Two in Five Doctors Want Out. An On-Call Cardiologist Built the Note That Lets Them Stay

Abridge is seven years old, and Shiv Rao tells Sarah Guo and Elad Gil that the thesis has not moved. Doctors and nurses will not be fully automated in the next decade. The first signal in care delivery is a conversation—a dialogue between a professional and a patient—and those dialogues sit upstream of clerical work, orders, billing, and much else. Clerical work is the wedge. The company, which just closed a $250 million Series D, processes medical conversations so clinicians can stop drowning in paperwork and look at the person in the room.

The reason he started shows up as a set of ugly numbers. Two out of five doctors do not want to be doctors in the next two to three years. 27% of nurses, according to a JAMA article published last year, do not want to be nurses in the next 12 months. Patients drive five or six hours from rural settings to see a clinician who might save their life. Hiring another clinician costs close to a million dollars and takes a long time; hospitals have been shutting down because they cannot staff them. Rao calls it a public-health emergency. The whole industry, he says, finally agrees it has to assist, augment, or automate these workflows.

The product is meant to feel almost too simple. Walk in, hit Abridge, have a normal conversation in whatever order, hit stop, swivel the chair: the note is there. It is not the note Rao's 14-year-old daughter could generate from an off-the-shelf model. It has to clear who the clinician is, what specialty, which system, who the patient is, what insurance plan, what geography—and it has to be both a clinical note and a billable note. Gil asks for the difference. In this country, Rao says, doctors are not compensated for the care they deliver. They are compensated for the care they *documented* that they deliver. "Every single one of these notes is actually a bill."

Healthcare is not one market. On one end of Rao's spectrum is a direct-primary-care doctor taking cash off the insurance grid; in the middle, independent PCPs and small groups; on the other end, large health systems, integrated delivery networks, and academic medical centers. Most startups start small, hope to gather data and people, and swim upstream later. Years ago Abridge did the opposite: run into the large-system end. The barrier to "good enough" is extreme. You have to work for every doctor, every specialty, outpatient and inpatient, urgent care and the ER, and every spoken language. That is where the company's scientific muscle can actually show.

CTO Zach Lipton is a Carnegie Mellon professor, full-time at the company, and has recruited a research team that can reach deeper into the stack. The point of clearing that bar is competitive: at the hard end Abridge has faced pretty much one other company, while a crowd of startups fight over mid-market and individual primary-care doctors. Rao is not a tourist in the workflow. He went to Carnegie Mellon as an undergrad, became a cardiologist, then spent time as a corporate VC at UPMC—"I played VC. I was a faux VC"—funding startups and a machine-learning-and-health program where he met Lipton. Abridge is not a UPMC spinout. He quit to start it with people from CMU.

He still sees patients. The weekend before this taping he was on call in the hospital—about one weekend a month, plus Thursday nights on phone call for heart attacks. That "for clinicians, by clinicians" stance, he argues, is how they sequence product and go-to-market. They are live in Kaiser, Sutter, and over 110 health systems. When he was investing at UPMC, a handful of logos in a year meant champagne. The speed of these multi-year agreements, he says, is historic. The day before the episode, they had just announced Memorial Sloan Kettering.

Gil has watched health tech move a decade behind for years. What is different now? Rao sees two stars lining up. One is post-pandemic burnout that is no longer lip service: clinicians leaving, hospitals closing, replacement cost near a million dollars. The other is generative AI and ChatGPT in early 2023. Abridge started in 2018, three months after "Attention Is All You Need." Lipton would be quick to say the research community already knew transformers. The company's job was to interrogate pretrained models on clerical workflows. They published; Lipton's team won best paper at EMNLP in 2021. They shipped something that worked on BERT, BioBERT, Longformer, Pegasus.

