Enterprise knowledge search

It retrieves what the asker may see and answers with citations

Enterprise knowledge search connects the company's approved repositories. For each question it retrieves the relevant passages, applies the asker's existing access rights and returns results or a cited answer.

Success is a permitted answer supported by current material. The fit is strongest where recurring questions have documented answers.

Value comes from less searching and fewer repeated requests

Employees stop trying keywords across several tools and stop reconstructing old decisions from memory. Specialists get fewer routine requests once colleagues can find a cited answer first. Teams also repeat less analysis, because prior work becomes findable.

Measure these effects separately, because they overlap in the data: search time saved and an avoided expert interruption may describe the same episode, and adding both counts it twice. Faster onboarding is a further effect, and it needs its own causal evidence.

The primary KPI is time to a supported answer. The clock starts when the need arises and stops at an answer that is relevant, current enough, cited and accessible under policy.

Adoption tells you whether the search is used, not whether it is worth anything.

A freed hour is worth what it becomes, not what it costs

Start from the pool of freed hours:

Freed-hour pool =
  employees in scope
  × hours saved per employee per week
  × 52
  × active adoption

The pool describes scale, not value, and an hourly cost is not the right price for it. A freed hour is worth what the person does with it next, so classify the hour before you value it:

• An hour that removes real spend, such as a hire not made, overtime not paid or a vendor fee avoided, is worth its loaded cost. This route is the floor of the case.
• An hour that reaches a binding constraint where demand is waiting is worth the incremental contribution profit it creates there. That is normally more than the person costs, because a company only employs people whose output is worth more than their pay.
• An hour that dissolves into scattered convenience has no P&L value until it converts into one of the routes above.

Duplicate work avoided counts through the first route as rework hours that no longer happen, and onboarding cost avoided needs its own causal evidence before it is added.

A fictional ROI range, floor to base

These assumptions are illustrative only. They are not a benchmark, a forecast, a guarantee or a quote.

Freed-hour pool = 1,400 × 1.2 × 52 × 75% = 65,520 hours

Floor, cost route only =
  65,520 × 10% × 520 SEK = 3.41 MSEK

Base, cost route plus constraint route =
  3.41 + (65,520 × 5% × 1,100 SEK)
  = 3.41 + 3.60 = 7.01 MSEK

Year-one cost = 2.80 + 1.50 = 4.30 MSEK

Fictional net first-year value = -0.9 to +2.7 MSEK
Fictional payback = about 15 months on the floor,
  about 7.5 months in the base case

The floor claims only removed spend and defends itself on payroll and vendor data alone. The base case is where the real return sits, because a constraint-linked hour is worth about twice what it costs, and it is also the part that needs internal evidence: the named constraint, the waiting demand and the marginal contribution per hour.

Permissions travel with every record and are enforced twice

Connectors read approved content and metadata, and incremental indexing handles changes and deletions. Access-control lists, group memberships and document restrictions travel with each record. Authorisation is enforced during retrieval and again before return, across embeddings, caches, citations, conversation history and answer caches.

Source documents are untrusted data, never instructions, which limits indirect prompt injection. Authorisation and tool permissions stay outside the model. The model receives only permitted passages, cites material claims and declines when the evidence is thin. Apply output data-loss checks and provider retention controls where required.

Chunks keep their titles, timestamps, owners and source links, so an answer can always point back to where it came from. Keyword and semantic retrieval can be combined and reranked, and freshness rules expose dates and warn on expired material.

Build an evaluation set of real questions, including ambiguity, conflicts, missing answers and required refusals, and score retrieval, relevance, citation support and freshness against it. Maintain a permission matrix across roles, regions, teams and exceptional grants. Test direct queries, paraphrases and follow-ups against restricted material, and test revocation, freshness and deletion lag against agreed service windows.

Compare historical questions under both processes using one rubric. Monitor connector failures, indexing and access-list lag, unsupported claims and stale citations. Privacy-respecting logs should reconstruct the user, the query, the passages and the permissions behind an answer.

Start with one domain that has repeated questions and named owners

Choose a domain with credible documents, named owners and baseline tasks. Connect the necessary sources, normalise the permissions and release to a small group.

Set the thresholds before the pilot: supported-answer time, citation support, freshness, permission tests, revocation and deletion lag. Expand only after meeting them for an agreed period, and only when the next domain has reliable ownership and access rules of its own.

Do not build when:

• Most valuable answers depend on undocumented judgment or sensitive political context
• Source permissions are inconsistent or cannot be synchronised reliably
• Documents are mostly obsolete, duplicated or ownerless
• Search demand is too low to justify connectors and governance
• A configured feature in an existing platform meets the tested need at lower total cost
• Leadership expects distributed minutes saved to appear automatically as EBIT

In those cases, clean the sources first, redesign access, capture the expert knowledge, or test a narrower assistant.

Sources and methodology

The value model and the implementation recommendations are Epicube analysis, and the worked economics are illustrative. These independent references inform the retrieval, evaluation and permission controls:

• NIST TREC 2025 RAG Track documents evaluation work for retrieval-augmented generation
• OWASP: Vector and Embedding Weaknesses covers access control, data leakage and retrieval risks
• NIST AI Risk Management Framework provides a general framework for governing and measuring AI risks