Shiv Rao and Abridge: Building Clinical AI Around the Draft Note
On this page 11 sections
Shiv Rao built Abridge around a narrow but costly problem: clinicians spend part of a medical visit documenting the conversation and more time afterward turning it into a usable record. Abridge listens with patient consent, creates a transcript, and drafts structured clinical documentation for a clinician to review. It does not replace the clinician, make the diagnosis, or sign the note.
That boundary explains both the company’s adoption and its risk. A draft note fits into an existing workflow and can reduce clerical work. Yet a fluent draft can omit context, misstate certainty, or add text a clinician would not have written. A health system therefore has to evaluate the product as a clinical workflow component, not as a generic meeting recorder.
The short answer
Rao’s advantage was not simply that he was a physician who founded an AI company. His public record connects three kinds of work: practicing cardiology at UPMC, investing through UPMC Enterprises, and building a product for medical conversations. Abridge’s own biography identifies him as founder and CEO and a practicing UPMC cardiologist. It does not establish that every design decision came from his clinical practice or that his background guarantees product quality.
Abridge won large health-system deployments because it packaged speech recognition, note generation, source traceability, and electronic health record integration into a reviewable workflow. The strongest public evidence suggests ambient documentation can reduce perceived burden and burnout. It is not yet evidence that the category improves diagnoses, patient outcomes, or total cost of care.
What Rao actually built
Abridge began with a patient-facing idea: record a medical conversation and make it easier to revisit what was said. Its 2020 launch announcement described a service that helped people record, review, and share details from care, and said the company had raised $15 million to date. That is a company disclosure from the period, not an audited history of the business.
The enterprise product moved the economic buyer from the patient to the health system. During a visit, the product captures audio, processes the conversation, drafts a note, and lets a clinician review the output before it enters the record. This matters because the product does not need to persuade a hospital to delegate clinical judgment. It asks the hospital to change who produces the first draft.
The operating model can be summarized in five controls:
| Stage | System task | Required human or institutional control |
|---|---|---|
| Consent | Start ambient capture | Explain recording and honor refusal rules |
| Capture | Transcribe the encounter | Check language, speaker, and audio limitations |
| Draft | Generate structured documentation | Keep output clearly marked as a draft |
| Review | Edit and approve the note | Assign the licensed clinician final responsibility |
| Monitor | Track quality after deployment | Audit errors, overrides, complaints, and specialty variance |
Kaiser Permanente proved distribution, not autonomy
Kaiser Permanente’s August 2024 announcement is the clearest scale marker. The system made Abridge available across 40 hospitals and more than 600 medical offices. Kaiser called it assisted clinical documentation and said it had tested the system before wider availability.
The announcement did not say that Abridge could act without review. Kaiser later stated in its 2024 annual report that AI does not make medical decisions and that doctors and care teams do. That distinction should follow the product into procurement documents, training, interface copy, and incident reviews.
Coverage described the rollout as reaching more than 24,000 physicians. The official release instead names facilities and availability to clinicians. Those measures are related but not identical: eligibility, activation, regular use, and successful use are four different adoption metrics. A buyer should ask for each one.
Evidence on burden is promising but bounded
A 2025 multicenter quality-improvement study in JAMA Network Open evaluated Abridge use across six US health systems. After 30 days, reported burnout fell from 51.9% to 38.8% among participating clinicians, while measures of cognitive task load and documentation burden also improved. The authors described the findings as associations, not proof from a randomized controlled trial.
That caveat is important. Clinicians who adopt a new tool may differ from those who do not. Survey respondents know they are using the intervention. A short follow-up can show immediate relief without resolving whether the effect persists or whether documentation quality changes over time.
A separate JAMIA Open study at UPMC surveyed clinicians after implementation. Of 181 invited clinicians, 133 had used Abridge for at least five encounters, and the authors reported improvements in perceived documentation burden and job satisfaction. The study also identified its adopters and survey design, which makes it useful operational evidence but not a universal estimate for every specialty or health system.
Note quality is the harder test
Burnout and time savings answer whether clinicians value the workflow. They do not fully answer whether the generated note is accurate, concise, clinically appropriate, or better than the previous note.
