Skip to content
Aplora
Consulting and professional services
AI Workflow Automation

Client under NDA

Answers stopped living in people's heads: search across company documents, with the source attached

A B2B IT integrator with more than 50 staff. Years of work had accumulated standards, instructions, solution write-ups and project reviews — across different folders, systems and formats.

Published: 2026-08-24 · updated: 2026-08-24

A row of labelled binders on an archive shelf
Illustrative photo of the working context, not a screenshot of the client's system

  • Time spent finding information

    −40–60%

  • Errors caused by out-of-date instructions

    −20–35%

In short

  • Problem: knowledge had accumulated but was scattered across systems; searching took time, and what surfaced could be stale.
  • Solution: a single document index, search on the meaning of the question, an answer returned as a fragment with its source, and access from inside the CRM and helpdesk.
  • Result: 40–60% less time spent searching, 20–35% fewer errors from stale instructions — per the client's figures.
  • Systems: document storage, CRM, helpdesk.

Context

  • A B2B IT integrator; name under NDA.
  • Size: more than 50 staff.
  • Process volume: standards, instructions, solution write-ups and project reviews accumulated over years.
  • Systems: document storage, CRM, helpdesk; some under NDA.
  • The constraint became visible as the team grew: a new hire took months to get up to speed, and the people who held the knowledge became the bottleneck.

Baseline

Baseline metrics with their sources. Without them, any later result has nothing to be compared against.
  • Before the work started we fixed: typical time to find an answer across several working scenarios, and the number of incidents traced to a stale instruction.
  • Data source: the client's incident records and a scenario-based measurement.
  • The assessment also drew on team feedback — that is not a measurement, and the numbers reflect it as a range rather than a point value.
  • There are no measurement dates or formula behind the numbers, so they are published as reported results.

Diagnosis

Which hypotheses were considered, why this one was chosen, what was assumed, and the condition under which we would have stopped.
  • Three hypotheses were on the table: build a corporate portal, rewrite the documents, or learn to search the ones that exist.
  • The third was chosen: the documents are mostly correct and the problem is access. Rewriting the archive would take months and go stale along the way.
  • The assumption: the answer to a working question exists in the documents. Tested against a sample of real questions from correspondence — it held for most, but not all: some knowledge genuinely is written down nowhere.
  • Stop criterion: if search returns the right fragment for fewer than half of historical questions, the work stops and the conversation moves to content rather than search.

What we implemented

  • Data sources: internal company documents only. External sources never enter an answer — otherwise nobody is accountable for the fragment.
  • Preparation: collecting, classifying, cleaning and normalising — the longest part of the work and the least visible in the result.
  • AI components: search on the meaning of the query and assembly of an answer from the fragments found.
  • Business rules: every answer must carry a link to its source. That is the defence against invention: a statement with no source does not reach the screen.
  • Integrations: CRM, helpdesk — access where the question arises.
  • Human checkpoints: the employee sees the source and decides. The system answers a question, not for the client work.
  • Access rights: scoped by role — not every document is visible to everyone.
  • Time to first effect was 2–4 weeks: that is how long collecting and structuring took before search went live.

How the process changed

Before

5 steps
  1. The question comes up in client work
  2. A search across folders and systems
  3. What turns up is checked for currency by hand
  4. If nothing turns up — ask a colleague
  5. The answer to the client is written from memory

After

5 steps
  1. The question is asked inside the CRM or helpdesk
  2. Search runs against the single index
  3. The answer comes back as a fragment with its source
  4. The employee sees the source and its date
  5. Nothing found — the question enters the queue to extend the base

What was stuck

Finding an answer took time, and what turned up could be out of date. Different people answered the client differently, and the surest route was to ask a colleague — that is, to interrupt one more person.

  1. 1The question is asked inside the CRM or helpdesk
  2. 2Search runs against the single index
  3. 3The answer comes back as a fragment with its source
  4. 4The employee sees the source and its date
  5. 5Nothing found — the question enters the queue to extend the base
  6. Measured result

Results

Time spent finding information

−40–60%

Reported by: Client figures: before-and-after process comparison plus team feedback

Errors caused by out-of-date instructions

−20–35%

Reported by: Client figures: their own incident records

Economic impact

  • The effect is freed time, twice over: the person asking does not search, and the person who used to be asked is not interrupted.
  • Fewer errors means losses prevented — the least reliable kind of effect, since it compares against an event that did not happen. No money figure is published for it.
  • Cost of ownership: the base needs feeding. With no update process, six months in the search starts answering confidently with stale content.

Adoption

  • Search lives in the systems people already work in: a separate portal would need a separate habit.
  • The base grows from the questions that found no answer — the queue builds itself.
  • New hires get up to speed faster because they no longer depend on whether the knowledge-holder is free.
  • The solution owner is the head of service.
«Answers used to live in people's heads. Now anyone finds what they need in a minute — with a link to the document, not “I think that's how it works”.»

Client under NDA — Head of service, B2B services (NDA)

What's next

  • Scaling: adding per-client project documentation, with stricter access scoping.
  • Next initiative: drafting the client-facing answer from what was found — with mandatory review.
  • What we decided against: showing an answer with no source. A statement without a link looks more confident, but it cannot be checked, and the first error destroys trust in the whole base.

Have a similar workflow? Let's check whether the hypothesis transfers

Another company's result is not a promise. It does show where to look.

All cases