The Documentation Crisis in Healthcare
The average physician spends 49% of their working time on EHR documentation — more than they spend with patients. For this 200-physician network, that translated to 420 hours of physician time wasted on documentation every single day. At an average physician cost of $150/hour, that's $63,000 per day in documentation labor — $16M per year.
The organization had tried traditional dictation software, but accuracy on medical terminology was poor (82%), and the workflow still required physicians to dictate explicitly rather than simply having a natural conversation with the patient.
Our Technical Approach
We built a three-layer system: (1) A custom-trained Whisper v3 model with a medical vocabulary extension that achieves 97.8% accuracy on clinical terminology — including drug names, procedures, and ICD-10 codes. (2) A GPT-4o pipeline that takes the transcript and generates a structured SOAP note, extracting chief complaint, HPI, ROS, physical exam, assessment, and plan into the correct Epic fields. (3) A physician review interface that shows the draft note alongside the original transcript, with one-click approval and inline editing for corrections.
Epic EHR Integration
The most complex part of the implementation was the Epic FHIR API integration. We built a custom SMART on FHIR application that authenticates with Epic's OAuth2 system, retrieves the patient's existing chart context (medications, allergies, problem list) to inform the AI's note generation, and writes the completed note directly to the encounter record. The full round-trip — conversation end to note appearing in Epic — takes under 45 seconds.
Results After 6 Months
Documentation time dropped from 2.1 hours/day to 40 minutes/day per physician. Physician satisfaction scores (measured by the Maslach Burnout Inventory) improved by 41 points. The organization has not lost a single physician to burnout-related resignation since deployment. Revenue capture also improved by 8% because the AI consistently documents the level of complexity that justifies higher E&M billing codes — something physicians were undercoding due to time pressure.