Ask anyone outside medicine what AI is doing to healthcare and you’ll hear about diagnosis. The algorithm that catches the tumor a radiologist missed. The model that reads a chart and names the disease nobody guessed. That story has collected a decade of magazine covers, keynote slides, and funding rounds. Still, it remains mostly a story.
Meanwhile, something far less photogenic went into production. Software now sits in exam rooms at hundreds of health systems, listens to the conversation between a doctor and a patient, and drafts the clinical note. It diagnoses nothing. Prediction isn’t part of the job either. Instead, the software transcribes and summarizes, which happens to be the one thing current language models do reliably enough to ship.
Ambient documentation is the least glamorous AI product in medicine. Yet it’s the one that stuck. Map the full range of AI use cases in healthcare and the same split turns up in every department: the clinical showpieces stall in pilot, while the operational tools go live and stay live.
What Does FDA Clearance Data Actually Show About Diagnostic AI?
Start with what regulators have signed off on. The FDA’s public list of AI-enabled medical devices reached 1,451 authorizations by the end of 2025, counting back to 1995. Notably, radiology accounts for 1,104 of them, roughly 76% of everything cleared. On a slide, that looks like proof that diagnostic AI won.
But read what those devices do and the picture flips. The overwhelming majority are narrow image-processing tools. Measure this lesion. Flag this scan for earlier review. Quantify calcium in this artery. Segment this organ. Even then, each one assists a radiologist who still reads the study and signs the report. Almost none render a diagnosis alone, and so far the FDA has authorized no general-purpose system that takes an undifferentiated patient and returns an answer.
Of course, narrow does not mean trivial. AI reading a routine ECG catches structural heart disease that a cardiologist’s eye misses on the same tracing, and that work matters. Even so, a tool that spots one signal in one clean data stream sits a long way from the machine that thinks like a clinician.
Why a Cleared Device Isn’t a Deployed One
A second gap rarely makes the coverage. After all, a clearance is permission to sell, not evidence of use. Getting a device cleared, getting a hospital to buy it, wiring it into a radiology workflow, and getting anyone to bill for it are four separate problems. The last three kill more products than the first. So a market with 1,104 cleared radiology tools and no comparable count of daily deployments tells you exactly where the friction lives.
The concentration in radiology is itself a clue, though people misread it constantly. Imaging dominates the list because the input arrives as a clean digital file with a defined format, a labeled archive going back decades, and a regulatory pathway that predates machine learning. Those are logistics advantages. They are not intelligence advantages. Pick any specialty where the raw signal is a messy conversation, a physical exam, or a patient’s own account of what hurts, and the clearance count collapses toward zero.
So when a headline announces that AI now diagnoses disease, the accurate translation usually runs narrower: a model flags a suspicious region on a CT scan so a human looks at it sooner. Genuinely useful. Also a triage improvement rather than a replacement for clinical reasoning. Treating those two things as the same thing is how a decade of expectations went wrong.
Why Documentation, Not Diagnosis, Was the Real Bottleneck
While the diagnosis conversation ran, doctors drowned in paperwork. A time-motion study funded by the American Medical Association found that for every hour of direct face time with patients, physicians spend nearly two additional hours on EHR and desk work during the clinic day. The physicians who kept after-hours diaries logged another one to two hours most nights, nearly all of it charting.
So that’s the actual bottleneck in outpatient medicine. Not diagnostic accuracy. Typing.
Ambient scribes aim straight at it, and they have company. Look at what health systems put into daily production rather than pilots and a consistent shape appears:
- Appointment scheduling and no-show prediction
- Prior-authorization drafting
- Medical coding support
- Symptom-based triage routing
- Draft replies to the patient messages that flooded in after portals went mainstream
Every one of those jobs is administrative. In fact, clinics outside medicine reached the same conclusion from the other direction, which is why conversational agents now answer the calls nobody picks up at front desks across dozens of industries.
