Healthcare / Hospital Operations

Prior Authorization AI: 74% Faster Approvals

An agentic AI system using RAG over payer policy documents that automates prior authorization clinical criteria review — cutting processing time from 4.2 hours to 65 minutes and denial rates by 31%.

Healthcare AIPrior AuthorizationRAGClaude APILangChainFHIR R4.NET 8
74%
Reduction in PA processing time (4.2 hrs → 65 min)
31%
Reduction in administrative denial rate
3,500+
PA requests processed monthly
$1.8M
Estimated annual revenue recovered from avoidable denials

The Challenge

The Challenge: Prior Authorization Was Destroying Clinical Productivity

Prior authorization is one of US healthcare's most expensive administrative burdens. A large regional health system with 12 hospitals and 200+ employed physicians was spending an average of 4.2 staff hours per prior authorization request — across 3,500+ requests per month. The process required a clinical authorisation specialist to manually cross-reference physician order documentation against payer-specific clinical criteria — criteria that varied by payer, plan, procedure code, and diagnosis code, and that were buried in PDF policy documents updated without notice. The team was processing thousands of PDFs per month. Specialist productivity was collapsing, physicians were frustrated by delays, and the health system was losing $2.1M annually in revenue from avoidable administrative denials — where the clinical criteria were actually met but the documentation submitted was inadequate. The prior auth team was at capacity. Hiring more staff was cost-prohibitive. The organisation needed a fundamentally different approach.

The Solution

Solution: Agentic AI with RAG over Payer Policy Documents

I designed and built an agentic AI prior authorization assistant that combines a FHIR R4 clinical data extraction layer with a RAG system over the health system's library of payer policy documents, orchestrated by a Claude 3.5 Sonnet agent that performs the clinical criteria matching and generates structured documentation recommendations.

1

FHIR R4 Clinical Data Extraction

Built a .NET 8 service that extracts relevant clinical data for each PA request from the health system's Epic FHIR R4 API: the patient's clinical history (Conditions, Observations, Procedures, MedicationRequests), the ordering physician's documentation (DiagnosticReports, ClinicalNotes), and the specific procedure or medication being authorised. All extraction respects SMART on FHIR scopes and HIPAA minimum necessary access requirements.

2

Payer Policy RAG System

Built a RAG pipeline over 8,000+ payer policy documents — coverage determination policies, clinical criteria guidelines, and formulary requirements — for 47 commercial, Medicare Advantage, and Medicaid payers. Documents are chunked, embedded using text-embedding-3-large, and stored in pgvector. The retrieval layer uses hybrid search (dense + BM25) with reranking to surface the specific clinical criteria sections relevant to each PA request's procedure code, diagnosis codes, and payer-plan combination.

3

Claude AI Orchestration Agent

A Claude 3.5 Sonnet agent that receives the structured clinical context and retrieved payer criteria, performs clinical criteria matching, identifies documentation gaps, and generates a structured PA recommendation: whether criteria are met, what supporting documentation is sufficient, and — where criteria are not clearly met — what additional clinical documentation would strengthen the request. The agent's output feeds a staff review interface where PA specialists validate and submit.

Technology Stack

Tools & Technologies Used

.NET 8C#Claude 3.5 SonnetClaude APILangChainpgvectorPostgreSQLtext-embedding-3-largeFHIR R4Epic FHIR APISMART on FHIRAzureAngular 17Azure Service BusPython

We went from our PA team being completely overwhelmed to processing the same volume with a third of the staff time. The AI doesn't replace clinical judgment — it gives our specialists exactly the right information to make faster, better-documented decisions.

Director of Revenue Cycle
Regional Health System

Related Services

Services Used in This Project

Have a similar project?

Let's discuss your requirements. Book a free 30-minute consultation.

Book Free Consultation