The healthcare industry’s focus on AI in medicine, and the Food and Drug Administration’s plan to clear 1,000 new tools this spring, may suggest that AI in healthcare is pervasive, but that is not necessarily the case.
“While we’re seeing a lot more AI-enabled tools, we are seeing variable adoption of these tools within actual clinical workflows,” according to Sonya Makhni, MD, MS, MBA, medical director of applied informatics at Mayo Clinic Platform and hospitalist at Mayo Clinic in Rochester, Minnesota.
AI-supported administrative solutions for clinical documentation, coding, and other uses are seeing more extensive use. Uptake in clinical care is the highest in radiology, the field that has by far the most FDA-cleared AI tools, some of which have validated benefits in improving efficiency or reading accuracy.
As for the hundreds of other AI tools for clinical care on the market? Makhni explains that “the reasons behind limited adoption of these solutions are multifactorial.” They include a lack of evidence that demonstrates meaningful benefit in real-world use, challenges to understanding the risk for algorithmic bias, and interoperability or integration challenges that can be costly to overcome.
Indeed, Elsevier’s most recent Clinician of the Future survey, including 2,206 survey responses from 109 countries in 2025, found just 16% of respondents were using AI tools to actually help make clinical decisions.(1) Still, an additional 48% expressed a desire to use AI to support clinical determinations, and 41% agreed that, within two to three years, “clinicians using AI tools will deliver higher quality care than those that don’t.”
The inevitability of AI being ever more important to healthcare delivery is indisputable, says Robert Wachter, MD, professor and chair of the Department of Medicine at the University of California, San Francisco (UCSF).
“We probably couldn’t hire new doctors here if we said we don’t have an AI scribe. It’s almost become an expectation,” he says. “I think that’s going to become true for more and more different tools.”
Wachter, author of A Giant Leap: How AI is Transforming Healthcare and What That Means for Our Future, co-chairs UCSF’s strategy committee. “I don’t think it’s over-hype to say that in five years, if you have not implemented AI thoughtfully and effectively, you’re going to be behind other institutions that have done so because you will not be able to deliver access, quality, safety, patient experience, or efficiency in a way that patients expect, and your competitors probably will be able to do.”
ARE PHYSICIANS EAGER OR WARY?
Makhni, with her clinical practice and expertise in applied informatics, concentrates on responsible AI design, development, deployment, and monitoring of AI solutions within clinical care. In working with other clinicians, she finds many are open to using AI if it can help them and their patients, but there is still uncertainty in how to effectively leverage tools.
“Based on my experience, clinicians want to solve problems, and they welcome innovative solutions that can help them address the very real clinical and operational challenges they face day-to-day,” she says. “However, they need to be able to trust these solutions. And these solutions need to lead to real positive impact.”
Claudine Lott, MD, physician executive for clinical solutions commercial transformation and implementation for Elsevier, which sells an AI-powered clinical decision support solution, says, “Any time you’re introducing something new into the workflow of a busy clinician, there’s going to be challenges, there’s going to be resistance, and you really do need to expect that and have a plan in place.”
Physicians are not change-averse so much as they are overburdened. “The barriers to any new technology for healthcare providers are that they are so pressed for time, they have so much else on their plate, and the stakes are so high,” Lott explains. “Tech solutions really have to have a meaningful, positive impact on users to be adopted.”
Additionally, usability and workflow integration are essential to adoption. “For a lot of the physicians we have spoken to, we find that ‘if it’s not easy for me to access in my workflow, it’s not going to be adopted because it’s just giving me an extra step,’ ” Lott says.
She offers one anecdote that illustrates the point: She asked a clinician who had access to various clinical decision-making tools how they decided which one to use. “And this clinician said, ‘Whichever one is already open — that’s the one I’m going to use.’ ”
Another success factor is human relationships that support the technology. That means clinical champions who can explain why they consider an AI tool worthwhile and collaborative partnerships with technology vendors. Vendors must be able to explain how the tool was developed and use clinician feedback for continuous improvement.
“When we have users say, ‘This is great and here’s something else that I would love to see,’ that really helps build trust in the product and helps the clinicians to adopt it,” Lott says.
The person or department that should be in charge of managing AI adoption depends on the organization, according to James I. Merlino, MD, executive vice president and chief operating officer for The Joint Commission. “The transformation of healthcare with AI has near limitless potential, and while the promise is great, like any new technology that touches patients, we want it to be safe,” he says.
“I don’t think every organization needs to have a chief AI officer,” Merlino continues. “It’s important to have a designated individual or group that has, as part of their responsibility, monitoring these tools and helping to guide decisions around how they are used.”
Of course, healthcare organizations have varying levels of expertise and capacity to introduce and monitor AI tools, and The Joint Commission wants to support those who need help.
“Everyone doesn’t possess the same level of resources, and one of our responsibilities is to ensure that we are getting good information out to everyone, especially the organizations that don’t have people who can dedicate 100% of their attention to it — level the playing field with information, so to speak,” he adds.
