Donna Roach had 100 licenses and no budget when she walked into her board retreat two years ago.
She brought a demo of the ambient AI technology with her anyway. One of her physician informaticists and her CMIO sat down in front of the University of Utah Health board and staged a patient visit — talking through an exam while the software listened and assembled the note in front of them.
“They’re looking at it going, ‘Really? Does it really do this?'” said Ms. Roach, the system’s chief digital information officer. “And we’re like, ‘Yes, it does.'”
The board backed her. The hundred licenses went almost immediately — “everybody wanted it” — and because Ms. Roach had no enterprise funding, individual divisions could buy their own. Some passed. Then they watched their peers and returned. Utah went to 300, then 400 licenses, and eventually to an enterprise agreement split between the school of medicine and the health system.
These were the arguments health systems were having about ambient AI for clinical documentation two years ago: does it work, and will clinicians actually use it? Across the six systems in this reporting, that two-part question is now a case closed. Adoption isn’t the constraint anymore; neither is enthusiasm.
Two things have changed in that time.
The first is what Abridge is able to do. The company that sold health systems a documentation technology now works the full arc of the visit. Its job is not just recording what was decided, but playing a role in care teams’ deciding. Context-aware clinical decision support that is patient-specific marks a different experience for clinicians who have long sought ways to be more present and better serve their patients. Care Signals, co-designed at Kaiser Permanente, surfaces a patient’s key conditions before the appointment, prompts the physician when one goes unassessed, and closes it out in coded documentation afterward. In June, it announced a foundation model for clinical conversations built with NVIDIA, a strategic investment from Eli Lilly aimed at surfacing clinical trial candidates at the point of care, and interest from Aetna and Cigna in tying clinicians’ documentation to claims adjudication.
If a scribe documents what happened, then this is technology that no longer begins or ends there. Abridge is earning progressively harder assignments around the visit and developing a deep bench of industry partners to tackle them. Growth like this is about staying power.
“Success depends on the network effect of a multistakeholder ecosystem — the best want to partner with the best to be able to anticipate and build for the future,” said Barry Stein, MD, chief clinical innovation officer and CMIO, and founder of the Center for AI Innovation in Healthcare at Hartford HealthCare. “There are only a few startups, healthcare systems, hyperscalers that actually have figured out that to do this well, you’ve got to do it together.”
The second shift is how health systems are grading the technology. Scrutiny in the C-suite and the boardroom is increasingly about economics, as some AI vendors move from predictable per-seat licensing toward utilization-based pricing tied to compute consumption — a variable cost no one has learned to forecast yet. That shift is arriving as hospital operating margins run near 1%, payer mixes shift and the bulk of Medicaid reductions are set to land across late 2026 and 2027. Systems are also asking whether one platform can absorb enough work to justify cutting others as they rethink their tech stacks.
Both shifts have produced a bar that did not exist when most ambient AI contracts were first signed. Being a capable health IT vendor no longer covers it. The job is to be an AI partner — which is an entirely different job, one with a description that systems are actively learning and writing for themselves.
Interviews with informatics and AI executives at six systems that recently renewed or expanded with Abridge — UNC Health, BJC Health and WashU Medicine, Hartford HealthCare, University of Utah Health, NYU Langone Health and WVU Medicine — describe what that bar looks like from the buyer’s side of the table.
The trial period
Each of these relationships started as a test, but none stayed one.
UNC Health began a small pilot in August 2024 and expanded twentyfold between February and April 2025. It renewed in February 2026 and is now using 103% of its contracted seats. As of June, UNC had 4,803 provisioned users and more than 2.5 million notes generated.
David McSwain, MD, UNC’s system CMIO, attributes the overage to demand he never had to influence. “The level of adoption has been incredible, without having to really push it,” he said. “Adoption and engagement is a challenge with new technologies, and in this case, it’s been one of the easier parts of the process.” He paused. “It is, I think, without question, the most impactful technology that I’ve ever deployed in my career.”
About 350 miles north, WVU Medicine built a curve the same way. The system went live in May 2025 with about 50 clinicians who had already been using a different ambient tool and were crossing over. What followed required no campaign.
