AI for medicine is no longer waiting outside the clinic door. In early 2026, 63% of U.S. physicians surveyed by Doximity said they were using AI, up from 47% in early 2025; 94% were either using it or interested in doing so.[1] That is adoption at clinical scale, not a pilot tucked into an innovation lab.
The harder fact sits beside it: only 2.4% of FDA-authorized AI/ML medical devices had clinical studies supported by randomized trial data, according to the Stanford HAI 2026 AI Index medicine chapter.[2] Those two numbers are the honest starting point. AI is already changing how many clinicians search, write, triage paperwork, and move through the day. It has not, as a class, earned the evidentiary confidence that many procurement decks imply.
Where Physicians Are Actually Using AI
The current clinical AI story is less glamorous, and more plausible, than the diagnostic supercomputer storyline. The most common reported uses are literature search tools, used by 35% of physicians in the Doximity survey, and ambient documentation or AI scribes, used by 29%.[1] Administrative workflow support follows the same practical logic: clinicians are reaching first for tools that remove friction from tasks they already know how to judge.
That matters because the clinician remains close to the output. A literature-search assistant may retrieve papers faster, but the physician still decides whether the evidence applies to a patient with renal disease, pregnancy, frailty, or a medication interaction. An ambient documentation tool may draft the note, but the clinician signs it. A prior-authorization or inbox workflow tool may reduce queue time, but someone still has to own the decision when the generated text is wrong.
The adoption pattern also has a primary-care signal. Among AI adopters in the Doximity report, family medicine physicians were the most frequent daily users, at 88%.[1] That is not surprising. A high-volume clinic session is where documentation, messages, forms, refills, results review, and patient counseling collide. If a tool reliably trims even part of that load, physicians notice quickly.
| Use case | What the tool changes | What the evidence can and cannot prove from current survey data |
|---|---|---|
| Literature search | Finds and summarizes medical literature faster | Shows physician uptake; does not by itself prove better decisions or outcomes |
| Ambient documentation and AI scribes | Drafts visit notes from the clinical conversation | Shows perceived burden relief; requires local review for note quality, omissions, and billing or medico-legal risk |
| Administrative workflows | Helps with forms, messages, routing, and other non-diagnostic work | May reduce clerical drag; survey results do not replace measured time, safety, or throughput data |
Perceived Benefit Is Real, But It Is Not the Same as Measured Outcome
The strongest near-term case for clinical AI is burden reduction. In the Doximity survey, 75% of physicians using AI reported reduced administrative burden, 69% reported better patient care, and 88% believed AI could meaningfully reduce burnout.[1] These are not trivial findings. Administrative burden is not a lifestyle inconvenience when it delays chart completion, fragments attention, and pushes cognitive work into exhausted evening hours.
But these are self-reported signals. They tell us what physicians perceive after using the tools. They do not tell us, without additional measurement, whether visit length changed, after-hours documentation fell, errors decreased, follow-up improved, or patient outcomes moved. The same caution applies to expectations about reducing “pajama time”: the 48% estimate reflects physician expectation of benefit rather than measured time savings.
That distinction is not academic hair-splitting. If an AI scribe makes a physician feel less buried but introduces subtle inaccuracies into the assessment and plan, the organization has traded one operational problem for another. If a literature tool speeds searching but overweights a non-applicable paper, the final clinical judgment still has to catch the mismatch. In medicine, usefulness at the keyboard and benefit at the bedside can overlap, but they are not interchangeable.
FDA Authorization Does Not Close the Evidence Question
Regulatory authorization is important. It can mean a product met a defined standard for a particular indication, population, data type, and performance claim. It should not be read as a blanket statement that the tool improves clinical outcomes across the messy environments where it will be installed.
The 2.4% randomized-trial figure from Stanford HAI is an aggregate measure across FDA-authorized AI/ML medical devices.[2] It does not mean every deployed tool lacks meaningful evidence, and it does not mean randomized trials are the right design for every workflow application. Some tools may need prospective silent-mode validation, reader studies, time-motion analysis, equity audits, or post-deployment safety monitoring more urgently than a conventional patient-outcome trial.
Still, the aggregate figure should make health systems careful. An FDA-authorized algorithm may perform well on the data used to support its submission and less well in a hospital with different scanners, documentation habits, referral patterns, disease prevalence, or patient demographics. The moment the model enters a local workflow, performance becomes partly a property of that workflow.
Accuracy Concerns Are Not Resistance to Innovation
The clinicians adopting AI are often the same clinicians worried about it. In the Doximity report, 71% of physicians cited accuracy and reliability as their top concern, and 47% said their institution’s AI policies were still evolving.[1] That combination is a useful reality check: physicians are using the tools, but many do not yet believe the governance environment has caught up.
