Skip to content
Services
Industries
Platforms
Insights
About Us
Contact

Banking & Finance · Case Study

AI-Powered Knowledge Assistant

Secure enterprise RAG that answers policy and product questions with citations, ACLs, and evaluation gates.

  • 20K+employees using it weekly
  • 93%answer accuracy on golden set
  • 100%answers cited and ACL-checked

Delivered

What we built

Instant knowledge access for 20K+ enterprise banking users.

Enterprise-wide RAG

Retrieval across wikis, drives, policies and tickets — millions of documents, permission-aware.

Cited answers

Every response links to its source passages; no citation, no claim.

ACL-aware retrieval

Source-system permissions enforced at query time — users see only what they could open.

Quality monitoring

Live scoring, drift alerts and a feedback loop that turns thumbs-down into fixes.

Inside the build

One assistant, twenty thousand users, zero leaks

A global professional services firm wanted every employee to find any answer — without a single permission boundary breaking. Two previous vendors had failed on their document mess.

The challenge

Knowledge everywhere, answers nowhere

Six versions of every policy, scanned PDFs, knowledge scattered across five platforms with different permission models. Employees spent hours hunting or asked colleagues — and two RAG vendors had already disappointed on accuracy.

  • Five platforms with incompatible permission models
  • Duplicate and stale documents poisoning results
  • Prior vendor accuracy too low for rollout
Knowledge pipeline
Ingest→ Chunk→ Embed→ Index→ Retrieve
PDFConfluenceSharePointSQLTicketsEmail
Grounded answer…with citations [1] [2] and permissions applied

The solution

Retrieval engineered for the mess that exists

Layout-aware parsing, deduplication with version resolution, hybrid search with reranking, and ACL mirroring from every source system — validated continuously against a 400-question golden set built with the firm's own experts.

  • Version resolution keeping only authoritative documents
  • Hybrid retrieval tuned to 93% golden-set accuracy
  • Per-source ACL mirroring enforced in every query
Enterprise context
DocsCRMWiki
↓
↓
Answer with citationsscoped to the user's permissions [1] [2]

The impact

Search time became answer time

Weekly active usage passed 20,000 within a quarter — driven by trust: answers cite their sources, admit uncertainty and never surface documents the asker couldn't open. Internal help-desk tickets for 'where do I find…' dropped by half.

  • 20K+ weekly actives within one quarter
  • 50% drop in knowledge-related help-desk tickets
  • Zero permission incidents since launch
Production dashboard
99.9%Uptime
1.2sp95 latency
$0.021Cost / task
QualityCostLatencyDrift

Outcomes

What changed for the firm

Onboarding speed

New joiners productive weeks earlier with instant answers.

Policy compliance

Current versions surfaced; stale guidance flagged.

Expert leverage

Senior staff interrupted less for known answers.

Trust in AI

Citations made the assistant credible enough to rely on.

Platform reuse

The retrieval layer now powers three other AI tools.

Security posture

Access boundaries proven intact under audit.

Delivery

From challenge to outcome

We publish the architecture decisions, delivery approach and measurable results — so you can see how similar outcomes could apply to your enterprise.

  • Measurable KPIs
  • Architecture transparency
  • Cross-industry patterns
  • Repeatable playbooks
  1. 1
    Challenge

    Business pressure, legacy process and constraints.

  2. 2
    Solution

    Architecture, agents and integration design.

  3. 3
    Delivery

    Build, evaluate, secure and launch.

  4. 4
    Impact

    Efficiency, cost, quality or revenue outcomes.

The citation model won people over — you can check every answer. And when it doesn't know, it says so. That honesty is why twenty thousand people trust it.
Chief Knowledge OfficerGlobal professional services firm

Under the hood

The stack behind the assistant

Deployed in the firm's Azure tenant with their identity stack end to end.

  • Anthropic Claude
  • Azure AI Search
  • Elasticsearch
  • Unstructured
  • SharePoint
  • Confluence
  • Okta
  • Azure
  • Braintrust
  • Datadog

FAQ

Common questions

How do you keep 93% accuracy as content changes?

Continuous golden-set evals plus live sampling — content drift or index staleness fires alerts before users notice. The golden set grows with every reported miss.

What happens when someone asks about restricted content?

The retrieval layer simply can't see it for that user — permissions are query-time filters, not post-hoc redaction. The assistant answers from what the asker is entitled to, or says it can't help.

Could this work with our document chaos?

The chaos is the normal case — dedup, version resolution and layout-aware parsing exist precisely for it. A pilot on your worst corpus is the honest test.

Next step

Want answers your people can trust?

Send ten real questions and a sample of your messiest documents — we'll show you what accurate looks like.

Book a Strategy Call
ISO 9001:2015 certificate issued to Skugal Technologies Private Limited, certificate number QD-98140/1125

Quality management system

ISO 9001:2015 Certified

Skugal Technologies Private Limited has been independently assessed and found to comply with ISO 9001:2015 for providing quality IT services and software development solutions.

Certificate No.
QD-98140/1125
Standard
ISO 9001:2015
Issued
24 Nov 2025
Valid until
23 Nov 2028
Certification body
Revolutionary Certifications Services (RCS Cert)

Client ID 95268. Validity is subject to annual surveillance audits (next due 23 Nov 2026). Verify at thesasl.co.uk.