The pattern is hard to miss once you look for it. Applications that scaled share two traits. A human was already going to review the output as part of the job. And a mistake costs a minute of editing rather than a malpractice claim.
What an Ambient AI Scribe Actually Does
The name oversells it, so precision helps here.
First, a microphone captures the visit. A speech model transcribes it, including the crosstalk, the interruptions, and the part where the patient describes symptoms in their own words. Then a language model reorganizes that transcript into the structure a clinical note requires: history, exam findings, assessment, plan. Finally, the physician reads the draft, fixes what’s wrong, adds what the conversation didn’t cover, and signs it.
That second model does most of the commercial work. After all, transcription alone stopped being the hard part some time ago, and a raw transcript of a fifteen-minute visit helps nobody. Turning it into a structured note does.
Nothing in that chain requires the model to know medicine. It requires the model to know English and the shape of a clinical note.
How Kaiser Permanente Ran Ambient AI at Population Scale
Pilots prove nothing. Volume does.
The Permanente Medical Group turned on ambient AI documentation for about 10,000 physicians and staff in October 2023. Within ten weeks, 3,442 physicians had used it across 303,266 patient encounters, spanning a wide range of specialties and sites, with results reported in NEJM Catalyst. In fact, nearly a thousand of those physicians used it in more than 100 encounters apiece. One crossed 1,210.
That last detail matters more than the headline number, because sustained repeat use by individual clinicians separates a tool people tried from a tool people rely on.
Still, the longer view held up. Across roughly 15 months running to December 2024, more than 7,000 Permanente physicians used ambient AI in about 2.5 million encounters, and the group reported saving close to 16,000 hours of documentation time. No single doctor experienced a dramatic transformation. Several thousand of them each got a few minutes back per visit, and the arithmetic did the rest.
Now name a diagnostic AI deployment at comparable scale in the same window. There isn’t one.
Why the Rollout Method Decided the Outcome
The mechanics deserve attention too, because the unsexy part drove the result. Kaiser skipped the two-site pilot and the accompanying paper. Instead, it enabled the tool broadly, let physicians opt in, tracked who kept using it, and studied the ones who dropped off. That structure surfaces the real adoption curve rather than a curated best case, and it produces the one metric enterprise software lives or dies on: retention. A tool that thousands of physicians tried and hundreds used in triple-digit volumes tells you something a controlled pilot never can.
Why Did the Boring Task Win?
Better engineering had nothing to do with it. Rather, ambient scribes shipped because someone picked a better task. Five properties made it winnable.
The reviewer was already in the room. A physician who just conducted the visit validates a draft note in seconds, because the ground truth still sits in their head. No separate QA process runs afterward. Nobody queues the note for expert review. A second opinion never enters the picture.
Errors surface immediately and cost almost nothing. A garbled sentence in a draft gets deleted. Meanwhile, a wrong diagnosis propagates into treatment decisions and can take months to surface.
The regulatory question mostly evaporates. A draft that a licensed physician edits and signs becomes that physician’s note. The AI never takes the decision, so it never becomes a medical device the way a diagnostic model does.
The payoff lands in a metric leadership already tracks. For example, time in notes and after-hours EHR minutes already sit in every large system’s Epic dashboards. Nobody had to invent a measurement framework to prove value.
Nothing about the visit had to change. No new hardware in the room, no extra step in the workflow, no retraining on a different clinical protocol. As a result, the doctor talks to the patient the way they always did.
Why Diagnostic AI Fails the Same Test
By contrast, diagnostic AI misses on all five counts. Verification runs expensive and slow, errors turn catastrophic and delayed, the regulatory burden is enormous, the benefit is statistical rather than immediate, and the workflow has to bend around the tool. Same underlying models. Wildly different odds of reaching production.
