AI-Native Healthcare: 100M Doctor Visits, 10–20 Hours Saved, Prior Auth in Minutes — Janie Lee & Chai Asawa, Abridge

What It Covers

Abridge co-founders Janie Lee and Chai Asawa explain how their clinical AI platform processes 100 million doctor-patient conversations to reduce physician documentation burden by 10–20 hours weekly, while expanding into prior authorization automation, clinical decision support, and real-time care intelligence across major U.S. health systems.

Key Questions Answered

• •Prior Authorization Automation: Abridge reduces prior auth approvals from 45 days to minutes by cross-referencing payer policies, patient EHR data, and real-time conversation context. During a visit, the system identifies which criteria are already met and prompts the clinician to address only the remaining requirements before the patient leaves the room, collapsing a multi-week administrative process into a single encounter.
• •Three-Level Personalization Architecture: Abridge structures personalization at individual, specialty, and health system levels. Individual clinicians set style preferences like note format and phrasing. Specialty-level tuning adjusts for cardiology versus dermatology workflows. Health systems embed their own clinical guidelines directly into decision support. Each layer requires separate evaluation calibration, with specialty-level evals described as particularly hard-earned to get right.
• •Proactive vs. Reactive Alerting: Over 90% of clinical alerts are ignored due to poor context and timing. Abridge counters alert fatigue by shifting interventions upstream — summarizing patient history and visit-relevant considerations before the clinician enters the room, rather than interrupting during sensitive patient conversations. Alerts are reserved for high-stakes moments where real-time action, like prior auth criteria, produces measurable care or financial outcomes.
• •Proprietary Data Flywheel at Scale: With nearly 100 million recorded medical conversations, Abridge post-trains models on de-identified transcripts to optimize for cost and latency without sacrificing quality. At this scale, using the most capable off-the-shelf model becomes cost-prohibitive. The edit history and conversation data also feed a memory sub-agent that builds persistent clinician preference profiles, enabling personalization that improves continuously across sessions.
• •Clinician-Scientist Staffing Model: Abridge embeds MDs with technical skills — ranging from full-stack engineers to skilled prompt engineers — directly within product teams. These clinician-scientists define evaluation criteria, assess clinical usefulness, and calibrate LLM judges across specialties. This role is described as foundational to catching long-tail errors before production and is positioned as increasingly valuable as AI tooling lowers the barrier for technically-oriented clinicians to contribute directly to product development.
• •Progressive Rollout as Evaluation Strategy: Abridge treats real-world deployment as a core evaluation mechanism, not just a shipping milestone. A subset of health system customers now operate outside monthly release cycles, accepting early-access builds in exchange for rapid feedback. Offline eval sets are sized deliberately — hundreds versus thousands of responses — based on product risk level. This mirrors autonomous vehicle testing logic: expanding real-world distribution data continuously improves model reliability over time.