Who do I ask?" — how a corporate expert finder works

There is a class of question corporate search can't answer: questions whose answer isn't in a document but in someone's head. "How is the VAT exemption handled on export invoices to Germany?" is probably written down nowhere — but somebody in the company knows.

An expert finder exists to find that person.

Why isn't people search enough?

A classic employee directory works on literal matching: it returns a result if your word appears in a name, a job title or a department. That is a narrow form of search.

Someone typing "accounting" won't find the person whose title is "chartered accountant". Someone typing "export" won't reach the person whose profile says "foreign trade operations". When the words don't overlap the search fails — even though the meaning does.

Semantic matching

The solution is to compare the question and people's profile text in the same meaning space. Technically that means turning texts into vectors and finding the nearest ones; practically it means this: if the words don't match but the topic does, the person is found.

That has a precondition: there has to be text on people's profiles to match against. A directory containing only names and titles gives weak results with even the best algorithm.

How is skills data collected?

This is the hard part, and it isn't solved by a one-off campaign. Three approaches work:

Ask during onboarding. Asking a new joiner "what can you help with?" in their first week is easier than chasing them later.

Route it through managers. Having managers review their team's areas of expertise fills gaps and keeps the data realistic.

Feed it from usage. If someone repeatedly receives questions on the same topic, that's a signal of expertise.

Profile completeness should be tracked regularly; the quality of this module depends directly on it. We cover the directory side in a separate article.

What belongs on the results page?

Finding the right person isn't enough; the user has to be able to decide. A result card that works carries four pieces of information:

1. Who and where — name, title, department.
2. Why this person — matching skills and profile text.
3. Available right now? — in the office, on leave, in a meeting.
4. How to reach them — email or extension.

The third matters more than expected: finding an expert who is on leave delays the answer by two weeks. If availability is visible, the user tries the second name.

Privacy and dosage

An expert finder turns people into search results, which creates a sensitivity. Two boundaries matter.

No popularity ranking. The amount of recognition someone has received can be displayed, but it shouldn't determine the ordering; presenting a much-recognised person as the answer to a topic they don't know misleads the user and puts the person in an awkward position.

No personal data leakage. Matching must not rely on someone's leave records, performance rating or private data — only on the skills and biography text they wrote themselves.

The relationship with the AI assistant

The two answer different questions and complement each other. The assistant finds what is written in a document; the expert finder finds the person who knows what nobody wrote down.

A practical division: if the answer is in a document the assistant gives it, otherwise the expert finder points to a person. The questions the assistant can't answer are also a signal — that topic should be written down. We cover the assistant's scope in a separate article.

Where does the question go?

There is one thing an expert finder deliberately doesn't do: send the question. The user finds the person and sees how to reach them, but the channel is email or the phone.

That should be seen as a deliberate boundary rather than a gap. Handing over cleanly to existing channels produces less friction than reinventing internal messaging.

Measurement

• Search volume and no-result rate — queries returning nothing show the gap in profile data.
• Profile completeness — the share of employees with a filled skills field.
• Click depth — do users go to the first result, or scroll down the list?

The first indicator is the most valuable: the list of unsuccessful searches tells you directly whose profile needs updating.