Case Studies

---

• Problem: Every query went to a frontier model, so costs rose faster than use.
• What we did: 
  • Built caching and short- and long-term memory.
  • Added a small model that routes each query to the cache, a smaller model or the frontier model only when needed.
  • Retuned the routing in Continuity as models and prices changed.
• Results: 
  • 35% lower monthly token bill
  • 26% → 40% of queries kept off the frontier model in six months
  • 95–100% accuracy held on the agreed test set
• Note: Built at three organisations.

• Problem: By month six, a compliance assistant on bank regulatory requirements gave inconsistent answers, and users stopped trusting it.
• What we did: 
  • Found the cause: chunks of 200 tokens were splitting the meaning of paragraphs.
  • Redesigned retrieval with chunks of 450 tokens and 20% overlap.
  • Reviewed the embeddings.
• Results: 
  • 60% → 85% answer accuracy
  • Adoption stabilised
• Note: Firm of about 1,200 staff.

• Problem: The challenge was to grow use of an AI assistant after go-live without accuracy slipping.
• What we did: Ran monthly value and quality reviews, with routing and content tuned on the evidence.
• Results: 
  • +144% AI-handled requests in 12 months (from about 2,500 to about 6,100 a month)
  • Accuracy held at about 90%

 Siloed systems were producing inconsistent records ahead of MAS and Basel III audits. Manual reconciliation was slow, error-prone, and couldn't scale.

What we did: Built a data governance layer standardizing metadata across transaction systems. Deployed an AI reconciliation model that auto-matched transactions and flagged anomalies. Integrated RAG-based compliance lookups so auditors could query by jurisdiction, reporting period, and product type.

30% — Reduction in reconciliation effort Hours — To answer regulator queries, down from weeks

7-month engagement · MAS & Basel III context

 Results: 

• 30% less reconciliation effort
• Hours, not weeks, to answer regulator queries

Note: Built at three organisations.

A rules-based fraud system missed subtle patterns and generated too many false positives—slowing claims and frustrating customers who had done nothing wrong.

What we did: Built a data governance and stewardship layer aligned with APAC insurance regulatory requirements. Deployed an AI-driven underwriting and fraud intelligence platform combining supervised ML with RAG-based compliance lookups. Integrated jurisdiction-specific metadata filtering for regulator queries.

20% — Improvement in underwriting decision accuracy 

25% — Reduction in claims cycle time

 Results: 

• 20% better underwriting decision accuracy
• 25% shorter claims cycle

6–8 month engagement · APAC insurance context

Rising fraud in subsidy disbursement, compounded by high false positives that consumed investigative resources — and a system that couldn't explain its own decisions.

What we did: Designed a graph-based fraud analytics engine to surface relationships between applicants, providers, and claims. Layered RAG compliance checks against PDPA and Singapore's AI governance guidelines. Embedded explainability dashboards so investigators could defend every flagged case.

15% — Uplift in fraud detection rate; 

40% — Reduction in false positives

 Results: 

• 15% higher fraud detection rate
• 40% fewer false positives

5–6 month engagement · 

Referenced as a Singapore AI governance pilot

Leadership wanted to move beyond copilots to agent-driven automation — but needed confidence that the organization was ready to scale safely, with consistent controls and clear accountability for agent actions.

Tags: AI TrustX™ · Agentic AI · Maturity Assessment · Governance

Outcomes: · Clear, board-ready answer to "Are we ready to scale Agentic AI?" · Standardized trust gates and evidence requirements across all AI initiatives · Prioritized use-case sequencing tied to measurable outcomes · Improved auditability — action logging, traceability, and accountability for agents

Strong AI momentum — pilots, vendor POCs, growing business demand — but no repeatable enterprise model to scale AI safely, audibly, and consistently across business lines.

Tags: AI TrustX™ · Data Governance · Responsible AI · Operating Model

Outcomes: · Faster pilot-to-production via standard trust gates and reusable templates · Clearer decision rights and audit-ready evidence per AI use case · Reduced late-stage rework through upfront assurance design · Confidence to scale GenAI responsibly across the enterprise