Cutting claim-triage cycle time by 62% at a top-5 US health insurer
A top-5 US health insurer was drowning in claim-triage backlog during open enrollment. pronix.ai designed and shipped an agentic triage workflow that took cycle time from 11 minutes to under 4, tripled throughput, and stayed inside healthcare and CMS auditability guardrails throughout.
- Client
- Top-5 US health insurer
- Industry
- Health Payers
- Platform
- Amazon Bedrock · Kore.ai · Salesforce Health Cloud
*Representative outcome; results vary by client, scope and platform configuration.
The challenge
Claim-triage was a 7-step human workflow averaging 11 minutes per case, with 22-day backlogs during open enrollment and $18M/year in avoidable adjudication cost. Prior copilot rollouts had moved the needle by ~9% — nowhere near what was needed.
Our approach
Bounded the agent to one intent
We scoped the agent to first-pass triage only — routing, coding validation, missing-document detection — with a mandatory HITL step before any denial. That single scope decision made the safety review tractable.
Built the tool layer on existing systems
Nine tool schemas over Health Cloud, the claims platform and the document store. All idempotent, all logged, all replayable. No net-new systems of record.
Two-provider routing on Bedrock
Claude and an Amazon Titan fallback with prompt caching and evaluation gates. Cheap-first cascade brought per-case model cost from $0.34 to $0.09.
HITL as a first-class surface
Every denial and every low-confidence case routed to a claims examiner with a pre-populated review view — 40 seconds instead of 4 minutes.
Evaluation harness in CI
A golden set of 1,200 historical cases with LLM-as-judge plus human spot-check. Every model or prompt change ran the harness before promotion.
“pronix.ai treated our regulator like a stakeholder from day one. That is why this shipped — and why it stayed shipped.”
Illustrative case study. Scenarios, metrics, quotes and client details are representative composites based on Pronix engagements and industry benchmarks unless a named client is shown with written consent. Outcomes vary by client, scope, data quality and platform configuration. Nothing on this page is a guarantee, warranty or professional advice. See our Terms of Use for the full disclaimer.
AWS, Salesforce, Microsoft, NICE, Genesys, Kore.ai, Google, ServiceNow, Amazon Connect and logos referenced on this site are trademarks of their respective owners. References are for descriptive purposes only and do not imply endorsement, sponsorship or partnership beyond stated partner relationships. See our Disclosures.
Talk to the team that shipped it.
Book a working session with a pronix.ai lead on this practice — we'll map the approach to your platform, industry and constraints.