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

The podcast starts with an introduction to Abridge, a company building a clinical intelligence layer for health systems. It emphasizes the importance of conversations in healthcare—the core workflow where care is given and received—and discusses how Abridge's technology aims to reduce the documentation burden on clinicians, who spend 10-20 hours weekly on paperwork, often working after hours ('pajama time'). The goal is to use AI to streamline clinical workflows, save time, reduce costs, and ultimately save lives.

Abridge initially focused on ambient documentation—helping clinicians by automatically generating notes from conversations. The platform has evolved into a broader clinical intelligence layer that offers decision support, pre-visit summaries, and real-time alerts. The expansion aims to assist clinicians before, during, and after patient interactions by integrating various data sources like medical literature, patient history, payer policies, and hospital guidelines, transforming healthcare delivery.

Chai, formerly at Glean, discusses parallels between search products and healthcare AI, emphasizing that 'context is king.' Unlike at Glean, healthcare's high safety and precision requirements mean AI models must be highly accurate and reliable, with rigorous evaluation processes. Unlike the horizontal scope of Glean, healthcare solutions are often more narrow and specialized, allowing for focused, specialty-specific development. The discussion highlights the importance of ambient, always-listening AI and proactive support in clinician workflows, contrasting it with traditional reactive alert systems.

The episode explores the challenge of alert fatigue in healthcare, advocating for proactive, context-aware AI that supports clinicians seamlessly and only interrupts when critical. It discusses multimodal form factors such as mobile, desktop, in-room devices, and emerging AR glasses that could deliver real-time information while maintaining clinician focus. Privacy and HIPAA compliance are crucial; Abridge employs de-identification models and strict data access controls. The conversation underscores that sophisticated AI solutions need to be deeply integrated with EHRs and workflows to ensure adoption and safety.

As AI models improve rapidly, the conversation addresses how Abridge leverages its unique data—specifically, hundreds of millions of clinical conversations—to train better models and create a competitive moat. Proprietary data allows for higher quality, lower latency, and lower costs. Operating at scale introduces challenges around infrastructure reliability, efficiency, and managing a vast amount of data, with a focus on optimizing tokens and computational resources.

The podcast emphasizes the importance of personalization at the individual, specialty, and organization levels to produce clinically accurate and style-consistent outputs. It discusses the complex evaluation process involving internal clinicians, third-party evaluators, and specialty-specific standards, which enables rapid iterative development. Trust with healthcare providers has allowed faster releases and more collaborative development, moving from months to weekly updates, building confidence and refinement.

The speakers note that healthcare's high stakes create rigorous regulatory guidelines, which serve as a moat but also as an accelerator for solving tough AI problems like zero-error and multi-step workflows. They observe that recent regulatory tailwinds, such as FDA guidance updates, support innovation while emphasizing safety. The discussion suggests that many advanced AI challenges—like real-time, low-latency support—must be addressed to make the impact real, pointing toward a future where AI facilitates a unified, conversation-driven healthcare system that benefits all stakeholders—patients, clinicians, payers, and pharma.

Chai elaborates on building event-driven, durable, real-time AI systems using tools like Kafka and Temporal, designed for scalability and multi-agent orchestration. The move toward more agentic, autonomous workflows involves advanced data manipulation and context layering, akin to the development of complex agent frameworks. They also highlight the importance of embedding AI into existing systems, such as EHRs, to support clinicians seamlessly without adding friction.

Both hosts and guests reflect on the importance of clear judgment, well-defined problem scope, and strategic focus over just prototyping. They emphasize that successful development depends on understanding the core problem, differentiating through unique data and models, and building for real-world complexity. Furthermore, they share how their backgrounds, particularly clinical expertise, profoundly influence product quality. They discuss the necessity of high-quality, high-confidence evaluation pipelines and the importance of deep partnerships with healthcare systems and regulatory compliance to build trustworthy AI tools.