Advances in artificial intelligence are moving from machine learning models that analyze data to algorithms that can act on it with minimal human involvement across clinical, administrative and patient-facing workflows.
Agentic AI, or autonomous agents, level up from standard generative AI by being able to make decisions autonomously to achieve a goal, rather than just producing output from an input to aid human decision-making or streamlining tasks.
The growing momentum behind agentic AI in healthcare was a hot topic HIMSS25, from the show floor to education sessions and keynote addresses. But all the discussion pointed to a key challenge for healthcare: Using agentic technology to improve operations is one thing, but carefully navigating the serious complexities of employing autonomous AI in clinical care is quite another.
AI agents can help physicians meet the demand for care, as accelerators and automators of habitual tasks. But they’re not a replacement for physician decision-making. As healthcare organizations look to agentic AI to handle more complex, multi-step workflows within the EHR and beyond, care must be taken when deploying these tools in clinical settings.
Health AI governance and policy
A controversial bill introduced in the U.S. House of Representatives in January and referred to the House Committee on Energy and Commerce. H.R.238 would amend the Federal Food, Drug, and Cosmetic Act to enable AI and ML that’s been authorized by the Food and Drug Administration to qualify as a practitioner eligible to prescribe drugs.
But many technology developers believe autonomous AI isn’t ready to be used in prescribing and other aspects of care
Several clinical and IT leaders we spoke with before, during and after HIMSS25 cited the importance of building trust and ensuring transparency in its functionality as the biggest hurdles to adoption in care management systems.
And at HIMSS25, AI experts urged caution. At the HIMSS AI in Healthcare Forum earlier this month, for instance, Dennis Chornenky, chief AI advisor at UC Davis Health, pointed out that governance has not kept pace with the technology’s advancement, and that current health AI regulations were designed for machine-learning models – not for autonomous AI decisions.
Move too fast with AI agents, and the result could be biased decision-making and potential risks to patient safety. Healthcare chief information officers are under pressure to adopt them, “but they’ve got to ensure safety” in preparing for the AI and automation disruption to come, Chornenky said.
“How do they do that if they don’t have the proper governance mechanisms set up?”
Agentic AI and patient follow-up
A primary driver for adopting agentic AI across all organizations is to alleviate the computer workload on clinicians and administrative staff.
Epic is integrating AI across applications and positioning agents to improve clinical efficiency, Seth Howard, executive vice president for R&D at Epic, told Healthcare IT News ahead of the HIMSS25 global conference and exhibition.
“For instance, we expect AI agents to help with pre-visit prep by chatting with patients about their needs, identifying missing tasks (such as labs), helping schedule and complete those tasks and creating an easy-to-read summary,” he said.
Garrett Adams, Epic’s vice president of research and development, used the example of a post-surgical patient assistant agent at the conference.
A multimodal follow-up from a wrist surgery that uses voice and image reaches out to the patient, “chatting with them in a conversation way,” he said. It asks the patient about the wrist recovery and to share an image of it or a video of it.
The agent can deduce the angle the patient can bend their wrist back post-surgery and compare it to patients at the same point in their recovery in Epic’s Cosmos database. Based on that point in their recovery, the agent suggests that a follow-up appointment may no longer be needed, because the patient is “trending incredibly well.”
The agent asks the patient if they want to cancel the appointment, and sends the patient’s request to cancel to the care team to confirm.
“So still having that human verification for that follow-up stuff, make sure the guy’s right,” said Adams.
While the post-surgical patient assistant agent is still with the patient, it can also offer a flu shot appointment.
“Combining the video input, patient communication, conversational tone, Cosmos data insights, care gap closure and scheduling functionality in the agentic workflow — that patient had a pretty seamless interaction,” he said.
Also at the HIMSS25 this month, Zoom workplace for clinician launched a public beta of ambient notes for use during telehealth appointments as well as for in-person appointments.
“When you are done with the appointment, as a doctor, you quickly get the note, and then you review and edit,” said Ritu Mukherjee, Zoom’s head of product, business acceleration and readiness, in an interview on the HIMSS25 show floor. “And then with the press of a button, you just send it to the EHR system,” she said. “This is just the starting point in automating the clinical workflow.”
