Key Moments
Inside Abridge: The AI Listening to 100 Million Doctor Visits — Abridge's Janie Lee & Chai Asawa
Latent Space PodcastKey Moments
Abridge uses AI to listen to 100 million doctor visits, aiming to streamline healthcare documentation and decision support. Their success hinges on integrating vast, complex data, but the high-stakes nature of healthcare demands extreme accuracy and a unique approach to AI development.
Key Insights
Doctors spend 10-20 hours per week on "pajama time" for documentation, with Abridge aiming to alleviate this burden.
Over 90% of medical alerts are ignored, prompting Abridge to focus on proactive, context-aware intelligence rather than reactive alerts.
The prior authorization process, which can delay care for weeks, is a key area where Abridge aims to reduce latency by providing information at the point of the doctor-patient visit.
Abridge has collected "on the order of 80 million or hundreds of millions" of medical conversations, creating a unique dataset for training AI models.
Abridge has "multiple stats professors on staff" contributing to rigorous evaluation and research, including a white paper on reducing hallucinations.
Clinicians using Abridge have reported significant life improvements, including avoiding early retirement and spending more time with their families.
Reducing clinician burden through ambient listening
Abridge's core mission is to build a clinical intelligence layer for health systems, starting with ambient documentation. Clinicians currently spend an estimated 10-20 hours weekly on "pajama time," which refers to after-hours work on notes and documentation. This burden, coupled with a significant doctor shortage, highlights the critical need for AI solutions. Abridge's technology acts as a "clinical intelligence layer," primarily by listening to doctor-patient conversations. This ambient approach aims to reduce the documentation workload, allowing clinicians to focus on patient care and potentially avoid burnout or early retirement. The company sees the doctor-patient conversation as the "most important workflow in healthcare," as nearly 20% of GDP is spent on healthcare, and most processes are derived from these interactions.
From documentation to proactive intelligence
Initially focused on saving clinicians time through documentation, Abridge is expanding its scope to become a more proactive intelligence tool. This involves leveraging the vast amount of contextual data captured from patient interactions to offer support before, during, and after a visit. The goal is to shift from reactive alerting, which is notoriously ineffective in healthcare (over 90% of alerts are ignored), to proactive intelligence that intervenes at the most critical moments. Abridge likens its ideal product to "air conditioning"—present but unobtrusive, making things better in the background. However, significant clinical risk necessitates a more active intervention when crucial. Instead of interrupting a sensitive conversation, the aim is to prepare clinicians beforehand by summarizing relevant patient history and suggesting discussion points based on the visit's purpose.
Addressing healthcare latency with intelligent automation
A key area of focus for Abridge is reducing "latency" in healthcare, particularly in processes like prior authorization. Traditionally, obtaining approval for procedures like MRIs can take weeks, leading to delays in patient care. Abridge's solution aims to address this by providing real-time, context-aware alerts to clinicians during a patient visit. For example, if a patient is prescribed an MRI, Abridge can quietly inform the doctor about specific criteria (like physical therapy history or pain duration) required by the patient's insurance plan, which can be confirmed with the patient before they leave the office. This proactive approach, leveraging integrated EHR data and payer policies, can ensure approvals happen before the patient departs, significantly collapsing the timeline from weeks to minutes. This not only improves patient outcomes but also helps health systems save money by reducing administrative friction.
The challenges and advantages of healthcare AI
The healthcare domain presents unique challenges for AI development. The stakes are extraordinarily high; errors can be fatal, unlike in more consumer-facing AI applications. This necessitates a rigorous evaluation strategy with progressive rollouts. Abridge views its focus on healthcare as a "vertical" play, allowing for more specialization compared to "horizontal" solutions. The company's starting point with ambient listening is highlighted as a key differentiator, enabling a truly seamless AI experience. The complexity of integrating diverse data sources—including EHRs, payer policies scattered across various documents, and medical literature—creates a significant "moat" for Abridge. The need for high AI model accuracy and deep workflow integration further solidifies its position.