“It was natural growth. It was friends telling friends,” David Rich, MD, CMIO of WVU Medicine, said. The system went enterprise-wide that November and has since recorded more than a million notes — close to 13 million minutes of captured conversation between clinicians and patients.
BJC’s pilot began in July 2024 and has since reached more than 2,000 clinicians across all specialties and care settings. Hartford HealthCare signed an unlimited enterprise agreement with Abridge in July for clinical note documentation and nursing workflows, reaching more than 1,600 licensed clinicians across ambulatory, emergency department, inpatient and community settings.
At NYU Langone, Abridge was available in ambulatory practices last September, then the emergency department, then inpatient in July, and now has more than 1,400 clinicians using it regularly. Adam Szerencsy, DO, NYU Langone’s medical director of ambulatory informatics, measures that expansion against 16 years of implementing software for physicians. “You never talk to a doctor and they’re like, ‘Oh my God, thank you so much. Abridge has made my life so much easier. I’m able to get my days back,'” he said. “You never hear that. Never ever.” He is careful not to oversell it — “I’m not saying that this is everyone’s experience” — but he is not ready to downplay the fact that clinicians who once spent hours on notes now spend minutes.
Checking the references
The narrative for health systems on ambient AI to date has been to manage expectations on its financial returns. It’s not a technology meant to make clinicians see more patients, most CIOs would tell you, but to repair a problem healthcare spent years naming. The American Medical Association minced no words: “Getting Rid of Stupid Stuff” was the title of its campaign against needless clerical work.
BJC Health and WashU Medicine took that premise at face value. The health system and medical school enterprise put just under 500 primary care providers on ambient AI and pointedly declined to measure whether it made money. Fix the clinician and patient experience, the theory went, and the finances would follow.
They did, slowly, and on three different clocks. Note-writing time fell about 7% almost immediately, then roughly doubled to 15% by month five — more than an hour back a week. Pajama time, the 5:30 p.m. to 7 a.m. window, moved not at all for a month, then declined steadily. Revenue came last: no gains at first, then a 3% weekly lift in work RVUs among physicians, or close to $2 million a year across the group’s 200-plus doctors.
“At a minimum, the technology is self-sustaining from a financial standpoint,” said Philip Payne, PhD, BJC’s chief health AI officer.
But what changed after hours interests Dr. Payne even more than how much did. Clinicians had been rebuilding and recalling encounters, a “cognitively intense task.” By month five they were editing drafts. Two changes arrived unprompted: more documentation of social factors, from clinicians asking and listening rather than typing, and heavier users adding encounters to their own schedules.
At WVU, adopting the Abridge AI platform cut in-room human scribe reliance by over 30% through natural student attrition rather than layoffs. In clinician survey data supplied by Abridge, WVU users reported 61% less cognitive load, 77% more satisfaction at work, 78% more undivided attention to patients and a 43% gain in accommodating urgent cases.
What it has to prove
AI pricing in healthcare is raising a lot of questions in 2026. “Everyone is very anxious about trying to understand and plan for a shift towards utilization-based pricing models for AI, whether that be tokens or some other indicator of use and volume,” Dr. Payne said. Between tokenization that varies by platform, usage that varies by use case and scaling effects that vary by volume, his assessment was flat: “I don’t believe anybody has solved the problem of accurate and reproducible forecasting of token utilization.”
Systems are hedging in different directions. Utah put several million dollars into an on-premises NVIDIA environment that keeps token economics inside its own four walls. BJC went the other way, choosing a utilization-based budget model and building micro-billing capability to track consumption at a fine grain.
Asked how UNC manages its token exposure, Dr. McSwain redirected to the people. “The most effective strategy is ensuring that you are optimizing how the human beings within our healthcare system are functioning,” he said, “rather than just having AI solutions take over the work in as many areas as possible.”
Shiv Rao, MD, Abridge’s co-founder and CEO, sees it the same way. “We have an opportunity right now not to sprinkle AI over all of these workflows and hope things are just going to get better naturally,” he said in June. “In fact, it can get a lot worse — AI versus AI, agents versus agents, would probably race to some dystopic future nobody wants to live in. Our opportunity right now is to use AI to actually refine, redesign the system.”
That framing sets the terms for everything Abridge builds next. If the return comes from making clinicians more effective, the value of the platform depends on how much of the working day it can reach — and how fast it gets there.