Developing agentic AI with caution
Meanwhile at HIMSS25, longtime interoperability and IT infrastructure leader InterSystems debuted a new EHR powered by AI called IntelliCare. In addition to ambient listening and generative AI features to simplify administrative tasks, an AI assistant can generate automatic patient history summarizations and agentically prepopulates the codes required for billing within the revenue cycle management system.
The vendor is taking a cautious approach to using agentic AI in direct clinical decision-making, emphasizing human oversight and focusing initial efforts on administrative tasks. For example, adding to the diagnostic list.
“My feeling is that agentic AI and open source models are the two topics of the year for 2025, from an AI point of view,” said Don Woodlock, InterSystem’s head of global healthcare solutions, ahead of HIMSS25.
There is potential to put a bigger dent in clinicians’ administrative burdens by automating tasks like sending letters, processing orders and handling prior authorizations, by enabling AI to think about the necessary set of steps and allowing it to pull them together, he said. It’s going beyond the ambient note, “doing follow-up actions and updates to the chart and we’ll see how far we get.”
For higher-order workflows in the EHR, developers may “not necessarily” push the agency of AI, he added. “What are the pros and cons of that? What are the dangers? Are we removing a level of doctor oversight?”
With automating processes like sending prescriptions and referrals, eClinicalWorks CEO Girish Navani said he is more inclined to pump the break on AI agents handling them.
“There’s a time and place where there will be agents even in clinical decision support,” Navani told Healthcare IT News at HIMSS25
“But that’s not the problem statement for today. We’re focusing on revenue cycle management, taking away the administrative overload and the work that is done today – which has no clinical decision-making – and creating autonomous agents to work while everybody else is sleeping.”
eClinicalWorks showcased practical applications of agentic AI, such as its AI document intelligence tool that automatically extracts patient data from incoming documents – in PDF, fax or CCDA formats. The company said it can accurately match data to patients for 75–85% of these documents. There’s also the healow Genie agent which can provide patients with instant answers to common inquiries 24/7 by voice, text or chat.
However, for complex questions, Genie escalates queries to human agents.
Athenahealth has also been gradual over several years in how it integrates advanced features into its AI-enabled products. Recently, the company has offered its EHR customers ambient listening, summarization tool and Salesforce’s Agentforce, which offers agentic workflows for in patient visit summaries, missed appointments, care gaps and pharmacy orders.
In clinical workflows, AI agents automate provider tasks like lab result follow-up and appointment prerequisites.
But Alicia Bassolino, athenahealth’s vice president of analytics and AI, also stressed the importance of human oversight.
“I view AI, agentic or not, as the ability to remove some of the mechanically oriented thinking, the task creation, the things that they have to enter into the system versus the thinking that is very much their decision making and their application of medical knowledge,” she said after HIMSS25.
RCM driving agentic AI
Now that healthcare AI technology has progressed to a place where it can make decisions autonomously, several EHR vendors are bolder on implementing the administrative advantages in RCM.
“You can really orient the generative AI around a task or a goal, and it can navigate through that autonomously or pseudo-autonomously with a set of parameters that you give it,” Bassolino said.
Automation across a broader spectrum of across payers, providers, specialties and other revenue cycle permutations is now possible with AI, and vendors are looking to tackle that friction.
“It allows us to work with payers or automate certain types of tasks where there’s a set of options that aren’t perfectly linear and don’t necessarily flow,” she said.
Navani said eCW is also heavily focused on implementing agentic AI in its RCM products.
“We’ve got a large RCM group ourselves, a thousand employees,” he said. “We came across seven or eight agents that could basically start doing the work what we have off and then run it. The productivities are ridiculously off the charts. You can talk about an entire group of 50 people that do something else now, because you can have the agent do that work.”
In addition to autonomously processing electronic and paper-based data coming in, the potential to manage claim coding verification and compliance checks on its own is enormous, he said, noting the company will have more data on efficacy later this year.
Andrea Fox is senior editor of Healthcare IT News.
Email: afox@himss.org
Healthcare IT News is a HIMSS Media publication.