The demos in 2021 and 2022 still landed as "cool story, bro." People put their hands up—*was that real?*—and said they would call in five years. Rao felt they were eating glass. Only in 2023 did he realize they had been pre-selling. When ChatGPT landed, CIOs and CMIOs called back: you were talking about generative AI at a dinner in 2022; we get it; let's pilot. Then the company YOLO'd the hard end of the market. WhatsApp groups among CMIOs and CIOs are daily and viral. Screw up one or two of those systems and you may not get another shot for years. They started with University of Kansas Health System, then Emory, then Yale. Home runs. Then it went viral at the enterprise-executive level.

Trust is the only currency that matters, and the stakes are core infrastructure. Rao had just come off a red-eye from HIMSS in Las Vegas. An executive there told him Abridge is now part of the health system's stack: if they go down, the system goes down, because the notes are bills and the system stops making money. Almost every deal is a three-to-four-week head-to-head with Microsoft. So far, Rao says, they have not lost one. In 2022 they demoed up and down Epic to earn an integration. The large competitor's offering at the time was humans in the loop: teams in Bangalore listening to audio, writing the note, and Wizard-of-Oz-ing it back into the record. That is why buyers kept asking if the demo was real. LLMs in late 2022 and 2023, he says, turned the game on.

The bottom of the stack is in-house speech recognition, a set of models for healthcare conversations. Outsiders say there is an API for that. Rao's answer is that a 3–5% error rate is a huge difference when you have to hear a new oral oncology drug the way *this* doctor mispronounces it. No clinician, he jokes, knows how to say these medications, and each has a private version. The system still has to catch symptoms, meds, diagnoses, and procedures across every specialty, and it has to keep up with the latest FDA approval.

The old world was dictation. After clinic, Rao would grab a Dictaphone and rattle: 25-year-old female, past medical history diabetes and hypertension, presents with shortness of breath, next line, capital B, past medical history, colon. Twenty or thirty of those in thirty minutes, off chicken-scratch paper. Sometimes the prompt to his own memory was "tall guy in the Mets hat." At night, in what doctors call pajama time, two patients with similar symptoms would smear together. Bad for doctors, terrible for patients, a disaster for billing. The new world is not a monologue. You have to lean into whatever the patient speaks.

Scale makes the language problem concrete. On a given day in California, Abridge will do at least 50,000 conversations in Vietnamese and Haitian Creole. In Boston, thousands in Brazilian Portuguese and Spanish. In Indiana, a doctor at Reid Health speaks Punjabi with her truck-driver patients. Whatever the language, the job is an English note in seconds, dropped into the discrete fields of the medical record for the clinician to trust and verify. The models even sample audio so a polyglot visit—hypothetically ten languages in one encounter—can still bias toward whatever is being spoken in the moment.

Above speech is text. One conversation has to be distilled three ways: what the doctor owes other clinicians, what the patient should see in an after-visit summary, and what revenue cycle needs, because these are bills. Burnout, in Rao's telling, is partly the tax of serving all those stakeholders at once and never focusing on the person you went to school to serve. Information-extraction models pull symptoms, medications, diagnoses, and procedures into data dictionaries. Summarization is style transfer. If Rao writes transcatheter aortic valvuloplasty in the chart and never said it out loud, the patient Googles it, panics, and blows up his inbox. Same dialogue: jargon for colleagues, plain language for patients, codes for billing.

If conversations are the original signal, the roadmap does not stop at notes. After a visit Rao might say "let's start metoprolol" or "let's get a CT." Those orders can be extracted, structured, and placed. After orders come claims, codes, and bills. Conversations also hide clinical-trial matches: inclusion and exclusion criteria for something that could save the person in the chair, surfaced at the point of care. The holy grail clinicians talk about is decision support. Imagine the system saying this patient looks like 10,000 recent California cases that turned out to be amyloidosis, not sarcoidosis, so skip the CT, go to cardiac MRI, and look at this *New England Journal of Medicine* paper. Rao used to think that was a three-year plan. They are building all of it now, across customers, at a scale of millions of conversations every couple of days.