NYU Langone is one health system taking serious stock of vendors’ tempo. “When we look at [Abridge’s] pace of development historically and beyond, we realize that this is the right company for us to move forward with,” Dr. Szerencsy said. “That’s why we significantly increased our number of licenses — because we see the development and what their roadmap looks like.” He has not stopped taking note: “We continue to be surprised at their pace of development. I’ve never seen it.”
Reach is one distinction of a partner. Abridge has described its own strategy as shifting left — solving problems upstream, in the conversation, before they become downstream problems in need of cleanup — while moving the cursor right, out of the note and into the rest of the visit. Every inch of that rightward movement is not just a new job the company is asking to take on, but a new function that health systems decide to trust it with.
Abridge’s context-aware clinical decision support is where capabilities and trust are already being tested. Conventional decision support fires on rules built for populations. Abridge has delivered the capability for clinicians to go granular based on the conversation transpiring in the room and relevant patient history, in addition to a content library anchored by The New England Journal of Medicine, JAMA, the American Diabetes Association, the American Academy of Family Physicians, Neurology, Journal of Clinical Oncology, the CDC, U.S. Preventive Services Task Force and more. If health systems are already customers of UpToDate, Abridge can leverage that knowledge repository as well.
“Now we’re taking decision support from a cohort of patients at a population level and tailoring it to this specific patient in front of me,” Dr. Szerencsy said. At NYU Langone, clinical decision support carries over to the next step. Abridge is now able to queue up orders from the conversation and provides billing recommendations. Custom prompts have pushed the tool past physicians entirely, including a go-live with a physical therapy group spanning 50 to 60 offices.
Nursing is a significant expansion that health systems guard carefully. Nursing documentation lives in flowsheets and discrete fields, structured that way so data can be extracted for quality reporting — a harder engineering problem than narrative notes and a far larger seat count if solved.
Demand is not the obstacle. “The age-old story, right? Doctors can dictate. Why can’t I dictate?” said Rebecca Mitchell-Perry, WVU Medicine’s inaugural CNIO and a nurse of more than 35 years. Her nurses watched the physician rollout and wanted in. WVU kicked off its pilot in June on a single unit, with an October go-live and a six-month evaluation. Ms. Mitchell-Perry has assembled super users and written a mission statement, and is still settling on what to measure.
What interests her most is what the technology might do to nursing thought. Flowsheets arrive with drop-downs and branching logic that shape how a nurse documents and, she suspects, how a nurse thinks. “Speaking and focusing on your patient and talking opens up your mind to think differently,” she said. “[This technology] gives you that space to do that — to use your critical thinking and not focus on just what this prompted for me to document.”
Hiring for a job description that keeps changing
At UNC, Dr. McSwain has developed his own test to spot something that is not evident in any contract.
When a technology company’s leadership team starts talking, he begins drafting the questions he will use to find the Achilles’ heel — whether they actually understand what they are asking a clinical workforce to absorb. He remembers the conversation with Abridge Co-Founder and CEO Shiv Rao, MD, where this particular habit stopped being useful.
“Before I could ask those questions, I remember Shiv addressed them on his own, without prompting,” he said. “That’s where I realized it’s a part of their culture.”
He kept testing it, waiting to see whether it was luck. “I remember the point in the conversation where I realized I don’t have to ask those questions. Because they already answered them,” Dr. McSwain said. “It’s been consistent in the relationship, and that is rare.”
Executives described different traits and tells of a strong AI partner, with shared appetite between systems for a partner who first “falls in love” with the same problems they have long worked to solve before unveiling brilliant solutions. Each leader has a way of testing for fit, and no two are the same.
Dr. Payne starts from the nature of the thing. “AI is not a static tool or platform that you deploy,” he said. “You don’t buy it, deploy it, and walk away and declare success. These are dynamic technologies that change constantly.” Capabilities change, performance changes, the training a workforce needs changes. “You have to take a product view over a life cycle that spans years, not just a few months.”
That is a demanding standard, and most vendors fail it in the first meeting. Dr. Payne noticed that ambient companies almost uniformly opened by demonstrating technology and benchmarking their foundation models, arriving at clinical problems at the end. “Every one of our conversations at Abridge started with them talking to us about what problems we were trying to solve,” he recalls. His live test is what happens when performance disappoints. “Their response is not to try to throw more technology at the problem,” he said. “Their response is, can we do a better job of defining why it’s not doing what we expect?”
