Internal Knowledge in Small Firms: When Only One Person Knows
In many firms knowledge sits with individuals. What an internal AI search realistically solves — and where these projects reliably come apart.
Almost every smaller business has that one person you ask. Which supplier has which terms, how the warranty process runs, why this particular customer is never invoiced up front. It is written down nowhere — or it is in a file nobody can find.
As long as that person is around, it works remarkably well. It is still the largest uninsured risk in the business.
What an internal search over your own documents does
The technical answer today is usually: a search that looks at content rather than filenames, and that answers questions in plain language — with a reference to the source.
It realistically works well for:
- Contracts and quotes — "What notice period do we have with supplier X?"
- Process descriptions and work instructions — "How do complaints work when a third party did the installation?"
- Product and technical documentation — data sheets, installation guides, approvals
- Recurring internal questions — onboarding, holiday rules, billing details
The benefit comes less from the complicated question than from the twentieth simple one. That one costs not only the asker's time but pulls the person being asked out of their own work.
Where these projects come apart
At the source, not the technology. If the same work instruction exists in three versions on three drives, a search will reliably serve the wrong one. That is not a model problem — it is a filing problem, made more visible than ever before.
At missing permissions. Salary data, personnel files, costings: if an internal search covers everything on the drive, it will answer questions it is not allowed to answer. Mapping permissions is the most laborious part of the project and the most frequently underestimated.
At missing citations. An answer without a source is worthless in a business context because it cannot be checked. A reference to the document and the passage is not a comfort feature; it is the precondition for anybody relying on it.
At expectations. A system like this answers questions about existing documents. It does not know what was never written down — and in smaller firms that is often the larger part.
The underrated side effect
The cleanup is frequently worth more than the tool. Deciding, for the sake of an internal search, which document is the valid version, where it lives and who may see it solves a problem that was costing money every day anyway — with no AI involved at all.
So the honest order is: first a clearly bounded document set, then search over it. Not the other way around.
Frequently asked questions
Isn't the search built into our file system enough?
For filenames and exact terms, often yes. The difference shows on questions phrased differently from the document — a full-text search for "withdrawal" finds nothing if the contract says "cancellation".
Do our documents go to an external provider?
That depends on the chosen setup and it is the single most important question. Processing location, the processing agreement, and whether content is used for training belong settled in writing before the first document is uploaded.
How small can a business be for this to pay off?
It is less a question of headcount than of repetition. If the same questions reach the same person several times a week, there is a case. If not, a well-maintained document is often the better answer.
What if the answer is wrong?
Then it has to be checkable — hence the citation. For binding statements to customers, human review stays, exactly as with quotes.
What you can do today
- For one week, note the questions that go internally to the same person.
- Mark the ones whose answer sits in a document — that is the addressable set.
- For those documents, settle: where is the valid version, and who may see it?
- Write the three most common answers down properly, once. That often handles a large share on its own.
- Only then consider a tool — and start with a clearly bounded set.
If what comes out is a document set clean enough to work with: how we build internal tools like this is described under custom solutions and AI & automation — or in a first call.