The most popular artificial intelligence (AI) tool in medicine right now doesn't diagnose anything. It doesn't recommend treatments. It listens to a doctor talk to a patient and writes a note. That simplicity is why ambient AI scribes have been adopted faster than almost any physician-facing technology in recent memory.
As of 2026, 70% of physicians in the University of California, San Francisco (UCSF) health system were using AI scribes in their daily practice. At Kaiser Permanente, 7,260 physicians used AI scribes in more than 2.5 million patient encounters over 14 months ending in December 2024, according to an analysis published in NEJM Catalyst.
In a broader sense, the American Medical Association (AMA) found that 66% of U.S. physicians were using some form of AI in their practice by 2024, up from 38% the year before. A January 2025 Medical Group Management Association Stat poll found that AI tools are now the top technology priority for health care organizations, with 32% ranking them first, ahead of electronic health record (EHR) usability and revenue cycle management. That figure was just 13% in late 2023.
Adoption has outpaced evidence for most of the technology's brief life. The first real-world studies are now arriving, though, and what they show is encouraging, complicated and worth reading carefully.
Why scribes came first
Robert M. Wachter, M.D., chair of the department of medicine at UCSF, has spent much of the past two years studying AI’s growing role in clinical practice for his new book, “A Giant Leap: How AI Is Transforming Healthcare and What That Means for Our Future.” He argues that AI scribes have seen success because they accomplish a task that is both relatively easy for the technology to do well and carries relatively low consequences when something goes wrong.
What to look for when choosing an AI scribe vendor
The market is crowded and moving fast. Eighty-five percent of health care artificial intelligence (AI) investment money is going to startups, meaning there aren't many vendors with proven track records for Health Insurance Portability and Accountability Act (HIPAA) compliance or validation testing. Before signing a contract, here's what practices should evaluate.
HIPAA compliance and security. Table stakes. Dan Silverboard recommends seeking "robust representations and warranties that the vendor is HIPAA compliant, including privacy and security policies and security risk assessments." Ask for documentation, not assurances.
Validation and bias testing. Does the vendor conduct ongoing validation of its AI output? A January 2025 study in the Journal of Medical Internet Research found errors in 70% of AI-generated notes, with omissions — the hardest type for physicians to catch — the most common.
Data governance. Contracts should specify how protected health information is used and prohibit vendors from attempting to reidentify deidentified data.
Liability and contract terms. Most contracts require the physician to sign off on all output and include broad liability disclaimers. CMS conditions of participation require that medical records be accurately written, and providers attest to claims compliance when they enroll with Medicare. AI documentation that goes unchecked creates regulatory exposure, not just clinical risk.
Track record and references. Ask for references from practices of a similar size and specialty. Kaiser Permanente's data showed that consistent users benefited most, so ask about typical adoption curves and onboarding support.
Cost and EHR integration. Pricing ranges from free to several hundred dollars per clinician per month, depending on electronic health record (EHR) integration depth and contract terms. Considering many AI scribe vendors do not post public pricing, it’s up to practices to press for transparent quotes. Ask whether implementation, training and ongoing support are included or billed separately.
Patient disclosure. In the UCLA trial, less than 10% of patients declined when told about the scribe. A brief verbal notification at the start and end of each recording is recommended, rather than just a signed waiver. Simple cues, like a sign reading "Digital scribe active,” reinforce transparency as a practice standard.
"You don't start with the hardest problem," Wachter said. "Diagnosis is both hard to get right, and if you get it wrong, you can kill somebody." He compared the path to autonomous vehicles, which didn't begin with driverless cars but with cruise control and automatic braking. AI in health care needed its own version of cruise control. Documentation was it.
The tools capture a patient encounter through an in-room recording and generate a structured clinical note, organized into standard sections such as chief complaint, history of present illness, review of systems, assessment and plan.
With the advancement of generative AI, most scribes now go beyond simply transcribing a visit. They parse the clinical content of the conversation, filter out small talk, suggest billing codes, flag care gaps and draft patient-facing visit summaries. Some can generate near-accurate notes from visits conducted in languages other than English.
The key distinction from older voice-to-text tools, and what Wachter describes as the real magic of generative AI, is the ability to take a sprawling conversation and figure out what belongs in the note and what doesn't, pulling clinically related details together — even when a patient mentions chest pain at the start of a visit and shortness of breath five minutes later after a digression about a grandchild’s soccer game.
What the studies show
One of the first randomized clinical trials of AI scribes, published in NEJM AI in November 2025, assigned 238 physicians across 14 specialties at the University of California, Los Angeles (UCLA) to use one of two commercial tools — Microsoft DAX or Nabla — or continue with their usual workflow.
Among approximately 72,000 encounters, Nabla users saw average note-writing time drop from 4 minutes, 30 seconds to 3 minutes, 49 seconds — a roughly 10% reduction compared with the control group. DAX users saw a smaller, statistically insignificant decrease. Both tools showed modest improvements in burnout and cognitive workload rates.