Some organizations are offering their AI expertise to providers who need help. For example, Mayo Clinic Platform, an initiative of Mayo Clinic, provides an ecosystem that connects innovators, clinicians, and health systems to advance the use of AI in medicine.
As AI tools flood the market, leaders must make careful decisions about which ones to acquire. “There is a fine balance between vetting or testing solutions and disseminating them widely into practice. I usually recommend investing enough time and resources to ensure that there’s proper alignment and outcomes,” Makhni says.
Like any other technology, a return-on-investment (ROI) analysis should be conducted before a decision is made to move forward. Equally important is getting input from the physicians and others who will use the solutions.
“Clinicians right now are being inundated by all sorts of new tools,” Makhni says. “We need to set up sustainable processes that include the various stakeholders so that we can better understand if these tools can fit well in their workflows and ultimately position them to successfully drive positive impact.”
At UCSF, leaders are evaluating AI tools and incrementally introducing a few. “When we can say ‘The AI is good enough, the problem is important enough, and we think it might make a difference,’ we roll it out in a limited way,” Wachter says. “We will not mandate it.”
In the case of AI scribes, for example, UCSF set up a competition between two tools, each of which was introduced in a few clinics, so leaders could compare the quality of notes each provided and physician acceptance. About 70% of its physicians are now using ambient listening.
“Beginning with something that is pretty low-hanging fruit, let us gain a level of acceptance,” Wachter explains. “Winning hearts and minds was actually quite important.”
GOVERNANCE STRUCTURE NEEDED
Elsevier’s survey of clinicians found that only 29% of respondents believed their organization was providing adequate AI governance.
The use of AI tools is a shared responsibility between individual physicians, who are ultimately accountable for the care they deliver, and their organizations. “Institutions definitely have a role in helping to define what are approved tools and what are some guardrails, without stifling the innovative spirit that drives medicine,” Merlino says.
Organizations should have an infrastructure in place that supports the use of trustworthy tools and helps physicians understand which tools are better than others. “It’s ‘Here’s why this tool is concerning to us in terms of decision-making or security of patient information, and here’s why we think this is a better choice,’ ” Lott says. “There needs to be a shared understanding between the user and the institution.”
And organizations need to know what AI tools are being used by their clinicians. “There’s risk with AI because it’s still new,” Merlino says. “Organizations have a responsibility to be mindful of what they are implementing and make sure that they are collecting any potential information regarding harm that is caused by it.”
Dictating which AI tools a physician can use — or must use — is not realistic, Merlino says. “I think AI is out of the box, and it’s running really fast and accelerates every day, and we should not try to stop that — we can’t stop that. The job of a leader is to make sure that you’re getting your people to pay attention to what’s important and you’re helping to guide those discussions.”
TRAINING NEEDS
While AI training opportunities for physicians are increasing, exactly what training physicians need and how it is best provided is not yet settled.
“From my perspective, it starts with providing healthcare professionals with the right amount of knowledge and practical tools so they can understand what solutions are available and how to vet them,” Makhni says.
For most AI tools, the “how to” is fairly straightforward, but understanding when to trust AI is not. Lott shares, “As an end user, you shouldn’t have to be a data scientist or an AI expert. But you do need an understanding of how the tool works, what it is, what it isn’t, what it can and can’t do, just to be able to have that trust in it. I think that this is a role more institutions are going to start providing.”
Perhaps the most important priority is making sure that clinicians understand when —and when not — to accept AI responses. “That’s where the physician expertise, the nursing expertise, and the experience of all healthcare professionals comes in because you want to be sure that the information that you’re using to assist in making decisions make sense against your training and what you bring to the delivery of care,” Merlino says. “It’s a tool to help solve problems, but it is not a tool to replace professional judgment and experience.”
Wachter has come to use AI as a curbside consultant, calling on its input rather than seeking out a specialist as he would likely have done in the past. His experience shows that using AI as a tool to support effective patient care requires the knowledge of a physician.
“When I put into the AI, ‘This is a 62-year-old woman with a history of lupus who comes in with a fever and a white count of 12.3 and infiltrates on chest X-ray,’ knowing that prompt is a pretty advanced cognitive act,” he says.
Equally important is the clinician’s ability to evaluate AI responses. “When [AI] comes out with diagnoses, you’ve got to know a lot of medicine to say, ‘Oh, that’s a really good idea — I hadn’t thought of that,’ or ‘That’s ridiculous. I’m not even going to think about that because it’s so off-base,’ ” he says.
Reference
Elsevier Health. Clinician of the Future 2025. Elsevier. 2025. https://assets.ctfassets.net/o78em1y1w4i4/T7F5sDDiUC8KJzLQXfJoy/004be7f43562d318115a294cf626be7f/ClinicianOfTheFuture_2025.pdf