May 14, 2026 · 1h 6m · latent-space
Inside Abridge: The AI Listening to 100 Million Doctor Visits — Abridge's Janie Lee & Chai Asawa
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this technical crossover episode, Abridge leaders Janie Lee and Chai Asawa explore how ambient clinical intelligence is transforming healthcare workflows, scaling from automated medical documentation to real-time clinical decision support across 100 million patient encounters.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 4.9 Guest teaching 5.2 Guest disagreement 0.9 The hosts pushing back 1.6
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 58:57 Janie's contrarian defense of PRDs against prototype culture
Janie directly challenges the prevailing AI startup consensus that PRDs are dead and rapid prototypes are sufficient, arguing that complex clinical enterprise software requires deep written strategic clarity.
Hardest push from the hosts ▶ 1:00:57 Shawn pushes back on Janie's high bar for PRDs vs doing it firstShawn refuses Janie's strict requirement for proving moats upfront in a PRD, countering that in fast-moving AI markets the winning answer is frequently simply executing first.
Biggest teaching moment ▶ 9:45 Janie educates on real-time prior authorization in clinical workflowsJanie walks through an intricate real-world clinical scenario showing how ambient AI transforms delayed multi-week insurance prior authorization into instant in-exam compliance checks.
The host holds their own ▶ 5:22 Jacob draws architectural parallels between Glean and clinical decision supportJacob demonstrates sharp technical and venture expertise by breaking down clinical decision support into a core enterprise search problem, directly connecting it to Chai's background at Glean.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Proactive Clinical Intelligence and Background Ambient Philosophy | 3 | 5 | 0 | 0 | Shawn opens by openly acknowledging he feels underqualified in healthcare, inviting Janie and Chai to define Abridge's core thesis as a clinical intelligence layer. | |
| Abridge's Evolution: Saving Time, Money, and Lives | 6 | 5 | 1 | 1 | Jacob demonstrates domain depth by comparing clinical decision support directly to enterprise search architecture from Chai's Glean background. Chai articulates the high downside risk and vertical focus of healthcare AI. | |
| Solving Alert Fatigue with Real-Time Prior Authorization | 4 | 7 | 1 | 1 | Janie educates the hosts on solving alert fatigue by shifting from reactive alerts to proactive real-time prior authorization while the patient is still in the room. | |
| Ambient Form Factors, Hardware, and In-Room Modalities | 5 | 4 | 2 | 4 | Shawn actively pushes back with skepticism regarding whether clinicians actually want AR glasses in clinical settings, prompting Chai and Jacob to clarify specialized surgical use cases. | |
| Enterprise Dynamics: Aligning Hospital Buyers, CFOs, and Payers | 4 | 5 | 0 | 1 | Shawn clarifies terminology around health systems, CFO incentives, and payers. Janie explains the multi-stakeholder dynamic balancing CMIOs, CFOs, clinicians, and patients. | |
| Machine Learning Frontiers: 100 Million Conversations as Traces | 6 | 6 | 1 | 1 | Chai discusses the ML engineering trade-offs of real-time conversational agents and conceptualizes Abridge's 100 million conversation traces as diagnostic debugging exhaust. | |
| Real-Time Agent Architecture and the Exam Room Dynamic | 5 | 5 | 1 | 2 | Shawn presses Chai on whether real-time agent execution is true streaming or batched intervals, and why voice-in voice-out agents are currently too intrusive for patient visits. | |
| Three Tiers of Personalization: Individual, Specialty, and System | 5 | 6 | 2 | 3 | Shawn challenges the assumption that hospital guidelines vary significantly, asking if they all converge to the same standards. Chai and Janie clarify how local specialty nuances drive customized decision pathways. | |
| Architecting Decoupled Memory Stores for Clinical AI | 5 | 6 | 0 | 1 | Janie and Chai break down their decoupled external memory store architecture and the rigorous 'Look at the F***ing Data' (LFD) eval process using in-house clinician scientists. | |
| HIPAA Compliance, PHI De-Identification, and Safety | 4 | 5 | 0 | 1 | Shawn inquires about PHI scrub models and whether de-identification is strictly one-way, while Chai and Janie explain contractually enforced privacy and anonymization pipelines. | |
| Scaling AI Infrastructure to 100 Million Encounters | 5 | 5 | 0 | 1 | Jacob observes that Abridge operates in post-scale optimization mode where token economics force post-training rather than burning frontier model tokens naively. | |
| EHR Interoperability and Deep Integration Moats | 5 | 5 | 1 | 2 | Shawn asks whether incumbent EHR vendors will eventually swallow this layer. Janie and Chai explain that interoperability and cross-stakeholder intelligence moats lie outside traditional EHR scope. | |
| Regulatory Tailwinds and Cascaded Fast/Slow Models | 5 | 6 | 1 | 2 | Chai explains how shifting FDA regulatory guidelines provide unexpected tailwinds, and details cascaded 'thinking fast and slow' model architectures to reduce inference latency. | |
| Clinician Scientists and Active Learning on Edge Cases | 6 | 5 | 0 | 1 | Shawn connects Abridge's hybrid clinician-engineer staffing model to active learning on edge cases, which Janie affirms as critical for high-stakes clinical evals. | |
| Reflections on Glean, Search, and Vertical AI Moats | 6 | 4 | 1 | 1 | Jacob and Chai reflect on Glean's engineering origins from Google Search and discuss how foundation model hyperscalers target broad horizontal knowledge work while vertical AI retains defensibility. | |
| Durable AI Infrastructure and Event-Driven Systems | 5 | 5 | 0 | 1 | Chai highlights durable infrastructure primitives like event-driven streaming (Kafka, Temporal) and collaborative CRDTs as lasting paradigms for real-time multi-agent systems. | |
| Product Philosophy Debate: PRDs vs Rapid Prototyping | 6 | 6 | 4 | 4 | Janie strongly rejects the popular Silicon Valley consensus that 'PRDs are dead and prototypes are everything,' arguing rigorous written clarity is essential for complex enterprise workflows. Shawn pushes back on the viability of that standard when speed matters. | |
| Developer Tooling and the Impact of Claude Code | 4 | 3 | 1 | 1 | Shawn and Jacob wrap up by discussing internal engineering adoption of Claude Code and Cursor, and Chai plugs Abridge's technical white papers and upcoming AI events. |
Statements from this episode (25)
Lee: Clinicians spend 10 to 20 hours weekly on documentation
“They're spending 10 to 20 hours a week on documentation.”
