Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Inside Abridge: The AI Listening to 100 Million Doctor Visits — Abridge's Janie Lee & Chai Asawa
TL;DR
Abridge is moving from AI scribes to a full clinical intelligence layer — Janie Lee frames the roadmap as three acts: save clinicians time, help health systems save/make money, and ultimately save lives by using the patient-doctor conversation as the core data source.
The product thesis is 'air conditioning,' not pop-up hell — With over 90% of healthcare alerts ignored, Abridge wants AI to stay ambient and only surface when context makes the intervention truly worth it, like catching missing prior-authorization criteria before a patient leaves the room.
Scale is the moat: Abridge says it has nearly 100 million medical conversations — Chai Asawa argues that this dataset captures the real 'debugging trace' between patient and clinician, which powers better transcription, note generation, personalization, and future agentic workflows.
Healthcare makes AI product design brutally hard because 80/20 doesn't work — A wrong answer in enterprise search is annoying; a wrong recommendation in care can be fatal, so Abridge emphasizes de-identification, specialty-specific evals, progressive rollout, and clinician-in-the-loop validation.
Personalization happens at three levels: doctor, specialty, and health system — Abridge tunes output for individual quirks like 'two spaces between sentences,' for specialty-specific workflows like dermatology vs. cardiology, and for local hospital guidelines that affect decision support.
The real opportunity is reducing latency in healthcare, not just writing notes faster — Both guests return to the idea that AI can collapse workflows that usually take weeks or 45 days—like prior auth, lab follow-ups, or coding queries—into minutes or real-time actions while the patient is still in the room.
The Breakdown
From 'pajama time' to a broader healthcare platform
Janie introduces Abridge as a clinical intelligence layer for health systems that started with documentation because doctors still spend 10 to 20 hours a week on notes. She shares the most human proof point right away: clinicians saying Abridge helped them stop eating dinner late, avoid early retirement, and even, in one quote, 'we're not divorcing anymore.'
Why the patient conversation is the center of everything
The company’s core bet is that the doctor-patient conversation is the most important workflow in healthcare because claims, payments, diagnoses, and treatment all flow downstream from it. Chai says Abridge is now pushing before, during, and after the visit: if you have the full context of the patient, payer rules, guidelines, and literature, healthcare can start to look fundamentally different.
The 'air conditioning' product philosophy
Asked how to avoid alert fatigue, Janie gives the line that sticks: Abridge should feel like air conditioning—quietly improving things in the background. Instead of bombarding doctors mid-visit, the product should often prep them beforehand with summaries and recommended discussion points, then reserve intervention for genuinely high-risk or high-value moments.
A prior-authorization example that shows why this is so hard
Their MRI example makes the promise concrete: if a patient on an Aetna plan in California needs six criteria met and four are already known, Abridge can nudge the clinician to ask the final two questions before the patient leaves. That turns a maddening four-week insurance delay into something close to instant approval, but only if Abridge can integrate EHR data, understand payer policies buried in PDFs and websites, and do it in real time.
Why healthcare AI is different from Glean-style enterprise search
Chai, who came from Glean, says the similarity is obvious—context is king—but the differences are what matter: the downside risk is much higher, the vertical is narrower, and Abridge started in ambient audio rather than search boxes. He calls that ambient approach closer to the real 'Jarvis' vision: AI that is always there, not demanding your attention.
Building with a constellation of models, not one magic model
Abridge uses off-the-shelf models where they help, but trains its own systems when proprietary data can improve quality or reduce cost and latency. Chai says the company now has on the order of 80 million to nearly 100 million medical conversations, which act like healthcare 'agent traces'—the raw material for better note generation, transcription, diarization, and future agents that can navigate the EHR like a file system.
Personalization, memory, and the weirdly hard details that matter
Janie breaks personalization into three layers: the individual doctor, the specialty, and the health system. That means handling everything from bullet points vs. paragraphs to specialty-specific billing requirements, and even health-system-specific clinical pathways; she jokes that some doctors demand 'two spaces between sentences or I refuse to use this tool.'
Evals, privacy, and why trust is earned in drops and lost in buckets
The eval stack is intense: in-house clinicians, third-party reviewers, specialty-specific criteria, LLM judges, and progressive rollout inspired partly by self-driving. Chai explains that any training or eval data has to be de-identified under strict PHI rules, while Janie says customer trust has become a huge advantage—enough that some large health systems now let Abridge ship outside normal quarterly release cycles and co-develop faster.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
