Health systems are moving from AI copilots to agents. Here’s where they’re starting

Advertisement

For the past few years, most health system AI has worked as a copilot. It drafts notes, summarizes charts and suggests codes, and a person checks the output before anything happens. The next step is AI agents that complete multistep tasks on their own.

At Becker’s 14th Annual CEO + CFO Roundtable this month in Chicago, health system technology leaders described where they are starting with agents. For most, that’s the back office, revenue cycle and a few narrowly defined clinical tasks.

Nabile Safdar, MD, chief AI officer of Atlanta-based Emory Healthcare, said the shift will push health systems to move faster than they’re used to.

“I think we have to transition in healthcare from chatbots and more kind of traditional embeddings of AI [in] EHRs or ERPs, et cetera, to figuring out how we can actually do agentic workflows, and the agents might be partially human, and maybe AI as well, and what that combination is, because the pace of that change is not a pace that we’re frankly used to operating at in healthcare,” Dr. Safdar said.

Jeff Gautney, senior vice president and CIO of Chicago-based Rush University System for Health, said most AI today still behaves like conventional software.

“The future is definitely autonomous agents. We’re in this stage right now where, yes, it’s AI, but it works a lot like our existing software. You buy a product, you bolt it on, you feed it and it puts out stuff,” Mr. Gautney said during a panel discussion at the conference. “But that’s not the future that we’re going to be in two years from now, where products are broken down into molecules and the molecules operate autonomously. And some of those molecules are developed by users and not by IT people anymore.”

At Louisville, Ky.-based ScionHealth, which operates mostly long-term acute care hospitals, the starting point is administrative work. Geoff Roten, chief technology and information security officer, said the organization is investing heavily in Microsoft tools for 2027 to modernize its technology stack and drive down cost.

“We’re actually making an intentional decision to move some of our back office operations to be more autonomous,” Mr. Roten said.

New York City-based Mount Sinai Health System is building agents inside its EHR. It is among the roughly 20 early-adopter organizations Epic is training to build with its Agent Factory platform, said Lisa Stump, executive vice president and chief digital information officer. Mount Sinai is starting with one clinical agent, for medication reconciliation, and one revenue cycle agent, she said.

Mount Sinai’s AI team is also building agents into its internal analytics. Leaders can now ask a dashboard follow-up questions and get new charts drawn from raw data, without filing a request with the data team, Ms. Stump said.

“So we built — it still looks like a dashboard — but then in that moment you can ask the AI agent, ‘Tell me more about that trend over the last three months at the level of our payers,’ and it’ll go back into raw data, and boom, here’s a new graphic by payer,” she said.

“A year from now, I think we’re going to look back on [static dashboards] the same way we look back today on the paper medical records.”

Scottsdale, Ariz.-based HonorHealth is looking at more mundane work. Its efforts to automate parts of its call center, which has about 20% turnover, have struggled to produce a return, according to Jim Whitfill, MD, senior vice president and chief transformation officer. Fax processing looks more encouraging.

“Right now we’re actually focusing on a lot of ancillary workflow, for example around processing incoming faxes. We have collections of people having to manage just that work,” Dr. Whitfill said. “That’s something that we’re seeing some good promise with, hopefully, being able to do some automation around. Because I don’t think anybody wants to sit for eight hours a day managing and running faxes.”

Dr. Safdar expects systems to push agents next into operations that cut across departments.

“I think what we’re challenged with is orchestration layers that are kind of touching more deeply specific systems and workflows, whether that’s scheduling or capacity command or things like that, and I think that’s where a lot of systems are going to be going next,” he said.

Vendors say the savings depend on getting humans out of routine steps. Ogi Kavazovic, CEO of HouseRx, which sells AI-driven specialty pharmacy services to clinics and health systems, said during a panel that his company’s prior authorization tool looked about 90% effective on paper. Real-world failures, such as payer forms that couldn’t be retrieved and chart notes stuck in draft, cut that to roughly 35%, he said, and it took about two years to work back into the 80s.

Even then, productivity initially dropped because staff double-checked the AI’s work. Costs fell only after the company stopped requiring staff to check routine cases, Mr. Kavazovic said.

“The way to get to dramatic cost reduction is not through AI enablement but through autonomy,” he said.

Handing AI that kind of autonomy raises security questions, Mr. Gautney said.

“These tools are weaponizable very easily in a lot of different ways. From a cyber standpoint, it’s going to be crucially important that they ask those questions and understand not who the human in the loop is, but who’s the adult in the room that’s making good decisions about how much power you’re handing over to, in many cases, unskilled folks.”

Advertisement

Next Up in Innovation

Advertisement