Where AI in healthcare actually works

Adoption, first

The AMA's survey of 1,692 physicians, fielded in early 2026, found 81% reporting awareness or use of AI professionally — up from 66% in 2024 and 38% in 2023, though the earlier figures counted incorporation rather than awareness, so the trend line is not quite like-for-like. The average physician now reports 2.3 use cases, against 1.1 three years earlier. (AMA, 2026)

What they use it for is informative. Summarising medical research leads at 39%. Then generating discharge instructions and care plans (30%), documenting billing codes and visit notes (28%), generating chart summaries (28%), and drafting patient portal replies (19%).

Note the shape: every leading use is a text transformation task. Nobody is topping the list with diagnosis.

Task by task, with the evidence

Ambient documentation — works, more modestly than advertised

Multiple studies show substantial improvements in burnout and cognitive load, and consistent but modest reductions in documentation time. The largest and best-controlled study — 8,581 clinicians across five health systems and three products — found 13.4 fewer minutes of total EHR time and 16.0 fewer minutes of documentation time per eight scheduled hours, and no significant change in after-hours EHR use. (JAMA, 2026) Heavy users got roughly double.

A three-arm randomised comparison of two commercial products found one reduced documentation time by 9.5% and the other by 1.7% — the second not statistically significant. Same category, same year, wildly different results. (preprint, 2025) "Ambient AI" is not one thing.

More on this in the documentation piece.

Clinical summarisation — works, with a measurable error floor

The largest manual evaluation of LLM clinical note generation to date — 450 consultation transcript and note pairs, nearly 13,000 clinician-annotated sentences — found an overall hallucination rate of 1.47%, of which 44% were classified major, and an omission rate of 3.45%. The best configuration produced two major hallucinations across 25 notes with zero major omissions. (npj Digital Medicine, 2025)

Two readings of that number are both correct. It is remarkably low for free-text generation. It is also not zero, and omission — the model quietly leaving something out — is both more common and harder to catch on review than fabrication.

A separate blinded comparison found hallucinations in 31% of ambient notes against 20% of human-written notes at the note level, with human notes rated better on accuracy and internal consistency. (Frontiers in AI, 2025) Worth noting that study's publication fees were funded in part by an ambient-AI vendor, which cuts in an interesting direction given the finding.

Speech recognition — a specific, documented failure mode

The transcription layer under several ambient products has a known hallucination problem. Peer-reviewed analysis found roughly 1% of audio transcriptions containing entirely fabricated phrases, with 38% of those including explicit harms, and hallucinations concentrated among speakers with longer non-vocal pauses — a symptom of aphasia. (FAccT, 2024) A broader safety review of LLM scribes reports 1–3% overall error rates, against 7–11% for legacy speech recognition, but with failure modes including fabricated exams, misattributed speakers, and systematically higher error rates for African American speakers. (npj Digital Medicine, 2025)

Medical coding — does not work, despite what is being sold

This is the clearest negative result in the field and it is being widely ignored. A benchmark in NEJM AI tested four leading models on reproducing codes from clinical descriptions across ICD-9-CM, ICD-10-CM and CPT, drawing on more than 27,000 codes. Every model scored below 50%. (Soroush et al., 2024) Work on evaluation and management codes specifically tends to report improvement over a baseline rather than absolute accuracy — the best purpose-built system in one 2025 study beat a single-prompt GPT-4o baseline by 33.7 percentage points, and the paper publishes no absolute accuracy for any system.

Set against a market where autonomous-coding accuracy above 95% is commonly advertised and, as far as we can find, never independently validated. Our reading is in the coding piece: models should assemble the medical decision making and a clinician should choose the code.

Diagnostic decision support — works until it doesn't, and then it does harm

In a randomised study of 457 clinicians, standard AI raised diagnostic accuracy from 73.0% to 75.9%. Systematically biased AI dropped it to 61.7% — and model explanations did not mitigate the harm. (JAMA, 2023) In mammography, incorrect AI suggestions cut accuracy among very experienced radiologists from 82.3% to 45.5%. (Radiology, 2023)

The asymmetry is the point: the upside of correct AI advice is a few percentage points, and the downside of incorrect advice is a collapse. That ratio is why we build approval gates rather than autopilots.

Revenue cycle — real adoption, unproven returns

Sixty-three percent of healthcare organisations report using AI and automation somewhere in the revenue cycle — and of those, only 15% report having seen a positive return. (HFMA, 2025) Among practices, 71% report some AI use in patient visits, but 47% use it in a quarter of visits or fewer, and 44% report no reduction in staff workload at all. (MGMA, 2025)

Regulatory status — one clearance, widely misdescribed

FDA cleared a 510(k) in December 2025 for software using a patient-facing large language model, in diabetes and insulin titration management. It was cleared against a drug-dose calculator predicate, not through a novel LLM pathway, and reporting has since noted the open question of whether the model is an interface or the decision-maker. (STAT, 2026) Describing this as "FDA approved an LLM" is wrong in at least three ways.

The thing physicians are worried about that nobody is measuring

In the AMA survey, only 13% of physicians worried AI would worsen their stress or burnout. But 88% expressed some level of concern about skill loss, and 70% were very or somewhat concerned specifically about loss of skills in current medical students and residents. There is, as far as we can find, no published measurement of whether that is happening.

It deserves to be taken seriously by everyone building these tools, including us. A trainee who never composes an assessment and plan from scratch is learning a different job than the one their attending learned. We do not know what that costs. We should probably find out before we have a generation of evidence.

How to read any AI claim in this field

• Ask what was measured, not what was reported. Perceived time savings and logged time savings have been shown not to correlate.
• Ask about the denominator. A result averaged across everyone who was given access is a different number from a result among people who actually used it.
• Ask who funded it. Vendor-run surveys are not worthless, but they are not studies. This applies to ours.
• Ask whether there is a control group. A meta-analysis of ambient AI documentation found zero randomised trials among 23 included studies.