Skip to main content
Agentic Online

Illustrative solution pattern

Grounded knowledge assistant

An assistant that answers staff questions from your approved documents, cites its sources, and declines when the corpus does not cover the question.

Discuss a solution like this
Illustrative solution pattern. This describes a solution we design and build, not a completed engagement. The expected outcomes below are design intent, not measured results, and we have not yet delivered this pattern for a client.
Agentic AIGovernment & public sectorProfessional servicesEducation

The challenge

The problem this solves

Staff spend significant time locating information that already exists somewhere in the organisation, and the answer they get depends on which colleague they ask. General-purpose chatbots make this worse: they answer fluently from training data that has nothing to do with your policies.

Existing process

The limitation being removed

Information sits across shared drives, intranet pages, email threads and document management systems. Finding the current version requires knowing where to look and who to trust, so answers vary between people and over time.

The solution

What was built

An assistant grounded strictly in an approved document corpus. Every answer cites the documents it came from so the user can verify before acting, retrieval respects the user's existing permissions so nobody sees material they are not entitled to, and the assistant says it does not know rather than inferring when the corpus does not contain the answer.

Implementation approach

  1. Agree the corpus: which documents are authoritative, who owns them, and how updates reach the index
  2. Build permission-aware ingestion so retrieval can never cross an access boundary
  3. Implement hybrid retrieval and tune it against a curated question set drawn from real enquiries
  4. Constrain generation to retrieved context, with an explicit refusal path when evidence is insufficient
  5. Measure answer accuracy, citation correctness and refusal behaviour before release, and re-measure on every change
  6. Pilot with one team, review the logged questions, then extend the corpus based on what people actually asked

Technologies used

  • Claude or an equivalent model, selected against your data-handling requirements
  • Hybrid retrieval combining vector similarity and keyword search
  • Permission-aware document indexing tied to your identity provider
  • Citation rendering linked to the source document and section
  • Evaluation harness with a curated question set and refusal testing

Services provided

Applicable sectors

  • Government & public sector
  • Professional services
  • Education
  • Enterprise

Security and governance

Controls designed into the solution

Decided before implementation. Every one of these is an architectural choice, which is why they cannot be added afterwards without a rebuild.

  • Retrieval filtered by the user's own permissions, enforced at the index rather than in the prompt
  • Answers constrained to retrieved content; unsupported claims are refused rather than generated
  • Every response carries citations so the user can verify the source before relying on it
  • Questions and answers logged for review, with retention agreed in advance
  • Prompt-injection containment on ingested content, which is treated as untrusted input
  • No customer data used to train external models; processing terms documented

Outcome

Expected outcome (design intent, not a measured result)

These are design expectations for this pattern, not measurements from a delivered engagement. We would agree how to measure them with you before building.

We have not yet delivered this pattern for a client. If that matters for your procurement, we would rather you knew now.
  • Answers to recurring questions become consistent and traceable to an approved source
  • Time spent locating current policy or procedure falls, because the source is returned with the answer
  • Gaps in documentation become visible, because refusals are logged and reviewable
  • Staff can verify before acting, which is what makes the assistant usable in a regulated setting

Related

Illustrative solution patternAgentic AI

Document intake and extraction

Structured data extracted from inbound documents, with confidence surfaced and a person confirming before anything is committed.

Read the full write-up
Illustrative solution patternAgentic AI

Automated compliance checking

Submissions checked against a written rule set, with every finding citing its clause and a competent person confirming the outcome.

Read the full write-up
Illustrative solution patternAgentic AI

Service desk triage and resolution

Inbound requests classified, enriched and routed, with straightforward cases resolved from approved guidance and the rest escalated with context.

Read the full write-up

Agentic AI & Automation

AI agents, assistants and automated workflows that carry out real work inside your systems, with human approval at the points that matter.

Data & Analytics

Dashboards, reporting and data foundations that make operational reality visible — and make your organisation ready for AI.

All solutions

Could this work for you?

Tell us how your situation differs from this example. Where a material uncertainty remains, a bounded proof of concept with a pass threshold agreed in advance is usually the cheapest way to find out.

We reply to every enquiry within one business day.

Discuss your project