Abridge has processed over 100 million clinical conversations. Every one of those is a doctor-patient interaction - diagnosis discussions, treatment plans, symptoms described in plain language - turned into structured medical documentation.
Co-founders Janie Lee and Chai Asawa walked through how they built clinical intelligence infrastructure that works at healthcare scale. This isn't a prototype. It's live in hospital systems, generating notes that doctors actually use, automating prior authorisations that used to take days, and doing it under privacy regulations that would kill most AI products before they launched.
The Latent Space interview covers the hard parts: evaluation pipelines for medical accuracy, integration with Electronic Health Record systems, and why healthcare's high-stakes environment forces you to solve AI problems most companies can ignore.
The Evals Problem in Healthcare
Standard LLM evaluation doesn't work for clinical notes. You can't just measure BLEU scores or ask GPT-4 to judge output quality. Medical documentation has regulatory requirements. A missed detail isn't a bad user experience - it's a liability risk and a patient safety issue.
Abridge built evaluation systems that check for clinical accuracy, completeness, and compliance. Every generated note is validated against structured criteria: Are diagnoses captured correctly? Are medication dosages precise? Is the documentation audit-ready?
They use a mix of rule-based checks, specialist review, and LLM-assisted validation - but the final call on accuracy still involves human clinicians. That's expensive and slow, but it's the only way to meet healthcare's evidence bar. The alternative is liability exposure that no hospital will accept.
EHR Integration is the Unglamorous Hard Part
Electronic Health Records are legacy systems built in the 1990s, running on infrastructure nobody wants to touch. They don't have modern APIs. They don't play well with external tools. And they contain the most sensitive data in healthcare, so access control is extreme.
Getting an AI system to write directly into an EHR means navigating vendor-specific integrations, hospital IT approval processes, and compliance audits that take months. Abridge's engineering team spends significant time on problems that have nothing to do with AI - data formatting, authentication protocols, fail-safe mechanisms for network outages.
This is why most healthcare AI stays in research labs. The infrastructure work required to deploy at hospital scale is brutal. But it's also why companies that solve it have a moat. Once you're integrated into a hospital's workflow, replacing you means ripping out systems that doctors depend on daily.
Prior Authorisation in Minutes, Not Days
Prior authorisation is the process where insurance companies require approval before covering certain treatments. It used to involve manual paperwork, phone calls, and wait times measured in days. Abridge automated it.
The system pulls relevant patient history from the EHR, structures it into the format insurers require, and submits the request - all within minutes of a doctor deciding on a treatment plan. The time savings are real. Doctors get 10-20 hours back per week. Patients get treatments faster. Hospitals reduce administrative overhead.
This is where AI's impact in healthcare becomes tangible. Not in diagnostics that might replace doctors someday. In automation that removes friction from workflows that shouldn't require human effort in the first place.
Why Healthcare Drives Harder AI Problems
Lee and Asawa argue that healthcare forces you to solve problems that other industries can afford to ignore. Privacy regulations mean you can't just send data to OpenAI's API - you need on-premise models or secure cloud deployments. Liability risk means you can't ship "good enough" outputs - you need provable accuracy. Integration constraints mean you can't build a standalone app - you need to fit into existing workflows.
These constraints are frustrating, but they produce better systems. An AI that works in healthcare is robust, auditable, and defensible in ways that consumer AI rarely needs to be. The lessons learned building for hospitals apply everywhere else.
The other side of this: healthcare's willingness to pay for solutions that work. Hospitals have budget for tools that save clinician time, reduce liability, and improve patient outcomes. If you can prove your AI does those things, the economics work. That's rare in AI product development - most companies are still figuring out what people will pay for.
At 100 million conversations processed, Abridge isn't a research project anymore. They're infrastructure. And the next 100 million conversations will likely come faster than the first.