How We Automated 100 Daily Physician Referrals with AI

A regional sleep medicine practice came to us with a problem that sounds almost comically analog: they were receiving 50 to 100 physician referrals every single day, almost entirely via fax. Each one needed to be manually reviewed, matched to the correct patient template across multiple EMR systems, and routed to the right clinic location.

Their staff was spending hours every day on what was essentially a sophisticated sorting task. Referrals were getting lost. Patients were waiting days for callbacks. And the practice was turning away new referrals because they couldn’t process them fast enough.

The Challenge

The real complexity wasn’t just reading faxes — it was the matching logic. Each referral needed to be cross-referenced against multiple electronic medical record systems, each with its own template format and data structure. A single referral might contain a referring physician’s handwritten notes, a patient’s insurance information, prior sleep study results, and specific treatment requests.

The staff had developed an intricate mental model for routing these referrals, built up over years of institutional knowledge. Capturing that logic was as important as the document processing itself.

Our Approach

We started with a two-week discovery sprint, sitting with the referral coordinators and mapping every decision point in their workflow. We documented 23 distinct routing rules and 4 EMR template formats.

Then we built an AI document processing pipeline. The system uses intelligent document parsing to extract structured data from faxed referrals — regardless of format, handwriting quality, or layout. We paired that with a matching engine that maps extracted data to the correct EMR templates and suggests the appropriate clinic routing.

The key insight was building confidence scoring into every step. The system doesn’t just process referrals — it tells the staff exactly how confident it is in each match. High-confidence matches get auto-routed. Lower-confidence ones get flagged for human review with the AI’s best guess pre-populated.

The Results

Within six weeks of deployment, the system was processing referrals with 98% accuracy. The staff went from spending 3-4 hours daily on referral processing to roughly 30 minutes of reviewing flagged cases.

The projected annual savings came to $89,000 in labor costs alone. But the bigger win was patient experience: average callback time dropped from 2-3 days to same-day for most referrals.

The practice is now processing more referrals than ever, with fewer staff hours and fewer errors. That’s what AI should do — not replace people, but free them up to do the work that actually requires human judgment and empathy.