A company most people outside healthcare have never heard of is quietly processing more real-world clinical conversations than almost anyone on the planet. Abridge, founded in 2018 (years before ChatGPT existed), is projected to support 80M+ patient-clinician conversations this year across 250 large U.S. health systems. If you run a healthcare operation, this is the clearest signal yet that ambient AI isn’t a pilot anymore. It’s infrastructure.
The Latent Space podcast just sat down with Abridge’s Janie Lee and Chai Asawa to walk through how they went from clinical note-taking to what they call a “clinical intelligence layer.” The details matter if you’re anywhere near healthcare delivery.
What happened
- Abridge started as clinical documentation software in 2018. It listens to doctor-patient conversations, generates the clinical note, and cuts the clerical burden so clinicians actually go home at night.
- The company has expanded well beyond note-taking into clinical decision support, prior authorization, and real-time alerts that fire during the patient visit, not days or weeks later.
- Abridge now supports 28+ languages and 50+ medical specialties.
- They raised $300M at a $5.3B valuation in June 2025, after a $250M round earlier that same year.
- Their pitch: the patient-clinician conversation is the most important workflow in healthcare. Almost everything (claims, payments, diagnoses, treatments) is a derivative of that conversation.
The numbers
- 80M+ patient-clinician conversations projected this year
- 250 large U.S. health systems using the platform
- 10 to 20 hours per week saved per clinician on documentation
- 28+ languages supported
- 50+ specialties covered
- $5.3B valuation after raising $300M in June 2025
- Product opened millions of times a week before, during, and after patient visits
5 things healthcare operators should take from Abridge’s playbook
- Start with the conversation, not the data warehouse. Abridge’s entire moat grew from one insight: the doctor-patient conversation is the source of truth for everything downstream. Billing, prior auth, quality measures, follow-ups. If you’re thinking about where AI fits in your clinical workflows, start where the information originates.
- “Pajama time” is a real metric, and it’s killing retention. Doctors spend 10 to 20 hours a week on documentation after hours, often at home in their pajamas. Abridge has clinicians saying the product saved their marriages. If your health system is bleeding physicians, documentation burden is the first thing to fix.
- Prior auth is ripe for real-time automation. Today, a patient gets an MRI denied weeks after the visit. Abridge is turning that into real-time guidance while the patient is still in the room. That single workflow change (from weeks to minutes) has massive implications for revenue cycle and patient outcomes.
- Alert fatigue is the enemy. Context is the cure. Chai described their approach as making AI feel like “air conditioning.” Always in the background, only interrupting when it truly matters. If you’re deploying any kind of clinical decision support, the bar isn’t “more alerts.” It’s fewer, better, contextual ones.
- Think of the EHR as a filesystem for AI agents. Abridge frames the electronic health record not as a static database but as a filesystem that agents can read from and write to. That mental model changes how you architect integrations. Deep EHR interoperability isn’t optional. It’s table stakes for clinician adoption.
The hot take
Healthcare isn’t lagging behind in AI. It might actually be ahead of most industries. While tech companies ship chatbots and call them agents, Abridge is running real-time decision support across 50+ specialties at a scale most SaaS companies dream about. The reason? Healthcare can’t do 80/20. You can’t ship a “good enough” diagnosis. That constraint forces teams to build better evals, better privacy infrastructure, and better human-in-the-loop systems than anyone in lower-stakes industries. The hardest problems produce the most durable solutions. Five years from now, the best practices for deploying AI agents in any enterprise will trace back to what healthcare companies figured out first.
The Agency OS play
If you run a health system or a company that sells to health systems, here’s what to do this week. Look at your prior authorization workflow end to end. Map every step where a human is waiting on information that already exists somewhere in your EHR or payer policy documents. That gap between “information exists” and “the right person sees it at the right time” is where automation delivers immediate ROI. You don’t need to build an ambient scribe. You need to build the connective tissue between your existing data sources and the clinical moment.
Start small. Pick one specialty, one workflow (prior auth is the obvious candidate), and one EHR integration point. Build a prototype that pulls payer policy, patient history, and visit context into a single view at the point of care. Test it with five physicians. Measure time saved in hours per week, not in abstract “efficiency” scores. Abridge’s success came from obsessing over clinician adoption, not feature lists. Your pilot should do the same.
Finally, invest in your eval infrastructure now. Abridge uses LLM judges, in-house clinicians, third-party evaluators, and specialty-specific evaluation pipelines with progressive rollouts. If you’re deploying anything that touches clinical decisions, you need a similar layered approach. Don’t ship a model and hope. Build the feedback loop from day one. Get clinicians reviewing outputs before you scale. Healthcare AI that works at 100 conversations will break at 100,000 unless you’ve built the quality system to catch drift early.