Automating claims triage for a regional bank

Objectives

• Reduce time-to-first-touch on inbound claims from days to hours
• Establish a defensible, regulator-ready audit trail across every triage decision
• Lift analyst leverage without changing headcount
• Retire two outsourced workflows without service disruption

Implementation approach

1. 1

   Discovery & baseline instrumentation

   Four weeks of paid discovery — process shadowing, instrumentation of the existing flow, and a measured baseline that every later metric is compared against.

2. 2

   Governance design before build

   Maker-checker model, audit log schema, role matrix and escalation rules signed off by compliance and legal before a single agent was written.

3. 3

   Phased agent rollout

   Intake classification first, document extraction second, routing third — each behind a human gate before being promoted to default.

4. 4

   Parallel running and cutover

   Eight weeks of parallel running with legacy queues kept live, decision-by-decision comparison, then phased cutover by claim type.

Technologies used

• Vestval Flow
• Private LLM deployment
• RAG over policy documents
• Document extraction pipeline
• OIDC + role-based access
• Immutable audit log

Outcomes

• Time-to-first-touch reduced from days to hours (qualitative; specifics confidential)
• Analyst leverage materially improved — same team handling significantly more volume
• Cleaner audit trail across every triage decision
• Compliance and legal sign-off retained at every decision point

Lessons learned

• Baseline instrumentation before any build is the single most valuable phase — without it, ROI is debatable forever.
• Maker-checker encoded in workflow definitions, not in convention, is what makes BFSI AI defensible.
• Parallel-run windows feel expensive and are absolutely worth it.

Next steps