A separate study published in JAMA Network Open in August 2025 tracked more than 1,400 clinicians at Mass General Brigham (MGB) and Emory Healthcare. At MGB, the rate of burnout fell from 52.6% to 30.7% over 84 days. The program scaled from 18 pilot physicians in July 2023 to more than 3,000 with access by spring 2025.
"There is literally no other intervention in our field that impacts burnout to this extent," said Rebecca Mishuris, M.D., M.P.H., co-senior author and chief medical information officer at MGB.
At the American College of Physicians (ACP) Internal Medicine Meeting 2025 in New Orleans, Deepti Pandita, M.D., FACP, FAMIA, VP of clinical informatics and chief medical information officer at the University of California, Irvine, presented survey data showing that, when assisted by AI, physicians saved "anywhere from 12 to 20 minutes" per day on documentation.
“If you’re working 5 days a week, [20 minutes per day] is giving you back a couple hours of your life, which you can spend in any way you want,” Pandita said.
Wachter cautioned that the efficiency story has been somewhat overstated. "We thought it would save 5 or 10 minutes a visit. It turns out, it doesn't really save that much time," he said. The bigger shift is what happens with the time that is recovered. "Some of that time is repurposed into actual, genuine human contact, which feels better for the doctor and feels better for the patient."
On financials: Is it enough to reduce burnout?
A January 2026 study reported in JAMA Network Open, led by A Jay Holmgren, Ph.D., M.H.I., an assistant professor and researcher at UCSF, analyzed more than 1.2 million ambulatory encounters across 1,565 physicians. Those who adopted AI scribes generated 1.81 additional relative value units and saw roughly one more patient per week, with no increase in claim denials, compared to nonadopters. The authors estimated that translates to about $3,044 in additional annual revenue per physician, using the 2025 Medicare Physician Fee Schedule, which is roughly enough to cover the cost of the tool.
The throughput math tells only part of the story. In a separate viewpoint published in JAMA Internal Medicine in February 2026, Wachter and co-authors Lisa S. Rotenstein, M.D., MBA, M.Sc., and David W. Bates, M.D., M.Sc., argued that traditional return on investment calculations miss what matters most: reduced malpractice risk, improved patient retention, and the cost of replacing physicians who burn out and leave. The cost to replace a single primary care physician is in the range of $500,000 to upward of $1 million, according to the AMA.
"The AI scribe probably kind of pays for itself just in pure throughput and time savings," Wachter said. "But the real benefit is the joy in practice, recruitment, retention."
Shannon Sims, M.D., chief product officer at Vizient and an internal medicine physician, made a similar case. "Ambient listening reduces pajama time or evening documentation, that's well described," Sims said. "But there are also other tools around billing and coding and even access tools or things like medication refills." The compounding effects matter for retention. "A happier doc is going to be more productive, they're going to provide better care, and they might extend their career. I really encourage these physicians and organizations to think not just about the 12-month profit and loss, but also about the longer-term joys and rewards of health care."
Not every practice can afford the math, even when it works at scale. Pricing across the AI scribe market has stratified into clear tiers.
As of early 2026, budget-friendly stand-alone apps can start as low as $40 per clinician per month — and often offer free tiers — but with strict caps on monthly notes, users, and often limited EHR integration, they aren’t always practical for full-time clinicians.
Mid-range tools such as Nabla, which was one of two products tested in the UCLA trial, can run practices upward of $100 per month per clinician. Enterprise platforms with deep EHR integration, including Microsoft’s DAX Copilot and Abridge, run significantly more — often several hundred dollars per clinician per month, often favoring multiyear contracts.
The market is also beginning to see EHR vendors bundle ambient tools directly: athenahealth launched athenaAmbient, and Epic announced its AI Charting in February 2026.
Melissa Lucarelli, M.D., a family physician running a three-provider practice in rural Wisconsin, said getting a tool that integrates with her EHR would cost about $500 per provider per month. "$1,500 a month is a lot for a practice of my size, so we elected to hold off," she said during a Medical Economics and Physicians Practice panel discussion on the future of independent medical practice. "My hope is that AI is just eventually going to be part of the EHR."
On the same panel, David Eagle, M.D., president of the American Independent Medical Practice Association, pointed to a broader opportunity. "For revenue cycle management and how many personnel were directed to that, there's a huge opportunity for AI to help," he said. "I think this will be one of the most rapidly changing areas that we've ever seen in our career."
Lucarelli may not be wrong to wait. Wachter observed that AI scribe companies built "breathtakingly good" tools, but within two years, they became commoditized, and prices have already begun to fall. “Now Epic is going to do it themselves," he said.
