Company knowledge should be useful to the people allowed to use it.
An internal knowledge assistant should not create a second truth. It should help employees find relevant, current information with the same permission boundaries that govern the underlying source systems.
Can this employee verify this answer?
Answer only from current, authorized evidence, preserve its lineage and expose a no-answer state when those conditions are not met.
An answer needs a chain of custody.
The employee’s question
Who must review this supplier exception?
Supplier policy · current revision
Authorized + currentEligible evidence; citation can open for this user
Supplier policy · superseded revision
Authorized + supersededExclude from the answer; preserve version lineage
Restricted commercial terms
Outside this user’s accessExclude before model context is assembled
A bounded answer
The current accessible policy requires an authorized reviewer. If the exception cannot be resolved from that policy, return the question to its owner rather than infer an approval rule.
Keep an accessible source citation and its version with the answer.Now change the evidence.
- Group membership is removed
- Previously accessible content stops being answer evidence
- Two current sources conflict
- Expose the conflict and request an authoritative resolution
- No eligible source answers the question
- Say that the evidence is insufficient
Do not rebuild commodity search by default.
Native enterprise search may be sufficient. Custom retrieval is justified by a specific permission model, research workflow, data type, action integration or experience requirement.
Start with one domain and two permission groups.
- A realistic question set with retrieval judgments
- Citations that the requesting user can actually open
- Revocation, freshness and conflicting-source tests
The complete knowledge systems argument
For CIO, VP IT, Head of Knowledge, CISO
The situation behind the request
Teams work across SharePoint, policy manuals, product documentation, tickets and databases. Search results are hard to reconcile because documents have different owners, versions and expiration patterns. People may spend longer checking an answer than finding the information themselves.
Where the obvious solution falls short
A vector index on its own does not solve versioning, entitlement, source attribution or trust. More retrieval often makes an answer more confusing. The hardest design decision is what to do when trusted sources disagree or an employee asks a question their role should not answer.
How we would approach the work
We map authoritative sources, document freshness, identity, indexing cadence and expected user questions. We benchmark retrieval methods against a real evaluation set and design an interface that makes supporting evidence visible. The assistant needs an honest no-answer state.
What the engagement may include
- Connectors and indexing plan with source ownership and refresh strategy.
- Document and tenant permission checks at retrieval and application layers.
- Source-grounded answer interface, confidence/limitations and fallback workflows.
- Search/retrieval evaluation, error monitoring and operational runbook.
The choices worth making explicitly
Native Copilot or enterprise search may be sufficient. Custom engineering is justified by unusual permission models, specialized research workflows, action integration, data types or user experience. Do not rebuild commodity search simply because it is possible.
What we would test before calling it ready
Test confidential material against a lower-privileged account, revoked access, duplicate documents, stale policies, contradictory sources and questions with no answer. Source citations should point to information the requesting user is actually authorized to open.
A sensible first step
Start with one knowledge domain, a realistic set of questions and two permission groups. If retrieval cannot be trusted there, scaling it across the company will not solve the problem.