Above all, the second property carries the weight. Cheap, fast error correction is what lets an imperfect model earn its keep, and it’s the variable most AI strategy decks ignore entirely. Picture a scribe that gets some fraction of every note wrong. Still a large net win, because fixing part of a draft beats writing all of it from scratch. Hand a diagnostic model that same error rate and it becomes unusable at any price. No accuracy improvement changes the category it falls into. The threshold isn’t a percentage. It’s whether the person holding the output can recover from a wrong answer on the spot.
Where Ambient Scribes Still Fall Short
None of this makes ambient documentation a finished success, and the data pushes back in useful ways.
First, physician burnout is falling. The AMA’s national survey, drawing on nearly 18,000 responses in 2024, found 43.2% of physicians reporting at least one symptom of burnout, down from 48.2% in 2023 and 53% in 2022. Real movement. Then again, that’s a three-year trend across an entire profession, so pinning it on scribes specifically would be sloppy. The AMA’s own reporting notes that after-hours charting has barely budged even as burnout eased.
How Often Do Doctors Actually Use the Tool?
Adoption inside the studies looks messier than vendor decks suggest. A three-arm randomized trial ran 238 outpatient physicians across 14 specialties from November 2024 through January 2025, assigning them to Microsoft’s DAX Copilot, Nabla, or usual care. The DAX arm used the tool in 33.5% of 24,696 visits. Similarly, the Nabla arm reached 29.5% of 23,653. Physicians handed a working scribe reached for it in roughly a third of their appointments. Whatever limits uptake, it isn’t availability.
Beneath all of that sits a structural worry. A policy brief in npj Digital Medicine argued that scribes drafting notes also shape billing documentation, which creates pressure toward more thoroughly coded, higher-reimbursing notes without any change in the care delivered. Either way, a documentation tool that quietly turns into a revenue tool is worth watching before it becomes the norm.
There’s a broader version of that risk too. Across industries, plenty of deployments shifted labor instead of removing it, and workers now spend their days cleaning up what the model produced. Scribes avoid that trap only as long as editing a draft stays faster than writing one.
Real technology, real limits, real open questions. In short, that combination is what a working deployment looks like, and it beats a decade of promising demos.
How to Apply the Same Test to Your Own AI Roadmap
Healthcare is the extreme case, but the selection rule travels.
Ask three questions about any candidate use case:
- Is the person who checks the output already required to be there?
- Does a mistake show up in seconds or in months?
- Are you removing work nobody wanted, or replacing judgment someone gets paid for?
Ambient scribes answer well on all three. Diagnostic AI answers badly on all three, which is why it has been five years away for fifteen years.
The instinct is to point AI at the hardest, highest-value problem in the business. Medicine tried that and got pilots. Instead, the wins came from pointing it at the tedious, verifiable task sitting next to the hard problem, then letting the expert keep doing the hard part with more of their day back.
Now run the test against your own list. Legal review, underwriting decisions, and clinical judgment all sit in the diagnosis category: the expert’s call is the product, verification is slow, and a wrong answer travels. By comparison, contract summarization, meeting notes, first-draft support replies, data extraction from unstructured documents, and code review comments sit in the scribe category, because someone competent reads the output within minutes anyway and the failure mode is a rewrite.
That second list is less exciting to pitch, which is exactly why it survives contact with actual users. Meanwhile, the first list produces impressive demos and stalled procurement. If your roadmap leans heavy on the first list and light on the second, healthcare already ran that experiment at a scale you can learn from for free.
Key Takeaways
Diagnosis was never the near-term prize in healthcare AI. Documentation was, and the deployment numbers now say so plainly.
Three things worth carrying out of this:
- Judge AI progress by sustained production usage, not regulatory clearances or benchmark scores.
- Pick tasks where the verifier already sits in the loop and errors show up immediately.
- Expect the durable wins to look unimpressive in a demo. Ultimately, giving thousands of people a few minutes back never photographs as well as a machine that appears to think.
Related: How AI Is Predicting Falls in Senior Care Before They Happen