Data scale and proprietary models
Abridge has amassed a substantial dataset, with "on the order of 80 million or hundreds of millions" of medical conversations. This unique corpus serves as a "trace" between patients and providers, offering invaluable data for training AI models for transcription, de-identification, and note generation. While Abridge leverages third-party models, it also develops proprietary models where its unique data can provide an advantage in cost, latency, or quality. They anticipate that general AI models will continue to improve in healthcare knowledge, but their proprietary dataset allows for specialized optimization, leading to a "constellation of models" tailored for the best product experience.
Personalization and health system-specific needs
Personalization is crucial across three levels: individual clinicians, medical specialties, and entire health systems. Clinicians have preferences for note style (e.g., bullets vs. paragraphs, conciseness), specific phrases, and templates. Specialties like cardiology and dermatology require different workflows and output expectations. Abridge also works with health systems to embed their own guidelines and best practices into the AI tools, particularly for clinical decision support. This deep integration allows Abridge to act as a trusted partner for health systems looking to standardize care based on their unique protocols, further strengthening their competitive moat.
Evaluation, privacy, and building trust
Ensuring clinical safety and quality is paramount. Abridge employs internal clinicians for "Look, Feel, and Do" (LFD) evaluations and utilizes LLM judges calibrated with annotated data. The process of evaluation has evolved from months to weeks and days through operational efficiency and domain expertise. Privacy is managed through rigorous de-identification of data using specialized models, ensuring compliance with HIPAA. While this de-identification is irreversible, partnerships and specific data contracting with customers allow for close collaboration. The trust earned from health systems has enabled Abridge to move from quarterly release cycles to monthly, with some customers even co-developing outside these cycles, a testament to the value and reliability Abridge delivers.
The future of AI in healthcare: Beyond 80/20
Unlike many AI applications where an "80/20" approach suffices, healthcare demands near-perfect accuracy. This high bar, driven by regulatory requirements and patient safety, is seen as a catalyst for advancing AI innovation. Abridge is building towards more agentic capabilities, such as reacting to lab results or proactively managing background tasks on behalf of clinicians. The company envisions a future where a single conversation can serve multiple stakeholders—clinicians, patients, payers, and pharmaceutical companies—creating a unified intelligence layer. This ambition aligns with the "triple aim" of healthcare: improving quality, reducing latency, and lowering costs, all starting from the foundational patient-provider conversation.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Concepts
Common Questions
Abridge is a clinical intelligence layer for health systems that aims to reduce the burden of documentation for clinicians, improve patient outcomes, and help health systems save money. It uses AI to analyze conversations and provide actionable insights.
Topics
Mentioned in this video
Johns Hopkins University, mentioned in the context of Michael, a professor at JHU who contributed to Abridge's white paper on reducing hallucination.
Mentioned as an example of a large health system that would be a customer of Abridge.
Mentioned as an example of a large health system that would be a customer of Abridge.
An AI tool mentioned as being used by the engineering team at Abridge.
A previous company where Chai Asawa worked, mentioned in the context of rapid iteration and deployment at scale, contrasted with Abridge's current approach.
Mentioned as a provider of APIs, specifically referencing its real-time API capabilities in the context of latency discussions.
A clinical intelligence layer for health systems that started with documentation and is expanding into clinical decision support and other areas. They aim to reduce clinician burden, save health systems money, and ultimately save lives.
A previous company where Chai Asawa worked as an early engineer, focusing on enterprise search and information retrieval. It is mentioned as a comparison point for Abridge's technological foundations and problem-solving approach.
Electronic Health Record, referred to as a 'file system' for patient information and a critical integration point for Abridge's platform. The partnership with EHRs is essential for deep integration and clinician adoption.
An AI tool mentioned as being enjoyed by the engineers at Abridge, particularly for its utility in onboarding and learning.
More from Latent Space
View all 213 summaries
23 minSenior Dev: This "Grill Me" Prompt Is Going Viral Among Top Engineers
92 min🔬Top Black Holes Physicist: GPT5 can do Vibe Physics, here's what I found
61 min⚡️ Competing with ChatGPT and Sierra, building a $10M ARR company — Yasser Elsaid, Founder, Chatbase
75 minCI/CD Breaks at AI Speed: Tangle, Graphite Stacks, Pro-Model PR Review — Mikhail Parakhin, Shopify
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free