Research published in 2026 on clinician edits to Abridge drafts found recurring reasons for intervention, including factual correction, specialty-specific refinement, diagnostic-certainty calibration, restructuring, and removal of transcript-like material. The JAMIA analysis is useful because edits are not merely failures. They reveal where clinicians apply judgment that the model does not reliably supply.
A buyer should evaluate at least four error classes:
- Unsupported content that was not said or clinically established.
- Omission of a material symptom, decision, instruction, or caveat.
- Incorrect attribution of a statement to the patient or clinician.
- Overstatement of diagnostic certainty in assessment and plan language.
Average note-quality scores can hide a rare but serious event. Review programs need severity-weighted error reporting, not only a single accuracy percentage.
Funding measures investor confidence, not clinical value
Abridge announced a $150 million Series C in February 2024. That company disclosure describes a private financing event and investor expectations. It does not show profitability, renewal rates, clinical effectiveness, or market share. Later private valuation estimates are not used here as operating evidence.
The previous version of this article treated valuation and an unsupported share estimate as proof of an “empire.” That framing has been removed. There is no public audited dataset that establishes a stable market-share ranking for ambient clinical documentation across comparable deployments.
Abridge’s moat is workflow evidence
Speech-to-text and summarization models are increasingly available from multiple suppliers. Abridge’s defensibility therefore depends less on access to a generic language model and more on deployment work that is difficult to reproduce quickly:
- integration into the health system’s existing record and identity controls;
- evaluation across specialties, accents, languages, visit types, and noisy environments;
- traceability from draft text back to source conversation;
- configuration that matches local templates and documentation policy;
- implementation teams that can support training, escalation, and monitoring.
Competitors can build similar features. Microsoft, Nabla, Suki, Ambience, and health systems’ internal teams all contest parts of the category. A buyer should compare evidence for its own workflow rather than infer quality from customer logos or financing.
Privacy starts before the model runs
Ambient documentation captures one of the most sensitive conversations a person can have. A health system needs to know where audio and transcripts are processed, how long each artifact is retained, which subprocessors receive it, whether data is used for model improvement, and how a patient can decline without disrupting care.
Kaiser said its implementation processes and encrypts data and complies with applicable privacy laws. That is an institutional disclosure, not a substitute for a buyer’s legal analysis or technical verification. Contracts should specify retention, deletion, access logging, breach notification, secondary use, and the treatment of data generated when a recording starts accidentally or captures a third party.
A practical health-system scorecard
The decision should be made on local evidence. A useful pilot compares the new workflow against an established baseline for the same specialties and visit types.
| Decision area | Evidence to collect | Stop or revise trigger |
|---|---|---|
| Clinical fidelity | Blinded review of draft and final notes | Material unsupported statements |
| Workload | Time in notes, after-hours EHR time, edit burden | Savings disappear after training period |
| Adoption | Eligible, activated, weekly active, retained users | High activation with low continued use |
| Equity | Performance by language, accent, disability, and specialty | Meaningful subgroup degradation |
| Privacy | Data-flow map, retention tests, access logs | Unapproved reuse or unclear deletion |
| Patient experience | Consent acceptance, complaints, visit surveys | Patients feel unable to refuse |
| Economics | Total license, integration, support, and review cost | Cost rises faster than verified benefit |
The scorecard should be versioned. Model updates, new specialty templates, or changed retention policies can invalidate an earlier approval.
Public record gaps
Public sources do not disclose Abridge’s audited revenue, gross margin, customer retention, or deployment-level error rates. They do not allow a reliable comparison of active users across vendors. Funding announcements and customer press releases also cannot establish that the product caused better patient outcomes.
Rao’s medical and investment background helps explain why Abridge focused on a deployable clinical workflow. It cannot stand in for those missing measures.
The durable lesson from Rao’s strategy
Abridge narrowed the initial promise of generative AI to a task with a visible handoff: the system drafts, and the clinician reviews. That made the product easier to evaluate and easier to integrate than an autonomous clinical agent.
The next test is less photogenic than a large rollout. Health systems must keep measuring what clinicians correct, what the model misses, who benefits, and what happens after a model update. If those records remain available to buyers and care teams, ambient documentation can mature as clinical infrastructure. If they do not, a fluent note will remain an uncertain proxy for a better visit.