Cut claims-triage cycle from 9 days to 38 hours.

The denials team was buried. Appeals were aging out before anyone touched them. We rebuilt the triage queue around an evidence-graded agent that reads the chart, applies the medical-policy, and routes only the genuinely ambiguous cases to a human.

What was actually broken.

The payer's denials and appeals queue averaged 9.4 days from receipt to first medical-director touch, and 41% of appeals were aging past the contractual SLA. The clinical-review team of 22 was triaging by FIFO with no severity ranking. CMS had flagged the SLA breaches in the last audit. The CIO had built an internal rules engine in 2022 that nobody trusted because it could not explain its decisions, which meant clinicians overrode it 70% of the time.

What we did, and what we deliberately did not do.

We replaced the rules engine with an evidence-graded triage agent: every claim got a structured summary, the relevant policy passages cited inline, and a confidence score. Anything above 0.85 confidence flowed to fast-track. Anything below 0.6 went to a clinical reviewer with the agent's reasoning attached so they could agree, disagree, or correct in one click. The corrections fed back into a weekly retrain loop. We did NOT auto-approve or auto-deny anything in the first six months — every decision was advisory until clinical leadership signed off on the audit trail. Medical-director time dropped 64% in the first quarter.

“Our last rules engine was overridden seventy percent of the time. This one gets agreed with ninety-one percent of the time, because the clinician can see the policy passage right there. That is the whole game with clinical AI.”

What the numbers did, twelve months in.