Healthcare AI: what is getting real adoption now?

Has healthcare AI finally moved beyond pilots?

Healthcare AI has clearly moved beyond pilots in a few workflows, especially clinical documentation, medical imaging and medical coding.

The broader market is still much messier than the headlines suggest. KLAS Research has found that almost every healthcare organization it surveyed was either using or testing some form of AI, yet much smaller numbers had rolled the same tools across whole departments or health systems. That gap between “we have AI” and “people use this every day” is still central to the story.

The useful way to look at adoption today is to follow what happens after a pilot. UCHealth tested Abridge with roughly 250 providers before expanding it to more than 2,300. Geisinger moved beyond 1,000 clinicians. CommonSpirit Health has been scaling Abridge region by region across both Epic and Oracle environments after selecting the product in 2025. Ardent Health has already passed one million patient encounters using ambient AI.

Medical coding provides another kind of proof. CodaMetrix says its software now touches more than 50 million outpatient visits a year across more than 220 hospitals. In radiology, the American College of Radiology says AI use among surveyed members has climbed from roughly 30% around 2020 to close to 90% in its latest survey work, although the surveys do not use identical samples.

These are deployments being used thousands or millions of times inside normal healthcare operations.

The adoption is concentrated, though. Hospitals are moving fastest when AI removes repetitive work while a human remains responsible for the final result.

Healthcare AI use case	What adoption looks like now	Our judgment
Ambient documentation	Large enterprise rollouts and heavy recurring use	Very strong
Radiology and imaging AI	Years of production use across many narrow tasks	Very strong
Medical coding	High automation rates in selected specialties	Strong
Patient messaging	Widely available, much lower actual acceptance	Moderate
Prior authorization	Rapid product development and early deployment	Moderate
Broad autonomous clinical decisions	Mostly narrow or supervised use	Low

What actually counts as real healthcare AI adoption?

Real healthcare AI adoption starts when a hospital keeps using the product after testing it and the workflow becomes part of normal work.

Healthcare AI numbers often mix very different things. FDA clearance tells us a product is allowed onto the market for a specific medical use. A contract shows that someone paid for it. A pilot shows that a hospital was curious enough to try it. None of those tells us how often clinicians actually use the software.

Usage becomes much more convincing when several pieces line up: the hospital expands after a trial, clinicians repeatedly choose the tool, the product sits inside the EHR or another existing workflow, and someone can measure a result such as time saved, coding automation, fewer denials or faster treatment.

MultiCare gives us a good example. The health system evaluated three ambient-AI products with more than 550 clinicians across more than 20 specialties before selecting Ambience for a broader rollout. That decision carries more weight than an announcement that a hospital has “partnered with an AI company.” MultiCare tested alternatives and still decided that the software deserved a larger place in its workflow.

The same filter should be applied to radiology. The FDA has cleared a huge number of AI-enabled medical devices, with radiology accounting for the large majority of them. The more interesting evidence is that radiologists increasingly say they actually use AI in practice.

For the rest of this article, we give far more weight to repeated usage, expansion and measurable outcomes than to product launches.

Are AI scribes the first generative-AI product healthcare actually wants?

Ambient AI scribes are currently the clearest example of generative AI finding real product-market fit inside healthcare.

The scale has moved quickly. Abridge says it is now live across more than 300 health systems. The company expects to support tens of millions of clinician-patient conversations this year, and recent deployments continue to expand beyond the first wave of academic hospitals. Microsoft says Dragon Copilot is used by more than 100,000 clinicians. Ambience has reported utilization around 80% in some large deployments.

More important than the vendor totals is what individual hospitals do after testing the technology. UCHealth moved from about 250 Abridge users to more than 2,300. Geisinger passed 1,000 clinicians. CommonSpirit built a rollout across five regions and two different EHR environments. MultiCare evaluated several competing products before choosing one for enterprise deployment.

The workflow itself helps explain the speed. Doctors generate documentation again and again throughout the day. An AI scribe can prove its value several times during one shift. Clinicians also know exactly what the software is replacing: typing notes, dictating them later, or finishing documentation after work.

The category is already moving beyond basic note generation. Abridge recently added pre-visit and pre-admission summaries that pull together information from the medical record and link the summary back to its sources. Microsoft is expanding Dragon Copilot into nursing and other clinical workflows. Ambient vendors are also adding coding, orders and administrative tasks around the encounter.

