Enterprise Knowledge Assistant
An internal assistant that answers questions from your own documents with citations, respects the permissions people already have, and says it does not know rather than inventing something plausible.
Every organisation past a certain size has the same complaint: the answer exists, somebody wrote it down, and nobody can find it. Enterprise search has promised to fix this for twenty years and mostly returned a list of documents. The difference now is that a system can read the documents and answer the question, which raises the stakes on getting permissions and citations right.
An enterprise knowledge assistant is an internal system that answers employees' questions directly from an organisation's own documents and systems, citing the sources it used, filtered by each person's existing access rights, and designed to refuse rather than speculate when the answer is not in the corpus.
What an enterprise knowledge assistant has to get right
- Permissions, enforced at query time. The assistant must never surface content the person asking could not open themselves, including in a summary. This is the requirement that stops most projects at security review.
- Citations that can be checked. Every claim links to the passage and document it came from, so a person can verify in one click rather than take it on faith.
- An honest refusal. When the corpus does not contain the answer, saying so is the correct behaviour. Systems that always produce something teach their users to distrust everything.
- Freshness. A superseded policy quoted as current is worse than no answer at all. Retirement and versioning have to be handled explicitly.
- Conflict handling. Where two documents disagree, the assistant should surface both with dates and sources rather than silently choosing one.
How we build it
Audit the corpus before promising anything
We sample the content and assess how much of it is current, authoritative and internally consistent. This audit routinely finds that a meaningful share of the material contradicts itself or was last reviewed years ago. That is not a reason to stop, but it does change the scope, and it is far better known in week one.
Connect the systems people actually use
Document stores, wikis, ticket systems, shared drives and intranets, each with its own permission model, formats and update patterns. Content is parsed with structure preserved, indexed for hybrid search and reranked, following the approach set out in RAG system development.
Enforce access at retrieval, not at display
Permissions are resolved against your identity provider and applied as a filter during retrieval, so restricted content never enters the model's context. Filtering after generation is not a control, because by then the content has already been read and summarised.
Answer with evidence, refuse without it
Answers are assembled from retrieved passages with inline citations, dates and document owners. Where retrieval returns nothing sufficiently relevant, the assistant says so and offers the nearest documents and the person to ask, which is far more useful than a confident paragraph.
Close the loop with the content owners
Unanswered questions are the most valuable output the assistant produces. They are a live map of the gaps in your documentation, and we route them to the people who can fill them, then measure whether coverage improves.
The assistant will expose your content problems
Within a fortnight of launch you will discover which policies contradict each other and which pages nobody has updated since a reorganisation. Teams sometimes experience this as the system failing. It is the system working: those contradictions were already costing people time, silently.
What to expect in the first quarter
| Stage | What typically happens | What to do about it |
|---|---|---|
| Weeks 1 to 2 live | High usage, mixed satisfaction, gaps surface immediately | Triage unanswered questions daily |
| Weeks 3 to 6 | Coverage improves as owners fix documents | Feed failures into the evaluation set |
| Weeks 6 to 10 | Usage concentrates on a predictable question set | Optimise retrieval for the real distribution |
| Quarter end | Support and internal query volume measurably lower | Report against the baseline captured at the start |
We capture the baseline before launch precisely so that final row can be evidenced rather than claimed.
How the engagement runs
Content reality is established first, because it determines what the assistant can honestly be expected to do.
Corpus and permissions audit
Content sampled for currency and consistency, source systems and their permission models mapped, and a baseline of current query volume captured.
Ingestion and retrieval
Connectors, structure-preserving parsing, hybrid indexing and reranking, scored against a question set drawn from real internal queries.
Assistant and access control
Grounded answering with citations and refusal behaviour, permission filtering enforced at retrieval and tested across roles.
Pilot
A department-sized pilot with daily triage of unanswered questions and routing of gaps to content owners.
Rollout and handover
Wider release, monitoring, and handover of the pipeline, evaluation set and the gap reporting process.
What you receive
An assistant your security team can approve and your content owners can improve.
Knowledge assistant
The deployed assistant with connectors, retrieval, grounded answering and citations, as infrastructure as code.
Permission enforcement design
How access is resolved and applied at retrieval, with the cross-role test results.
Corpus audit
What is current, what contradicts itself and what should be retired, by source system.
Evaluation set and scores
Real internal questions with agreed answers, scored, including refusal cases.
Gap reporting
Unanswered questions routed to content owners, with coverage tracked over time.
Usage baseline and impact
Query volume and resolution time before and after, measured against the baseline captured at the start.
Is this the right engagement?
Worth being direct. Enterprise Knowledge Assistant is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- People routinely cannot find answers that exist somewhere in your systems.
- Support or internal helpdesk volume is dominated by repeat questions with documented answers.
- Content sits across several systems with different permission models.
- Answers must cite their source and respect existing access rights.
- Someone owns the content and can act on the gaps the assistant exposes.
Choose something else if
- The documentation does not exist. An assistant cannot answer from nothing.
- The corpus is known to be contradictory and nobody is willing to reconcile it.
- You need a customer-facing assistant, which carries a different risk and tone profile.
- You only need the retrieval layer for another product. That is RAG system development.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
How is an enterprise knowledge assistant different from enterprise search?
Search returns a ranked list of documents and leaves the reading to you. A knowledge assistant reads the relevant passages and answers the question, with citations back to them. The retrieval underneath is similar; the difference is that the answer, and therefore the responsibility for being right, moves to the system.
How do you stop it showing people documents they should not see?
Permissions are resolved from your identity provider and applied as a filter during retrieval, so restricted content never reaches the model. We test this explicitly across roles, including the case where someone's access changes, because filtering after generation is not a control at all.
What happens when the answer is not in our documents?
It says so, and offers the nearest relevant documents and, where we can determine it, the person or team who owns that area. Those unanswered questions are then reported as a coverage gap, which is one of the more valuable things the system produces.
How current will the answers be?
As current as your content, with incremental reindexing so edits appear quickly and retired documents stop being quoted. Where a source is stale, the assistant shows the document date alongside the answer so a reader can judge, rather than presenting everything with equal authority.
How do we measure whether it is worth it?
We capture a baseline before launch, typically internal query volume, repeat question rate and time to resolution, then report the same measures a quarter later. Usage alone is a weak signal; the question is whether the queue in front of your subject experts got shorter.
Often paired with this
Most clients combine two or three engagements from the Generative AI & LLM Engineering pillar. These are the ones that most often run immediately before or after.
RAG System Development
Retrieval augmented generation built as a measured retrieval problem, with chunking, hybrid search, reranking and evaluation.
Read more →AI Copilot Development
A copilot inside the tool people already use, scoped to real workflows, with the user always in control.
Read more →LLM Evaluation and Benchmarking
Golden sets, calibrated judges and a regression suite in CI, so quality is a number that moves rather than an opinion.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.