EU AI Act high risk obligations are now enforceable. Check your exposure
Insights About us Careers
Contact us
Generative AI & LLM Engineering

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.

8 to 12 weeks
Typical build
Fixed scope
Commercial model
Permission aware
Retrieval

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.

In one paragraph

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 A superseded policy quoted as current is worse than no answer at all.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.

Worth knowing

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

StageWhat typically happensWhat to do about it
Weeks 1 to 2 liveHigh usage, mixed satisfaction, gaps surface immediatelyTriage unanswered questions daily
Weeks 3 to 6Coverage improves as owners fix documentsFeed failures into the evaluation set
Weeks 6 to 10Usage concentrates on a predictable question setOptimise retrieval for the real distribution
Quarter endSupport and internal query volume measurably lowerReport against the baseline captured at the start

We capture the baseline before launch precisely so that final row can be evidenced rather than claimed.

Process

How the engagement runs

Content reality is established first, because it determines what the assistant can honestly be expected to do.

Weeks 1 to 2

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.

Weeks 3 to 5

Ingestion and retrieval

Connectors, structure-preserving parsing, hybrid indexing and reranking, scored against a question set drawn from real internal queries.

Weeks 6 to 8

Assistant and access control

Grounded answering with citations and refusal behaviour, permission filtering enforced at retrieval and tested across roles.

Weeks 9 to 10

Pilot

A department-sized pilot with daily triage of unanswered questions and routing of gaps to content owners.

Weeks 11 to 12

Rollout and handover

Wider release, monitoring, and handover of the pipeline, evaluation set and the gap reporting process.

Deliverables

What you receive

An assistant your security team can approve and your content owners can improve.

01

Knowledge assistant

The deployed assistant with connectors, retrieval, grounded answering and citations, as infrastructure as code.

02

Permission enforcement design

How access is resolved and applied at retrieval, with the cross-role test results.

03

Corpus audit

What is current, what contradicts itself and what should be retired, by source system.

04

Evaluation set and scores

Real internal questions with agreed answers, scored, including refusal cases.

05

Gap reporting

Unanswered questions routed to content owners, with coverage tracked over time.

06

Usage baseline and impact

Query volume and resolution time before and after, measured against the baseline captured at the start.

Fit check

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.
Questions

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.

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.