Dr. Stein’s test has parts. The most important one is whether the founders truly understand the clinical problem to solve. The second is technical depth. A demo does not move him. Sales team headcount does not influence him. He wants a good, hard read of the people who are actually building the core. “It can’t just be, oh, I’m really good at scribe,” he said. “Well, okay, how competent are you? What’s your core competency in the actual core technology? Deterministic AI, generative AI and agent AI?” The technology is moving too fast to bet on a company that cannot move with it. “I want to be with a partner who can pivot and adapt rapidly.”
The third for Dr. Stein is the robustness of their AI testing and monitoring capabilities. “Safe and responsible AI is key as a balance to unlocking AI’s potential to transform care,” he said. This is not an either-or; he needs all three demonstrated as an imperative before moving forward.
For Ms. Roach, the proof is narrower and harder. Feedback has to change something — and quickly. “We’ve had other AI, especially in ambient listening, where you would give them feedback and nothing would happen,” she said. “I have a very short and very small window to build up credibility with my clinicians. If they give me feedback and you don’t incorporate and fix that, they just won’t use it at all.”
Abridge came out to Utah recently with nothing to launch, but to dig into the next questions: how do you improve this for a sub-subspecialty, move it to the inpatient side, build it for nursing and the therapists behind them? It matters to Ms. Roach, too, that Dr. Rao still sees patients. “He has to eat his own dog food,” she said. “I see him all the time at conferences, and he’s very accessible. He doesn’t have an entourage of people surrounding him.”
Dr. Szerencsy sets the bar for AI partners at a willingness to be told no. “They need to be open and understanding and not just push back and say, ‘You know what, sorry, we know better than you,'” he said. What he describes in Abridge instead is a company that sits with the problem: “They seek to understand our needs, and they’re thinking through them with us rather than just dictating how things should be.” At WVU, Dr. Rich watched the same thing land in the release notes. “They very much wanted to hear our feedback, and have weaved some of our feedback into their interim functionality releases along the way,” he said. “So I think they’re a company that’s eager to listen and learn.”
Dr. Payne traces all of this to the top. “It starts at the top of the company and it percolates all the way down to the frontline folks that we work with in terms of client success and technology every day,” he said. “They’re curious. They always want to know what our experience is. They’re very humble about the technology.”
Bigger than a contract
Ms. Roach’s confidence has edges. Her concern is not whether the notes are accurate today, but what happens when clinicians stop checking them — when the tool has been right often enough that they take themselves out of the loop, and then the models underneath it change, or the data does. At Utah, the physician stays the final check on what gets documented, and she intends to keep it that way.
“It’s a partner with you,” she said. “It is not the decision-maker.”
She applies that standard to each new frontier. Trust in one capability does not transfer to the next. The documentation case may be closed — more than a million notes recorded at WVU, 103% of the seats UNC thought it would need, clinicians at NYU Langone thanking the people who rolled it out. But as the conversation moves on to decision support, revenue cycle, nursing, payer adjudication, clinical trial recruitment and more, Abridge needs to earn each job.
As high as health systems may hold the bar for their AI partners, it is also hard to miss something different unfolding in these organizations. Technology is no longer happening to the people using it. Clinicians, nurses, executives and even board members are rooting for all of this — the pilots, the science, the new use cases and new users, the pre- and post-visit supports, point-of-care guidance, real-time trial matching and payer collaboration — to work. And to win.
“For the first time, we are giving our clinicians not just hope, but tremendous excitement about what’s possible with AI-enabled technology to impact clinical outcomes in a joyous, meaningful, transformative way,” said Dr. Stein. “Seldom in my career have I experienced an opportunity like this.”
At the Becker's 11th Annual IT + Revenue Cycle Conference: The Future of AI & Digital Health, taking place September 14–17 in Chicago, healthcare executives and digital leaders from across the country will come together to explore how AI, interoperability, cybersecurity, and revenue cycle innovation are transforming care delivery, strengthening financial performance, and driving the next era of digital health. Apply for complimentary registration now.