The scribe is becoming the entry point for a much broader clinical assistant. That expansion is happening fast.

If you want more recent data on this point, please see our latest healthcare AI market report.

Do AI scribes actually save doctors enough time to matter?

Ambient AI is producing meaningful reductions in documentation burden, although the size of the benefit varies a lot by clinician and product.

A multicenter study published in JAMA Network Open followed 263 ambulatory clinicians across six U.S. health systems using Abridge. Among clinicians included in the burnout analysis, self-reported burnout fell from 51.9% before adoption to 38.8% after 30 days. Participants also reported spending about 0.9 fewer hours on after-hours documentation.

A separate randomized study involving 238 physicians found a more mixed result. Nabla reduced time spent in notes by 9.5% compared with the control group, while the DAX Copilot arm did not produce a statistically significant reduction on that particular measure. That difference is useful because every ambient product clearly does not produce the same result.

Hospital data show the same variation. Houston Methodist has reported documentation-time reductions of around 40% in its Ambience deployment. Other health systems report smaller gains, and a recent clinician survey cited by Heidi found that 42% of respondents experienced no time saving at all from AI scribes.

So one strong study or vendor case study should not become a universal “AI saves doctors two hours a day” claim. The narrower conclusion is still impressive: enough clinicians are saving enough time to keep using the product, and enough health systems are seeing a benefit to keep expanding it.

The burnout effect may also matter as much as raw throughput. Saving 30 or 50 minutes at the end of the day creates value even when the hospital cannot turn that time into another billable appointment.

Is radiology AI already more established than generative healthcare AI?

Radiology AI is currently the most mature clinical AI market, even though ambient scribes are growing faster and attracting more attention.

Radiology had a long head start. The FDA’s list of AI-enabled medical devices is dominated by imaging products, and the American College of Radiology says close to four-fifths of FDA-cleared medical AI devices are related to radiology. The tools cover CT scans, mammography, ultrasound, lesion detection, image reconstruction, segmentation, stroke detection and many smaller jobs.

Actual clinician use has followed. The ACR’s latest survey work found nearly 90% of responding radiologists using some form of AI. Earlier ACR research around 2020 found adoption closer to 30%. The samples differ, so we cannot treat that as a precise tripling of adoption, but the direction is too large to ignore.

The real-world products are also much older than most generative-AI tools. Viz.ai began with stroke detection and care coordination and says its platform now reaches roughly 2,000 U.S. hospitals and supports dozens of disease-specific pathways. Aidoc, RapidAI and other imaging companies have similarly spent years embedding narrow algorithms into hospital workflows.

Most radiology AI still handles very specific jobs. The software may flag a possible pulmonary embolism, push a suspected brain hemorrhage higher in the worklist, make measurements or highlight an area that deserves attention. The radiologist remains responsible for the interpretation.

That narrowness helped radiology AI get adopted. Hospitals could add one useful capability without asking doctors to hand over the entire diagnostic process to a model.

What we can measure	Radiology AI	Ambient generative AI
Time in production	Many tools have years of deployment	Mostly much newer
Regulatory footprint	Very large	Usually outside device regulation when used for documentation
Clinician penetration	High among surveyed radiologists	Growing quickly, but uneven across specialties
Typical job	Detection, triage, measurement, reconstruction	Documentation and encounter support
Human role	Radiologist keeps final interpretation	Clinician reviews and signs output

Does radiology AI actually improve care?

Radiology AI is showing real value when speed matters, especially when the software helps the right specialist see an urgent case sooner.

Stroke is probably the cleanest example. When an algorithm detects signs of a large-vessel occlusion and alerts the stroke team immediately, the main benefit can come from shortening the chain between imaging, specialist review and treatment.

Viz.ai has published and sponsored a large body of clinical evidence around that workflow. One study presented at the International Stroke Conference found a 44% reduction in door-in-door-out time at a primary stroke center using its platform for patients who needed transfer for large-vessel-occlusion treatment.

That kind of deployment is also why diagnostic accuracy alone is an incomplete way to judge healthcare AI. A model can be commercially useful without outperforming a radiologist. Saving several minutes, prioritizing the right scan or automatically performing tedious measurements may be enough.

There is a less comfortable side to the story. The ACR is putting more resources into monitoring AI after deployment because performance can change across hospitals, scanners, patient populations and time. Its Assess-AI registry exists to collect real-world performance data once models leave controlled validation settings.

