Traditional enterprise search looks for words. People look for answers. Between the two there’s a daily, silent loss of time.
Silent because it doesn’t generate any tickets. Nobody opens a ticket to say they spent twenty minutes finding a procedure. They adapt: they ask a colleague, they reconstruct the information, or they give up and proceed based on what they remember.
The last option is the most expensive, and it’s also the one that leaves no trace.
The limits of keyword search
Traditional search compares text strings. It returns the documents that contain the typed terms, ranked by frequency and position.
It works when the person searching already knows the document’s exact vocabulary. It fails in three situations that recur constantly in technical and administrative contexts.
When the same concept has different names. An employee searches for “early termination,” the contract talks about “withdrawal.” A technician searches for “gasket replacement,” the manual titles the procedure “hydraulic seal maintenance.” No text match, even though the document is in the archive.
When the answer is spread out. The condition is in one document, the exception in another, the update in a later circular. The search returns three results and leaves it to the person searching to piece them together, assuming they even notice the third one exists.
When the question doesn’t contain the words of the answer. “Who authorizes a business trip longer than five days” might be answered in a policy that never uses the word “trip.”
On top of this there’s a practical effect. When a search fails two or three times, people stop using it. From that point on, knowledge goes back to circulating informally, and the document system becomes an archive for storage rather than a working tool.
Semantic search and hybrid search
Semantic search compares meanings instead of words. Documents and the query are represented in a numerical form that places similar content in nearby positions. A question about “early withdrawal” retrieves the paragraph on “contract termination” because the system recognizes the conceptual proximity.
It solves the three cases above. But it introduces one of its own: it’s less precise on literal references. When searching for an article number, a protocol number or a regulatory code, meaning-based search can return conceptually related results instead of the exact one.
Hybrid search combines both modes. It runs both searches and merges the results with defined ranking criteria.
A concrete example of the difference. An employee searches “which documents are needed for a variance to the approved project.” Text search returns every document containing “variance” and “project,” in the wrong order. Semantic search identifies the relevant procedural section even if it’s titled differently. Hybrid search does both, and if the question contains a specific regulatory reference, it doesn’t lose it.
In practice, hybrid is almost always the right choice in a corporate and administrative setting, where conversational language and coded references coexist.
The permissions issue
This is the point on which an enterprise search project’s chances of reaching production are decided, and it’s addressed too late with some regularity.
A search system crosses boundaries that repositories keep separate. A user who couldn’t open a folder must not be able to receive its content through a synthesized answer.
Three minimum requirements.
The filter applies at retrieval, not at presentation. Unauthorized documents must not enter the material the answer is built from. Hiding them downstream isn’t enough: the information would still have influenced the output.
Permissions follow the source and stay in sync. The system inherits rights from the source repository and updates them when they change. A static copy of permissions drifts out of alignment within weeks.
Queries are logged. Who searched for what, which documents were retrieved, what answer was produced. This is needed for audits and for diagnosing errors.
An operational warning: in many organizations, permissions in existing repositories are inconsistent, layered over years. The search project makes them visible. That’s a benefit, but it has to be planned for as an activity, because the review is the organization’s responsibility and takes time.
How to measure the benefit
Three measures, captured before launch.
Average time to retrieve a piece of information, sampled on real requests with a stopwatch in hand. It’s a manual measurement, but it’s the only one that produces a comparable figure.
Rate of searches with no useful outcome, meaning sessions that end without opening a document or that get rephrased multiple times. It can be extracted from the current system’s logs, if they exist.
Volume of internal information requests: how often a person asks a colleague or a department for information that’s already documented. It’s estimated through a sample survey over a limited period.
The number of searches performed, on its own, measures nothing. It can grow because the system is useful, or because no search succeeds on the first try.
What search doesn’t solve
It doesn’t create the missing knowledge. If a procedure isn’t written down, no engine will find it.
It doesn’t establish which version is current. If three revisions coexist in the archive with no indication, the system will be able to cite the source and cite the superseded one.
It doesn’t replace document structure. Semantic search reduces the dependency on perfect classification, it doesn’t remove it: reliable metadata on date, type and validity remain decisive.
And it doesn’t solve adoption problems. A better tool introduced without support gets tried for two weeks and then abandoned in favor of old habits.
In closing
Time lost searching for information is among the least visible and most widespread costs in organizations with large archives. It doesn’t show up in any indicator, it’s spread across every person, and it grows with the age of the archive.
The path we recommend starts from a bounded document corpus with permissions already under control, measuring retrieval time before and after.
If you have an archive that people struggle to consult, we’re available for a conversation about the case: which sources would be involved, how permissions are managed today, and what benefit would be realistic to measure.
