Summary

Abridge’s core thesis is that conversations—not autonomous clinicians—will remain healthcare delivery’s first signal over the next decade. Those dialogues sit upstream of documentation, orders, billing, trials, and eventually decision support, making clerical automation a wedge into broader workflows. Rao’s goal is to remove work that “crushes their souls at night” while keeping clinicians in the loop.
The company deliberately entered large health systems, where the quality barrier creates both defensibility and concentrated distribution. Rao estimates Abridge is live in more than 110 systems, including Kaiser and Sutter; he says it has never lost a three-to-four-week head-to-head against Microsoft in recent years. Success at the University of Kansas Health System, Emory, and Yale then spread through CIO and CMIO networks—enterprise virality with severe downside because “you don’t get another shot on goal” after a failure.
Burnout created urgency, while ChatGPT converted years of market education into demand. Two out of five doctors reportedly may leave medicine within two to three years, and 27% of nurses within 12 months; Rao estimates replacing a clinician can cost close to $1 million. Abridge spent 2021–22 “eating glass” through demos, but Rao later realized it had been “pre-selling”: after ChatGPT arrived in early 2023, health-system executives called back asking for pilots.
The technical moat is not generic transcription but healthcare-specific recognition, orchestration, and audience-aware output. Rao argues that even 3–5% speech-recognition error rates can matter when doctors pronounce new oncology drugs idiosyncratically. Abridge must handle multilingual, polyglot conversations, then generate within seconds an English clinical note, patient summary, structured fields, and documentation sufficient to get “full credit for the care that you delivered.”
Scale turns clinician edits into a post-training flywheel. Abridge processes millions of conversations every couple of days, combining dialogue with medical records, insurance systems, and clinical textbooks through a “contextual reasoning engine.” Because its drafts remain imperfect, edits feed preference tuning, DPO, reward modeling, and reinforcement learning—the objective, in Rao’s estimation, is candidly to become “less imperfect,” not claim perfection.
Early measured outcomes are unusually strong, but the adoption wedge depends on a clinician remaining in the loop. Rao cites roughly 60% lower cognitive burden within six weeks and sometimes 50% lower burnout within the first few months. His framework favors lower-stakes, high-frequency workflows: they can prove productivity and ROI while clinicians verify drafts, unlike high-stakes autonomous care that health systems may absorb much more slowly.
The next prize is point-of-care intelligence, but Rao still expects seriously ill patients to want a live doctor using these tools. Abridge could surface trial eligibility or compare a patient with 10,000 similar recent cases, suggesting amyloidosis rather than sarcoidosis and a cardiac MRI rather than a CT. Yet Rao’s own use of GPT and Claude was sometimes immediately correct, but maybe just as often became a “dialectical experience” requiring three or four exchanges before reaching the right plan.

Deep dive

1. The clinical conversation is the wedge into healthcare’s operating system

Abridge began in 2018 with a thesis that has not changed: doctors and nurses will not be fully automated over the next decade, and the dialogue between professional and patient is healthcare delivery’s “first signal.” Because conversations precede so many workflows, automating clerical work can open paths into much more.
The immediate problem is capacity. Rao cites two out of five doctors considering leaving within two to three years and 27% of nurses within 12 months, while rural patients may drive five or six hours for lifesaving care. Burnout is no longer “lip service”; some hospitals have shut down because they could not staff themselves.
Abridge lets a clinician hold a normal conversation and receive a draft note within seconds. But the artifact must reflect the clinician, specialty, health system, patient, insurer, and geography: in the US, Rao says, “we’re not compensated…for the care that we deliver. We’re compensated for the care that we documented,” so every note is also a bill.

2. Running toward enterprise complexity created distribution and defensibility

Rao chose large health systems over independent practices because “the barrier to good enough” is much higher: one product must serve every specialty, inpatient and outpatient settings, urgent care, emergency departments, and multiple spoken languages. That demanded deeper science and, Rao said, let Abridge compete with pretty much one other company.
The timing required two aligned stars. Post-pandemic staffing pressure made clinician experience economically urgent; then ChatGPT made generative AI legible to buyers. After people dismissed 2021–22 demos as “cool story, bro,” health-system executives called back in 2023 saying, “I get it now. Let’s try it.”
Abridge “YOLO’d it” by starting with major academic systems, where CIOs and CMIOs constantly compare notes. Home runs at the University of Kansas Health System, Emory, and Yale produced executive-level virality; a failure at one or two institutions, Rao believed, could close the market for years.
Trust is “the only currency that ends up mattering in healthcare.” Abridge built relationships with ecosystem players such as Epic. An executive told Rao Abridge was now core infrastructure; Rao says that if it went down, the health system would go down and stop making money because the notes are essentially bills. Microsoft is the usual competitor, and Rao says Abridge has not lost a recent head-to-head.