Radiology has therefore reached the stage other healthcare AI categories are approaching now: the question has shifted from “can hospitals deploy this?” to “does it keep working well after thousands of real patients?”

Is autonomous medical coding really happening now?

Autonomous medical coding has quietly become one of healthcare AI’s strongest production use cases.

CodaMetrix now says its platform serves more than 60,000 physicians across more than 220 hospitals and processes over 50 million outpatient visits annually. KLAS named CodaMetrix the top vendor in its autonomous clinical coding category for 2026, with 93% of surveyed customers saying they would buy the product again.

The automation rates at individual systems are even more interesting. Mass General Brigham has reported around 85% automation in radiology coding and 73% in pathology through CodaMetrix, alongside a 58.7% reduction in coding-related denials. University of Colorado Medicine has reported a radiology automation rate of roughly 92%, compared with a previous ceiling around 46%, while cutting coding lag by 3.6 days.

Those numbers describe a workflow where software handles the majority of cases. Human coders increasingly focus on exceptions, audits and cases the model does not confidently process.

Fathom and Nym are pursuing the same market, while companies such as AKASA are pushing AI further into documentation review, revenue integrity and claim preparation. The economics are straightforward: every encounter has to be coded, coding labor costs money, delays hurt cash collection and errors can lead to denials.

Healthcare organizations can calculate the return much more easily than they can for a general “clinical intelligence” product.

The word autonomous should still be read carefully. Hospitals retain auditing, compliance controls and exception handling. In practice, bounded automation is exactly why these systems can scale.

If you want more recent data on this point, please see our latest healthcare AI market report.

Are hospitals really letting AI write patient messages?

AI-written patient messages are becoming part of normal EHR software, but clinicians still reject most generated drafts in the best published studies we have.

Stanford Health Care tested an EHR-integrated generative-AI system with 162 clinicians over five weeks. Clinicians used roughly 20% of the suggested responses. They reported lower task burden and work exhaustion, yet the study found no clear improvement in reply, reading or writing time.

Another deployment generated more than 21,000 potential responses across nine clinics and saw only around 12% accepted. Nurses were generally more receptive than physicians and advanced-practice providers.

Those percentages tell us more than the fact that the feature exists inside an EHR. A 12% or 20% acceptance rate means AI is helping with a minority of messages while clinicians still prefer to write or heavily edit most replies.

Patients themselves appear more comfortable with the idea than some clinicians may expect. A large Duke-linked survey involving 1,455 people found satisfaction above 75% across different message types, with respondents sometimes preferring AI-drafted replies when they did not know who wrote them. Later qualitative work found patients were generally comfortable with AI helping on routine messages when a clinician reviewed the answer.

The dividing line is intuitive. Scheduling questions, refill requests and simple follow-ups are easier places for AI to help. A frightening test result creates a much higher expectation that a real clinician has read, understood and personally handled the message.

Patient messaging is a real deployment category today, but usage still trails far behind ambient documentation.

Is healthcare AI taking over the back office faster than clinical care?

Healthcare AI is spreading extremely quickly through administrative work, and some of the largest deployments in the sector barely involve medical decision-making at all.

The NHS in England is one of the clearest examples. After an earlier deployment involving roughly 30,000 workers, NHS England expanded Microsoft 365 Copilot access to around 505,000 clinicians and support staff. The earlier trial reported an average self-reported saving of 43 minutes per user per day.

That 43-minute figure should be treated cautiously because reported time saved and objectively measured productivity are different things. The seat count is still remarkable. Half a million people getting access to an AI work assistant inside one healthcare system is adoption on a completely different scale from most clinical AI deployments.

Hospitals are applying the same logic to call centers, scheduling, revenue-cycle management, coding, documentation review, referrals and patient access. These jobs contain huge quantities of text, forms and repetitive information transfer, which makes them unusually compatible with current AI.

The operational risk is easier to control too. A bad meeting summary can be corrected. A coding exception can be reviewed. A scheduling mistake can often be fixed. A wrong cancer treatment recommendation carries an entirely different consequence.

That risk gradient explains much of the market today. Healthcare organizations are already willing to buy AI aggressively. Spending is moving first toward the parts of healthcare where mistakes can be caught cheaply.

If you want more recent data on this point, please see our latest healthcare AI market report.

