Health systems have spent years rolling out AI tools that draft notes, summarize charts, and answer patient messages. The next wave is different: AI agents that can take actions on their own — placing orders, flagging results, closing gaps in documentation — without a person approving each step.
Some CIOs told Becker’s the industry isn’t ready to watch over them.
“If we’re going to build agents in a healthcare environment, and we’ve got agents crawling across our workforce … we’ve got to manage them. We’ve got to monitor them. We’ve got to make sure that they’re not drifting and that they’re doing what they’re intended to do,” said Eric Neil, CIO of Seattle-based UW Medicine. “They’re not consuming too many tokens, there aren’t duplicate agents — a system to manage the agents is incredibly important.”
Mr. Neil compared it to managing a new hire. “You hire an employee, you train that employee, and then they have a manager who monitors their performance and what they’re doing on a daily basis,” he said. “When we build agents in the workforce doing some of the work for us, how are we going to do the same thing with them? You’ve got to have platforms to do that. You can’t build it, launch it, and forget about it.”
That oversight will likely be some mix of human and automated, he said — but either way, it can’t be an afterthought. “We’ll have systems that’ll help humans keep an eye on it, but I think it’s just as important as the human portion of our workforce,” he said. “If you don’t get that in place up front, it’s going to be incredibly difficult to come back in later when you have agents crawling across your environment, automating things and assisting people.”
Right now, access to actually build those agents is tightly controlled. Epic’s Agent Factory platform lets health systems “take” an existing agent, “shape” it to their own organization, or build a new one entirely — but only a small number of health systems are doing that work today, alongside Epic’s own team. “Everybody who’s building agents today is building them with Epic, because everybody’s learning how to do this,” said Rebecca Mishuris, MD, chief medical information officer and vice president of Mass General Brigham in Somerville, Mass. Her system was among the first to build agents through Epic’s early pilot, called Factory One, developing ones focused on radiology actionable findings and emergency medicine.
Her colleague, Eric Podradchik, vice president of digital clinical systems at Mass General Brigham, said that tight control needs to hold as access opens more broadly. “Agentic AI … is the most exciting thing, and not surprisingly, also the biggest risk,” he said. “We have to build governance around it before we expand.”
Broader building access is expected to open to more health systems by the end of the year — a timeline Epic itself confirmed at its keynote — which Mr. Podradchik said leaves a narrow window to get the guardrails right. “We have to build a governance [structure] to be able to understand what has been built, govern it, make sure it’s validated, and also measure the effectiveness and efficacy before deployment,” he said.
Not every CIO frames it as an unsolved problem. Ryan Smith, chief digital and information officer of Intermountain Health in Salt Lake City, described Epic’s approach as building governance into the system by design rather than bolting it on afterward. Agent Factory, he told Becker’s earlier in August, “gives organizations a governed way to build their own agents inside the system, addressing any number of potential use cases.”
The concern echoes a broader theme in how CIOs are reacting to Epic’s rapid AI expansion this year. Aaron Miri, DHA, chief digital and information officer of Jacksonville, Fla.-based Baptist Health, has said getting the “sophistication and maturity” to monitor new capabilities matters more than how fast they roll out. For AI agents specifically, several CIOs said, that maturity doesn’t fully exist yet — for Epic’s tools or anyone else’s.
Part of what makes agent oversight harder than monitoring a single AI tool, Mr. Neil said, is scale: a chatbot that gives a wrong answer is one bad interaction, but an agent empowered to take actions across thousands of patient charts can compound a mistake before anyone notices. Tracking that activity also means tracking cost — duplicate or malfunctioning agents can quietly run up token bills as easily as they can cause clinical errors, a concern that dovetails with the budgeting questions several CIOs have raised about Epic’s consumption-based AI pricing.
For now, the industry’s answer to who’s minding the machines is still being built alongside the machines themselves.
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.