3. Medical speech remains an unsolved, high-stakes systems problem

Elad’s challenge—voice looks solved because APIs exist—draws Rao’s key distinction: 3–5% error rates matter around symptoms, procedures, and newly approved drugs. The system must recognize each clinician’s peculiar pronunciation across specialties and continually update for new medical vocabulary.
Traditional dictation was a lossy monologue assembled later from “chicken scratch” during doctors’ “pajama time.” Rao might write “tall guy in the Mets hat” and hope it reconstructed the encounter, risking details from patients with similar symptoms blending together.
Ambient conversation instead requires multilingual recognition and polyglot conversations. Rao says Abridge may process at least 50,000 conversations in Vietnamese and Haitian Creole in California in a day, thousands in Brazilian Portuguese and Spanish in Boston, and Punjabi conversations with truck drivers in Indiana—then generate an English note in seconds.
Downstream models extract symptoms, medications, diagnoses, and procedures; map them to dictionaries; and create different artifacts for clinicians, patients, and revenue-cycle teams. Style transfer matters: a patient-facing summary should not suddenly introduce an unexplained phrase such as “transcatheter aortic valvuloplasty.”

4. Documentation expands naturally into orders, trials, and decisions

Once the conversation is treated as the upstream signal, the roadmap broadens. “Let’s start you on metoprolol” or “Let’s get a CT scan” can become structured orders in the medical record; those orders then lead to codes, claims, and insurance billing.
Trial matching is another adjacent workflow. A clinician may not realize that the patient in front of them satisfies inclusion and exclusion criteria for a potentially lifesaving study; Rao wants technology to surface that fact at the point of care, with enough information to discuss it immediately.
Clinical decision support is the “real holy grail.” Rao imagines Abridge identifying that Sarah resembles 10,000 recent California patients, for whom clinicians favored amyloidosis over sarcoidosis—then recommending a cardiac MRI rather than “screw around with the CT scan,” while surfacing a relevant New England Journal of Medicine study.

5. Deployment scale makes every correction a training asset

Abridge now handles millions of conversations every couple of days. Its contextual reasoning engine combines the dialogue with problem lists, medical history, insurance information, and clinical textbooks, orchestrating those sources “in the right way, in the right order” to produce the best available draft.
Rao explicitly rejects perfection claims: clinicians still edit drafts, but save hours daily. Those edits power preference tuning, DPO, reward modeling, and reinforcement learning, creating a feedback loop whose practical aim, in Rao’s estimation, is to become continuously “less imperfect.”
The reported user metrics are material: validated instruments show about a 60% reduction in cognitive burden within six weeks, while one Stanford survey sometimes shows roughly 50% lower burnout in the first couple of months. Rao argues no prior healthcare technology has produced that kind of impact.

6. Human verification unlocks adoption before autonomous care is ready

Rao’s adoption matrix separates stakes from frequency. High-stakes, high-frequency automation will enter healthcare slowly; documentation is comparatively lower-stakes and high-frequency because a clinician reviews the output. That open window widens when vendors can prove productivity, physician experience, patient experience, and CFO-level revenue capture.
His own weekend call shift illustrated the current boundary. GPT and Claude were sometimes correct immediately, but maybe just as often the process required three or four rounds—a “dialectical experience” in which the art was “getting there together with it.” Trainees may learn this collaboration faster than older attending physicians.
Minimum viable quality also differs by buyer. In early 2023, Abridge could satisfy the CMIO’s specialty requirements and the CIO’s integration concerns, though not yet fully run the table for the CFO; “two out of three” was enough to enter.
Workflow fit is specialty-specific. Emergency clinicians move repeatedly between rooms, so Abridge had to stitch discontinuous conversations into one encounter; it is now extending that concept across the broader care team. Oncology, cardiology, surgery, and primary care each demand different structure, content, and stylistic preferences.

7. Clinician-builders connect the product’s economics to its purpose

Abridge recently secured a $250 million Series D and has raised more than $500 million overall. Rao remains a practicing cardiologist, and Abridge has “mutants”—doctors who are also engineers, prompt specialists, scientists, or go-to-market operators—because they can hold interdisciplinary meetings “in their own mind” and skip translation steps. Rao says 80% of capital should continue going into R&D.
The founding patient story centered on agency. A woman with a 10-year history of breast cancer had relied on her husband to take notes, letting her remain present and later unpack the visit in language they understood—so they could feel “like the main characters as opposed to someone looking in from the outside.”
Clinicians face the same loss of control: Rao cites research estimating doctors need 30 hours per day to complete their work. Abridge’s mission is to bridge patient and professional agency by returning attention to the encounter and reducing the documentation debt paid after dinner.
His sharpest outcome came from a rural Tanner Health doctor whose son asked why she was not working right then. After she explained Abridge, her husband said, “Mommy’s gonna be able to eat dinner with us every night now.” Rao distinguishes hypergrowth’s sprint-oriented “dopamine hits” from these “oxytocin hits”—purpose and fulfillment that explain why the company is working so hard.

Code Context is King: Augment’s AI Assistant for Professional Software Engineers, with Guy Gur-Ari 30. The Revival of Chinese Stocks and Alibaba’s AI-Stock Story: China Assets Enter a Long-Awaited Era of Confidence