Is AI finally fixing prior authorization?

AI is beginning to take real work out of prior authorization, but the category has not reached the maturity of ambient documentation or medical coding yet.

The opportunity is huge because prior authorization combines several jobs that modern AI handles well: finding information inside a record, matching it against payer rules, gathering evidence and preparing a request.

The infrastructure around the process is also changing. CMS already requires affected payers to give specific reasons for prior-authorization denials and meet tighter decision timelines. Starting in 2027, many Medicare Advantage, Medicaid, CHIP and marketplace plans will also have to support standardized prior-authorization APIs. That creates a much cleaner technical environment for automation than the old mix of portals, faxes and phone calls.

Healthcare AI vendors are positioning themselves ahead of that change. Abridge has connected its clinical workflow with Availity to work on real-time prior authorization. The idea is to spot what a payer needs while the patient is still with the clinician instead of discovering days later that the documentation was incomplete. Revenue-cycle companies are building similar systems around chart review and evidence gathering.

The hard part is that authorization crosses organizational boundaries. The provider may automate its side beautifully while the insurer still uses different rules, systems or manual checks. One good AI agent cannot fix the entire workflow on its own.

Prior authorization could become one of the biggest healthcare-AI categories once the API infrastructure becomes standard. Today, there is real deployment, strong demand and obvious economics, but the evidence of broad routine usage remains thinner than it is for scribes or coding.

Are healthcare AI copilots becoming real, or are they still mostly product demos?

Healthcare AI copilots are starting to become useful beyond note-taking, but the strongest adoption still sits in tightly controlled tasks.

Abridge gives us a good view of where the category is moving. Its latest product releases now include pre-visit summaries, pre-admission summaries, coding recommendations and access to clinical evidence alongside the original documentation workflow. The company has also been building connections to prior authorization and other actions around the encounter.

Microsoft is moving in the same direction with Dragon Copilot. Epic is adding chart summarization, patient communication, documentation assistance and suggested actions directly inside its EHR. The three companies are gradually surrounding the clinician rather than offering a single isolated AI feature.

What remains much harder to prove is how often clinicians use those newer functions. We have strong evidence that doctors repeatedly turn on ambient documentation. Public evidence is much thinner on thousands of doctors routinely asking an AI system to plan care, prepare orders and coordinate the full next step of treatment.

The latest product moves make the ambition obvious. Abridge has even started pushing toward agents that can act inside EHR workflows rather than merely generate text. That is an important change in product direction, but recent launches should not be mistaken for established adoption.

For now, the healthcare copilot is becoming real one task at a time.

Is AI making diagnosis and treatment decisions on its own yet?

Autonomous AI diagnosis and treatment are still rare in everyday healthcare, especially compared with the scale already reached by AI documentation, imaging support and coding.

There are plenty of systems that influence clinical decisions. Imaging AI flags suspicious scans. Stroke platforms alert care teams. ECG models identify unusual patterns. Risk models highlight patients who may deteriorate. Newer generative products can retrieve medical literature or summarize a patient’s record before a clinician makes a decision.

In most of those workflows, the doctor still decides what happens next.

That gap shows where healthcare institutions currently draw the line. They are comfortable letting software listen, summarize, classify, prioritize and recommend. Trust drops sharply when the same system would independently diagnose a disease, prescribe a drug or decide that a patient does not need treatment.

Clinical errors also behave differently from administrative errors. A flawed generated note can be corrected before it is signed. A coding system can send uncertain cases to a human queue. A false treatment decision may immediately affect the patient.

Regulation and liability obviously play a role, but day-to-day clinical trust is the bigger obstacle. Hospitals need to know how a system behaves across their own patient population, what happens when the model is uncertain and who catches the mistake.

So the current adoption frontier sits much closer to “AI helps the clinician decide” than “AI becomes the clinician.”

Will Epic and the big EHR companies eventually crush standalone healthcare AI startups?

Epic has a huge distribution advantage in healthcare AI, but standalone companies are still winning major deployments when clinicians think their product is clearly better.

Epic can place new AI features directly inside software hospitals already use. It can connect them to patient charts, scheduling, orders and messaging without asking the customer to build another major integration. As Epic expands its own AI Charting and broader AI tools, standalone vendors will increasingly have to justify why a hospital should pay for another platform.