Behind the note is what they call a contextual reasoning engine: not just the transcript, but the EHR problem list, insurance systems, clinical textbooks, orchestrated by agentic systems in the background. The artifacts still get edited. That is the point. Validated instruments show cognitive burden down about 60% within six weeks. A Stanford survey of burnout dropped about 50% for some clinicians in the first couple of months. Rao says no healthcare technology has had that kind of impact. The drafts save hours a day. The edits feed post-training—DPO, reward modeling, reinforcement learning. They do not claim perfection. They claim to get less imperfect, every day, and at this end of healthcare that margin matters.

Gil's digital-health company, a decade ago, taught him that healthcare adoption runs ten years behind the technology. Google's Med-PaLM 2 could already outperform physicians on some prediction tasks two or three years ago and barely got used. Why is the adoption curve only moving now? Rao keeps a two-by-two in his head. High-stakes, high-frequency workflows will not be absorbed quickly. Lower-stakes, high-frequency work with a clinician editing the draft is the open window: visible productivity, better experience, and a CFO conversation. Use a tool that did not orchestrate the models and the system loses money. Use Abridge and you get full credit for the care you delivered.

He is a heavy user himself. On his last call weekend he put ChatGPT and Claude on differentials and treatment plans. Sometimes they were 100% right on the first pass. Just as often it was a dialectic—three or four rounds, two hands clapping, the art being getting there together. Residents and medical students, he thinks, will pick this up faster than older attendings. When they mature, the game is on. When most of us get *really* sick, he still expects we will want a living doctor to parse the tools and set the plan.

On the hard end of the market the product will be imperfect forever, so Abridge threads several buyer personas plus end users. The CMIO speaks for doctors and nurses. The CIO worries about sunk cost, integration, too many apps, and the fact that nobody gets fired for Microsoft. The CFO wants tangible ROI. In early 2023 they knew they could not run the table. Two out of three—CMIO and CIO—was enough to get in. The CMIO bar was every specialty. A cardiology note looks nothing like oncology, and Sloan Kettering's notes look nothing like primary care or surgery. Emergency medicine is *The Pitt*: room one, room two, back to one, order an x-ray, room three. They had to stitch discontinuous conversations into one record, and they are now extending that stitch across the broader care team.

Some of the people who make that possible are what the company calls mutants: doctors who are also engineers. One was a principal engineer at Meta and a clinician. Some live in prompt engineering; some go deeper into the science; some sit in partner success and go-to-market. Interdisciplinary meetings happen inside one person's head, which lets the company skip translation steps. They have raised over $500 million. Rao wants about 80% of it to keep going into R&D—reaching lower in the stack and opening new workflows at the top.

The company started because a problem became hard to unsee. In March 2018, a University of Pittsburgh English professor with a ten-year history of breast cancer came for cardiology clearance before doxorubicin, a chemo that can injure heart muscle. She was crawling out of her skin. Her husband usually sat quietly in the corner and took notes so she could make eye contact, then they went home, rewrote the visit in words they understood, and showed up at the next specialist as main characters. He could not come that day. The patient story, Rao says, is agency. So is the clinician story. A *Journal of General Internal Medicine* paper last year estimated doctors need 30 hours a day to finish the work. You are always paying down a debt. That is pajama time. Abridge is trying to build the bridge between those two people.

Inside Slack there is a channel called Love Stories, programmatically filled with feedback from doctors nationwide. Criticism is oxygen: a gastroenterologist writing that he would never sequence assessment-and-plan that way, that it makes him look terrible. The other kind of note is "I'm not going to retire anymore." Rao's favorite comes from a physician at Tanner Health, a rural system. At dinner her son asked why she was not working. She pulled out her phone and told him a new tool let mommy come home early. She teared up. Her husband said, "Mommy's going to be able to eat dinner with us every night now." Hypergrowth dopamine, Rao says, gets you through sprints. Oxytocin—purpose—is what you are actually after.