist, the manager’s short description of the area. It looks for gaps and contradictions before the conversation starts.
2. The conversation. Twenty to thirty minutes on one topic. The AI asks about the normal case, then the exceptions, then two real examples. When an answer contradicts a document, it asks which one is current.
3. Draft rules. From the conversation come short statements with a reason, like “Wholesale customers get 60 days payment terms because our contracts with them say so.”
4. Review. The expert reads the drafts and corrects them. Then the Owner of the topic approves them. Only approved rules count.

The expert spends half an hour talking and ten minutes reading. That’s a lot easier to get into a calendar than a documentation project.

What to watch out for

Interview the right person about the right thing. One topic per conversation, and the person who actually does the work, not just the one who manages it.

Keep the transcript out of the result. A transcript is raw material. Nobody reads two hours of conversation, and an AI that does will treat every half-sentence as a rule. What you want at the end are short rules someone has checked.

Ask before you record. People talk more freely when they know what happens with the recording and who reads the result. In the EU that’s also a data protection question: say what you store, where, and for how long.

Don’t stop at capture. Knowledge that isn’t used goes stale. The rules need to end up where people and AI actually work, and someone has to own them when things change.

A small first step

Pick the one person whose holiday everyone dreads. Pick one topic they handle. Book thirty minutes and ask about the last three cases that didn’t go the normal way. Write each answer down as a rule with a reason.

You’ll probably end up with more rules from that half hour than from the last documentation project.

In CtxCore Brain these interviews are built in: the AI prepares from what you already have, talks to your experts, turns the answers into draft rules and sends them to the right Owner for approval. Approved rules go to every AI tool you use. More on the idea behind it in What is a company brain?