That pressure is already visible in ambient documentation. Yet Abridge, Ambience and other specialists keep winning large health systems despite the obvious EHR advantage. MultiCare tested several products and selected Ambience. UCHealth expanded Abridge after a long pilot. CommonSpirit chose Abridge and then built a multi-region rollout around it.

Frequent usage gives specialist vendors some room to survive. A doctor may use an ambient system during almost every patient visit. Small differences in note quality, latency, specialty vocabulary or editing can become irritating very quickly when repeated 20 times a day. Hospitals will pay separately when those differences produce much stronger clinician adoption.

The next fight will be harder because Epic is steadily adding more of the surrounding workflow. A standalone scribe that only writes notes will become easier to replace. A company that controls documentation, chart context, coding, prior authorization and follow-up work has a much stronger reason to exist.

That is why the recent expansion of Abridge, Ambience and Microsoft beyond transcription matters more than another scribe feature. The market is already moving from “who writes the best note?” toward “who owns the work around the clinical encounter?”

If you want more recent data on this point, please see our latest healthcare AI market report.

What do the healthcare AI products that actually get adopted have in common?

The healthcare AI products getting real adoption today usually remove work people already hate, sit inside software people already use and make mistakes easy to catch.

Ambient documentation fits that pattern almost perfectly. The physician keeps talking to the patient while the software handles a task that previously happened during or after the visit. The doctor can read the note before signing it.

Medical coding works similarly. The model handles routine cases, while difficult or uncertain cases can move into an exception queue. Radiology AI often prioritizes scans or makes measurements while the radiologist keeps control of the diagnosis.

Patient messaging gives us the opposite test. AI can technically draft thousands of answers, yet studies showing 12% to 20% acceptance make clear that integration alone does not guarantee adoption. If reviewing the draft takes almost as much effort as writing the answer, clinicians simply skip it.

This may be the most useful pattern across the whole market. Model intelligence by itself tells us surprisingly little about adoption. Healthcare buys software when the product removes a repeated pain point with limited extra supervision.

The products struggling to move past pilots often ask humans to do something less attractive: open another dashboard, inspect another recommendation, supervise another agent or take responsibility for an output whose economic benefit is difficult to measure.

What helps healthcare AI get adopted	Why it works
Replaces a frequent existing task	Users feel the benefit immediately
Lives inside the EHR or current workflow	Little extra behavior is required
Produces something a human can review quickly	Errors remain manageable
Has an obvious cost or time saving	Buyers can defend the budget
Handles a narrow job reliably	Trust builds faster
Sends difficult cases back to humans	Automation can scale without pretending to be perfect

So what healthcare AI is actually getting real adoption now?

Healthcare AI is getting real adoption today, and the clearest winners are ambient documentation, radiology AI and medical coding.

Ambient AI is the strongest new category. Large systems have repeatedly expanded deployments after pilots, clinician populations have moved into the thousands, Abridge now reports more than 300 health-system customers, Microsoft says Dragon Copilot reaches more than 100,000 clinicians, and some individual deployments have already crossed one million patient encounters. Independent studies also show that documentation burden can fall meaningfully, even if results differ between products and clinicians.

Radiology AI has the deepest clinical history. As seen above, close to four-fifths of FDA-cleared medical AI devices sit in imaging, AI use is now widespread among surveyed radiologists, and companies such as Viz.ai have built footprints measured in thousands of hospitals. Radiology reached adoption through lots of narrow tools rather than one general-purpose AI system.

Coding is the quiet third winner. Automation rates above 80% in some radiology and pathology workflows mean AI is already doing most of a defined job inside real health systems. That is much stronger evidence than another hospital announcing an AI “partnership.”

The next group is less mature. Patient-message drafting is deployed widely but accepted much less frequently. Prior authorization has obvious economics and is getting a major infrastructure boost from upcoming federal API requirements. Broader clinical copilots are expanding quickly, although the newer decision-support and agent features still lack the depth of usage data we have for documentation.

Autonomous diagnosis and treatment sit much further behind.

The current healthcare AI market has a clear shape. Hospitals are handing over repetitive work while keeping people in control of consequential decisions. Documentation, imaging triage and coding already meet that test. More ambitious clinical copilots are moving in the same direction, but they have more to prove.

For now, the biggest healthcare AI adoption story is surprisingly practical: the technology winning inside hospitals is the technology that gives expensive people less work to do.

If you want more recent data on this point, please see our latest healthcare AI market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.