Accuracy is also too blunt a word unless the use case is specified. A wrong medication-history summary can harm a patient differently from a missed radiology flag, a fabricated citation, a mistranscribed symptom, or a biased risk score. Each failure mode has a different reviewer, time pressure, and consequence. A hospital AI committee that treats all “AI accuracy” concerns as one risk bucket will miss the operational details that determine whether a tool is safe enough to use.
For ambient documentation, local validation should include note completeness, clinically meaningful omissions, attribution of speaker statements, problem-list behavior, billing implications, and clinician editing burden. For literature search, it should include source retrieval, citation fidelity, specialty relevance, and the handling of uncertainty. For administrative workflows, it should include routing accuracy, turnaround time, denial or appeal effects, and whether automation shifts work onto nurses, medical assistants, or patients.
The Hype Cycle Is Starting to Look Like a Governance Problem
Mass General Brigham’s Hugo Aerts described the field as moving from Gartner’s “Peak of Inflated Expectations” toward the “Slope of Enlightenment.”[3] That framing fits the moment, as long as it is not used to soften the underlying issue. The problem is not that clinicians are excited too soon. The problem is that deployment can scale faster than evidence, monitoring, and accountability.
The first governance question should be mundane: what task is the tool allowed to perform? A product that drafts documentation is not the same as one that recommends a diagnosis. A model that prioritizes an inbox is not the same as one that changes triage. A summarization assistant used by an attending physician is not the same as an autonomous message generator sent directly to a patient.
The second question is who checks the output, and when. A human-in-the-loop label is not enough if the human is overloaded, lacks the right context, or sees the AI output at a point in the workflow where correcting it is unlikely. In some settings, AI will save time only if the review burden is light. In others, the need for review is the safety mechanism that makes deployment acceptable.
The third question is what happens after launch. Models drift, documentation templates change, clinical practice changes, and patient populations differ across sites. A health system that validates once at purchase but does not monitor performance after implementation is treating AI like static equipment rather than adaptive software embedded in clinical work.
The Jagged Frontier Is a Better Mental Model Than “AI Works” or “AI Fails”
The inaugural ARISE State of Clinical AI Report describes a “jagged frontier”: models can show very strong performance on controlled tasks while remaining brittle in recognizing their own uncertainty.[4] That is a more clinically useful frame than treating AI as either a breakthrough or a fad. Many tools will be impressive in one narrow lane and unsafe when the lane boundary is unclear.
This is where forward-looking systems need the most careful language. Multi-agent diagnostic systems and digital-twin approaches may produce striking results in single studies or curated evaluations, but that is not the same as broad clinical readiness. A diagnostic orchestrator evaluated on selected cases is not yet a service line. A digital twin trial in a defined diabetes population is not yet proof that the method generalizes across chronic disease management.
Emerging evidence should be welcomed without being inflated. Medicine has room for tools that search better, listen better, draft better, flag earlier, and help clinicians see patterns they would otherwise miss. It also has a long memory for interventions that looked compelling before they were tested in the places where patients actually receive care.
What a Sensible Deployment Standard Looks Like
For now, the most defensible deployments are the ones with a clear task, a visible reviewer, a measurable workflow problem, and a monitoring plan. That standard does not require every AI tool to prove mortality benefit before use. It does require the institution to know what claim it is making.
- For documentation tools, measure note quality, editing burden, after-hours work, clinician satisfaction, and safety-relevant errors.
- For search and summarization tools, test citation accuracy, omitted evidence, specialty relevance, and the handling of conflicting literature.
- For administrative automation, measure turnaround time, rework, denials, staff workload, and patient-facing consequences.
- For diagnostic or predictive tools, validate prospectively in the local population before relying on the output in care decisions.
- For all clinical AI, define ownership when the system is wrong, and keep monitoring after launch.
The Doximity adoption data are also bounded. They describe U.S. physicians, not nurses, pharmacists, therapists, community health workers, or clinicians practicing outside the United States.[1] That boundary matters. The burden profile, governance structure, staffing model, and regulatory environment may look different in other settings. The evidence conversation should not pretend that one physician survey describes all clinical AI adoption.
So, is AI in medicine living up to the hype? Partly. It is already living up to some of the workflow hype, especially where it helps clinicians search, draft, summarize, and move administrative work out of the way. It has not yet lived up to the broader claim that AI is a proven clinical-effectiveness layer across medicine. Adoption is not effectiveness. Authorization is not local performance. Enthusiasm is not patient benefit. Deploy where the use case is clear, validate locally, monitor continuously, and keep the clinical claim no larger than the evidence can carry.
References
- Doximity 2026 State of AI in Medicine Report, Doximity, 2026.
- Stanford HAI 2026 AI Index Report, Stanford HAI, 2026.
- Mass General Brigham 2026 Predictions article, Mass General Brigham, 2026.
- ARISE State of Clinical AI Report 2026, ARISE, 2026.
Comments
Join the discussion with an anonymous comment.