Automating 100 Daily Physician Referrals with AI

A regional sleep medicine practice receives 50 to 100 physician referrals per day. Almost all arrive via fax. Each referral requires data extraction, cross-referencing against multiple EMR systems, template matching, and routing to the correct clinic location. Their staff was burning 3-4 hours daily on what is fundamentally a classification and data mapping problem. Referrals were getting dropped. Patient callback times stretched to 2-3 days. The practice was rejecting new volume because the pipeline could not keep up.

The Engineering Problem

The hard part was not OCR. The hard part was the matching logic downstream.

Each referral contains a mix of structured and unstructured data: referring physician details, patient demographics, insurance information, prior sleep study results, treatment requests. Handwriting quality varies wildly. Layouts are inconsistent across referring practices.

That extracted data then needs to map against 4 distinct EMR template formats, each with its own field schema and data structure. The routing logic — built up over years of institutional knowledge — encompassed 23 discrete decision rules that the coordinators ran in their heads. No documentation. No flowcharts. Pure tribal knowledge.

Capturing that decision tree was the real engineering challenge. The document processing was table stakes.

Pipeline Architecture

We ran a two-week discovery sprint embedded with the referral coordinators. Every routing decision was mapped, validated, and codified into explicit rule definitions.

The processing pipeline has three stages:

Stage 1: Document Ingestion & Extraction. Inbound faxes hit an intelligent document parser that extracts structured fields regardless of source format, layout, or handwriting quality. The parser outputs a normalized data schema — patient demographics, referring provider, insurance, clinical history, requested services.

Stage 2: EMR Template Matching. The extracted data runs through a matching engine that resolves against the 4 EMR template formats. Field mapping handles schema differences between systems. Fuzzy matching catches name variants, abbreviation mismatches, and partial data.

Stage 3: Confidence-Scored Routing. Every match produces a confidence score. High-confidence referrals (above threshold) auto-route to the correct clinic location with the appropriate EMR template pre-populated. Below-threshold referrals get flagged for human review — but the system pre-fills its best guess, so the coordinator is validating rather than starting from scratch.

The confidence scoring is the critical design decision. Binary pass/fail automation breaks trust. A system that tells you exactly how sure it is earns adoption.

Results

Six weeks from deployment to production stability.

• 98% processing accuracy across all referral types
• Staff time: 3-4 hours/day down to ~30 minutes of reviewing flagged cases
• Patient callback time: 2-3 days reduced to same-day for the majority of referrals
• $89,000 projected annual savings in labor costs alone

The practice now processes higher referral volume with fewer staff hours and fewer errors. The coordinators spend their time on the cases that genuinely require human judgment — complex insurance situations, ambiguous clinical requests, edge cases the system correctly flags as uncertain.

That is the right use of AI in healthcare operations. Not replacing clinical staff. Eliminating the mechanical throughput bottleneck so they can focus on the work that requires expertise and empathy.