Lee: Patient-doctor conversations drive nearly all downstream healthcare spending
“Conversations between patients and clinicians are probably the most important workflow in healthcare. It's, Obviously, where care is given and received, but if you think about the 20% of our GDP that goes towards healthcare, almost everything is a derivative o…”
Lee: Abridge software is opened millions of times per week
“The fact that our software and our product is open millions of times a week before, during, and after a patient walks in the room,”
Asawa: Many more AI products will become ambient and always-listening
“We're ambient, and we're always listening in the background. And I think many more AI products will go that way, but it's actually how we started.”
Asawa: Proactive ambient AI without screens is the best form of AI
“And I think that's actually the, like, the greatest form of AI we can create. AI that's actually seamless. You're not actually looking at your screen. It's always there. It's always helping you out and being proactive.”
Lee: At Opendoor, every pricing outlier wiped out 30 transactions' margins
“When I worked at Opendoor, I worked on pricing models, like every outlier wiped out the margins of 30.”
Lee: In-room AI can collapse prior authorization latency to real time
“If we can pull forward The use of both the AI, but also the ability to solve problems when the patient's in the room. You can start to collapse what typically takes weeks or months after your visit ideally down to minutes or real time”
Lee: Hospital CFOs Require Direct Financial ROI Beyond Clinician Time Savings
“For CFOs, they care a lot more than just time savings. We have to show for every dollar you put into a bridge because you have more compliant documentation or because you have fewer queries coming from your billing team we actually save or add real dollars.”
Asawa: Abridge holds proprietary dataset approaching hundreds of millions of medical conversations
“So for example, we have on the order of eighty million or hundreds of millions actually now getting close to of medical conversation.”
Asawa: Almost every AI agent functions as a coding agent underneath
“Almost every agent is a coding agent underneath underneath the hood, right? So you give it whatever, a file system, it can write its own code and so forth.”
Lee: Clinicians and patients do not want a spoken AI voice in exam rooms
“Right now patients, clinicians don't want a third voice, at least in a literal voice in that room”
Asawa: AI 'slop' is fundamentally AI operating without context
“A writing that doesn't feel like your own, and then we call that slop, but the way I describe one framing of slop is, like, AI without context”
Asawa: Decouple User Preferences From Model Weights to Avoid Throwaway Fine-Tuning
“When you think the models are rapidly changing, whether it's in-house or third party, baking into the model weights, sometimes you worry that it could be a little throwaway. And so, like, how do you need to find a way that you decompose the problem, the prefer…”
Abridge Uses Separate Memory Stores and Background Agents for Clinician Personalization
“The thing we're right now, most But that's easiest to start with, and we're excited about, is having, like, a separate store for memory, where you actually have, for example, like, a memory sub-agent that's, like, working in the background, figuring out what a…”
Lee: Clinical AI Evaluation Velocity Requires Operational Gains Over Pure ML
“And so I think as much of it is an ML problem, like so much of it has also been operational gains that I think are hugely, hugely important where domain specific expertise is, is everything.”
Lee: Abridge Moved Health System Customer Releases to Monthly Cadence
“Historically, a lot of these health systems, when they bring on new vendors, their release cycles are quarters, sometimes twice a year. We've gotten our customers onto monthly release cycles, which is pretty fast for health systems”
Lee: Clinicians reject healthcare software adding two daily clicks
“Anytime we introduce a new product that, you know, adds two clicks for them in their day, they're like, we're not going to use it.”
Lee: Healthcare will solve the hardest AI problems first
“I think for the very reasons things are higher stakes or, you know, potentially considered more difficult in healthcare, I think it's where some of the hardest AI problems will get solved first, just because the bar is so high.”
Lee: Strict zero-error requirements force AI innovation in healthcare
“When you think about like zero error, Evals or multi-step workflows that have really, really low tolerance. I actually think a lot of the innovation will happen here just because we have to, or else we can't chip.”
Asawa: Past enterprise search startups failed on quality, not market viability
“I think a lot of times enterprise search startups failed because the quality wasn't great enough, but the learning that people took away from that is, oh, enterprise search is not good enough. And so, like, quality, I think, really changes the game of, like, i…”
Effron: Foundation model companies must win consumer, coding, and enterprise search
“Always my mental model is like, there's a few markets that are like the foundation model companies have to win or like big enough to go after. And it's probably like consumer code and that.”
Asawa: Infrastructure built for human collaboration will remain durable for AI agents
“So all these things that we've built for, I actually think the things we've built for humans are actually the things that are going to be continually durable.”