Where the tools fall short
A January 2025 study in the Journal of Medical Internet Research tested two commercial AI scribe products in a simulated clinical setting and found an average of roughly three errors per note, with errors in 70% of the notes reviewed. Omissions, where relevant clinical details were left out entirely, were the most common type and the hardest for physicians to catch — spotting what's missing requires recall, not just recognition.
The UCLA trial flagged similar concerns. Physicians in both groups reported encountering clinically significant inaccuracies on an occasional basis. Senior author and UCLA Health internist John N. Mafi, M.D., M.P.H., said the technology requires "active physician oversight, not passive acceptance,” adding that “physicians must remain vigilant in reviewing AI-generated documentation … embracing innovation while maintaining medicine’s fundamental commitment to patient safety through rigorous evaluation and ongoing monitoring."
Wachter was blunt about the limits of that oversight. "If I've used the scribe in the last 49 notes and they were perfect, am I really going to be fully attentive as I review note No. 50?" he said. "If I'm human, the answer is no."
He raised another concern that hasn't gotten much attention. Before AI scribes, typing a note post visit was itself a thinking process.
"Sometimes, while I was typing my note, I would say, ’I forgot to ask whether she has a pet at home,’” he said. “There's something in the process of writing that is probably more cognitively active than reading over a draft." Physicians are voting with their feet, he acknowledged, and the trade-off appears to be worth it. But he called it the biggest open question of the scribe era.
He also raised the question of enforcement. At UCSF, where 70% of physicians now use an AI scribe, there is no technical mechanism to ensure that a physician has actually read the note before signing it. And the remaining 30% who have not adopted the tools are worth watching: Some are simply resistant to change, sure, but others may be raising legitimate questions about whether the technology is mature enough for their specific workflow.
That tension between enthusiasm and oversight runs through the malpractice conversation. Richard Anderson, M.D., FACP, chairman and CEO of The Doctors Company and TDC Group, said that most physicians using sophisticated ambient listening programs "are actually quite pleased with them" because the tools effectively solve the documentation burden created by EHRs. But he noted the irony: The solution to a technological problem is another, more sophisticated technology. "We've been down that road before," he said, "and it would be nice if every new technology didn't require another technology to monitor it and regulate it."
Deepika Srivastava, chief operating officer at The Doctors Company, emphasized that transparency is central to managing the risk. "Patients feel comfortable and consent to technology use when the physician has taken the time to communicate," she said. Srivastava recommended that providers establish clear protocols for when ambient listening will or won't be used, how consent is documented and what alternatives are available if a patient opts out. "Be clear, be respectful and make communication the foundation of the relationship," she said.
Dan Silverboard, a health care attorney with Holland & Knight, framed the legal picture simply: "Health care providers may simply sign off on whatever recommendation the AI program makes or approve ambient documentation without verifying that it accurately reflects the encounter. Those are the key risks."
The physician, not the technology, bears responsibility. Texas, for example, already requires providers to verify AI-generated documentation, and other states have followed with similar guidance.
Silverboard also urged physicians to document their use of AI in the medical record and to remember that they are personally attesting to the accuracy of every billing code. "There is no 'get out of jail free card' if an AI recommendation results in a higher code than what was actually performed," he said.
Patient acceptance is another variable, though the data are more encouraging than many expected. In the UCLA trial, less than 10% of patients declined to have the AI scribe used during their visit. At Kaiser Permanente, roughly two-thirds of surveyed patients said they were comfortable with the technology. But Wachter noted that some patients remain "a tiny bit creeped out by it," and practices should understand their population before assuming universal acceptance.
What comes next
Sims emphasized that the payoff depends on redesigning workflows, not just deploying technology. Practices that layer a scribe on top of a broken documentation process won't capture the full benefit.
A rapid review published in JMIR AI in October 2025 screened more than 1,400 studies on ambient scribes and found that just six met rigorous criteria for real-world evidence. The tools show promise, but the evidence base is still thin.
At UCSF, physicians have embraced the technology. "We've almost gotten to the point where, if we turned it off, we might lose a fair number of doctors," Wachter said. "It's almost become an expectation of practice now."
At Kaiser Permanente, a clinician’s age or years in practice had no correlation with adoption success. Instead, it was the heaviest-documenting specialties — mental health, primary care and emergency medicine — that saw the highest uptake.
Ultimately, the success story of AI scribes has implications far beyond improved documentation workflows. “The biggest success of scribes, when that story is finally written,” Wachter said, “will be winning over hearts and minds and creating receptivity for bigger, broader and, ultimately, more impactful tools.”
In “A Giant Leap,” he says the success of AI scribes is a single in the health care ballpark. The home run is clinical decision support that helps physicians choose the right test and the right treatment. Those tools require trust. And in health care, trust has to be earned one step at a time.
The AI scribe era is here. The tools are improving fast, the research is building and the message is consistent: Review your notes, understand your vendor and don't treat the technology as a finished product.
