The most expensive search engine in your company is the person everyone interrupts

TL;DR — Most of what a team knows isn’t searchable. It’s scattered across a drive, a wiki and three people’s heads, and the fastest way to find any of it is to interrupt someone. What that costs, and what retrieval-based search with citations changes.

Thursday, 14:10

A new hire at a 22-person B2B services firm needs one fact: the standard payment terms for one client type. It has been answered in writing, twice — once in a wiki page last updated in March, once in a resolved ticket from June. Neither surfaces: drive search matches filenames, not meaning, and the wiki holds three different pages about terms.

So she does what everyone does: asks. Nineteen minutes later a senior colleague stops what he is doing, digs out the ticket, pastes the link. Her question cost her 20 minutes and him six — and that is a good outcome, because he knew where it was. Nobody logs it, so the next hire will ask it again, in the same channel.

That is one question. Sample a month of the ops channel and the ticket queue and the pattern holds: about 45 internal questions a week, two-thirds already answered in writing, somewhere.

Where the knowledge actually lives

Count it for this firm:

1. The scattered sources. A drive with about 4,000 files. A wiki a third of which hasn’t been touched in a year. 11,000 resolved tickets — the company’s real operating memory, every edge case anyone has hit. Two shared inboxes. None of them search each other; most barely search themselves.
2. The repeat questions. About 45 a week; two-thirds already answered in writing. At ten minutes combined per question — asking, waiting, context switching — roughly 30 hours a month go to re-deriving what the company already knows.
3. The single point of failure. Client specifics, contract quirks, promises made verbally: the load-bearing knowledge lives in two people’s heads. They get interrupted fifteen times a day. When one is on holiday, answers wait a week — and so does every decision behind them.
4. Onboarding by interruption. A new hire asks 200-plus questions in their first three months, most already answered somewhere. Three months to stop asking. A rite of passage for the hire, a tax on everyone else.

That is the shape of a job retrieval should own.

What retrieval-based search changes

Not a chatbot with opinions about your business. A search over the sources you already have, answering in plain language and showing its work:

• One index over your existing sources — drive, wiki, tickets, inbox — kept in sync. Nothing is migrated or rewritten; it reads what is already there.
• The receipt attached. Every answer carries a link to the exact document and section it was built from. Click through and check. If it can’t cite it, it doesn’t say it.
• “I don’t know” as a feature. It answers only from what it can trace to a source; when nothing supports an answer, it says so instead of producing a fluent guess. One invented clause cited to nobody, and the team stops asking it anything.
• Permissions respected. People get answers built only from what they may already read. HR folders stay invisible to the team.
• Your data stays yours. Documents stay inside infrastructure you control — not a third-party index someone else operates, and nothing is used to train a model.
• Measured before anyone relies on it. You supply 50 real questions from last month; the build ends with an evaluation pass — how many answered with a correct citation, how many correctly declined. That number exists before the team sees the tool.

The numbers, before and after

• Time to an answer: ~20 minutes waiting for a person → seconds, including 11pm.
• Questions answered without interrupting anyone: near zero → two-thirds.
• The experts’ interruptions: fifteen a day → the few that need judgement.

The arithmetic: ~30 hours a month of question-answering → about 10, mostly the genuinely new questions. At a conservative blended rate, roughly $11,000 a year — much of it senior time.

Then the number nobody can count afterwards: the questions never asked because searching wasn’t worth twenty minutes — so people guessed. You find out about those from the client.

How it starts

You hand over two things: read access to one or two sources, and a sample of real questions from last month. A first version over those takes two to three weeks and ends with the evaluation pass — the tool’s scorecard on your questions, not someone else’s demo data. If the number convinces, add sources; each next one is faster.

It does not replace the person. The questions that need judgement — which client will flex on terms, which promise is safe to make — still go to whoever has been there six years. The difference: the thirty questions a week with written answers stop landing on their calendar.

The test is simple: if the fastest way to find something in your company is to ask a specific person, your search index is their memory. It takes holidays, it changes jobs eventually, and it has never been backed up.

---

If the fastest way to get an answer on your team is interrupting the person who has been there longest, that is the thing to fix first. We build retrieval-based search over your own documents, tickets and conversations — plain-language answers, citations to the exact source, and “I don’t know” when it doesn’t know: Knowledge base & search.