What health systems large and small need to launch AI agents

Photo: Kan Kingpetcharat/Getty Images

Health systems are being warned that, even though they may have the budget it, they may not be ready for the realities of deploying autonomous artificial intelligence agents.

Complex requirements around compliance, security and validation need sorting out. And providers – especially those pursuing agentic AI driven only out of fear of missing out on the next big trend – need also to beware of the AI expertise shortage. And they could face significant risks thanks to "data drift" that reduces agent accuracy.

"The ones chasing the FOMO will arrive at a platform they are not prepared to use," said Dr. John Lee, an emergency physician and Epic consultant at HIT Peak Advisors and Taryn Shipley, an Epic architect at Sutter Health and principal at Lean Health Tech, in a recent opinion piece about readiness to use Epic Systems' Agent Factory.

That said, for those organizations that have "done the data governance work, built the team depth, and designed their workflows," success with Epic agents can be had, they say.

Two other health systems have shown how that's true..

Advocate Health, a multi-state health system running two of the largest Epic electronic health record instances in the world, says it had the scale and internal technical muscle – including an already-established internal AI and machine learning studio – and is now live on Agent Factory with autonomous workflows for inpatient pharmacy and infusion chart prep.

Meanwhile, ECU Health, a regional academic network in eastern North Carolina has a smaller IT team that did not have the resources to develop agents on its own. It now has two Epic-built prototypes successfully in use, focused on patient transfers and discharges. While the largely rural system will eventually need to take over maintenance of its AI agents, being an early pilot partner allowed it to implement custom AI for the first time.

These agents, all deployed in May, show what different healthcare organizations on Epic are doing to create custom autonomous AI in Agent Factory that helps to address long-standing issues of organizational friction.

In interviews with Healthcare IT News, Andy Crowder, Advocate's chief data and AI officer, and Jacob Parrish, ECU's vice president of clinical operations, both say that creating and deploying autonomous agents requires realistic expectations, an education in how to navigate the new math that is token budgeting and processes for responding to feedback in order to foster clinical trust.

Earlier this year Epic moved beyond integrating AI copilot assistance to agentic AI with the rollout of Agent Factory, a platform aimed at enabling healthcare organizations to build, customize and orchestrate their own autonomous AI agents that are fully integrated directly into the core Epic EHR environment.

The no-code tool with its visual sandbox allows developers to use pre-made native agent templates or invent and test custom agents, which are all automatically tailored to their healthcare organization's policies, standards and clinical protocols, the company said.

The key feature is that Agent Factory agents can orchestrate existing communication and operational tools in the Epic EHR without the need for external APIs or middleware with drag-and-drop functionality.

Using it makes agent deployments seamless, Crowder said in a recent discussion with Healthcare IT News about the fundamental considerations for approaching no-code AI in healthcare.

Integration happens "in the same pane of glass in the clinical workflow in the stride of how your clinicians do work, as opposed to giving some FHIR API or some other integration or some other pane of glass that they now have to navigate to get the benefit of the solution," he explained.

Most low-code tools also require added data protection, whereas Agent Factory has all the requisite access security integrated, Crowder added. Epic's no-code platform also contains monitoring tools to track, trace and audit autonomous agent decisions.

The third benefit, Crowder had said, is access to all of Epic's data sets. Neither a data abstraction step nor a data repository are needed.

Supporting patient transfers and rounds workflows

Epic developers have been working directly with a handful of health systems that participated in the company's annual build-a-thon this year, including rural ECU Health in North Carolina, Mayo Clinic, Sutter Health and Loma Linda University Health.

ECU now has two agents live on custom Agent Factory flows that Epic built for them – location validation and case summarization for the ECU Transfer Center and discharge planning for case managers' progression of care rounds, or POCRs.

While the health system had been ready to go live in April, it had to delay one month for a scheduled upgrade of its Epic EHR.

"We realized we were a version off," Parrish told HITN.

An agent serving the Transfer Center team – about 14 people – parses data inside its EHR module to generate highly focused, three-sentence summaries that follow a strict "capabilities grid" detailing which of its hospitals can perform specific clinical tasks, he said.

The summaries prevent incoming shift nurses from having to review full charts manually. Instead, the agent digests the knowledge and prompts the nurse with its patient transfer recommendation.

According to Parrish, the tool has been highly accurate, with zero hallucinations so far.

It's also shaved off roughly 20 hours of chart review time per week, saving the equivalent of 0.5 FTE in its first month alone, he said.

"It was so structured and so very specific," Parrish noted. "It's not doing full chart reviews and throwing that against ChatGPT to render some clinical opinion. It's limited data from the EHR that feeds this."

Despite the successful metrics, ECU has chosen to keep its AI agents focused on narrower source data, for now, and away from patient interactions.

"We're not even letting it listen to the phone calls," said Parrish. "We thought that it could, but we decided not to do that yet."

The POCR tool for case managers was rolled out system-wide across all nine ECU Health hospitals and is available to inpatient nurses to explore.

Parrish said multidisciplinary teams meet daily for these rounds to discuss barriers to patient discharges. Some cases might require extensive chart reviews before or during the rounds meetings.

The new agent Epic helped ECU develop generates a high-level chart summary meant to help the full team of about 160 case managers to quickly comprehend the patient's entire stay without missing important details buried in notes.

After deployment, changes needed to be made quickly in Agent Factory, according to Parrish.

On the first day of use, case managers overwhelmingly gave the tool a "thumbs down" in the automated feedback pop-up because the summaries were too long and read like a regular chart, he said.

That night, the IT team refined the prompt constraints – narrowing the summary down to key points – and by day two, 75% of the case managers returned a positive rating.

While the ROI on that tool is much harder to define, Parrish said the agent has greatly enhanced the quality of discharge discussions. Ultimately, case managers use it as a pre-round prep tool in the 30 minutes before meetings and they walk into rounds ready.

"It enhanced the quality of the discussion because historically, the case managers or the pharmacy, they come in with their printed list, their Epic report of this patient, this diagnosis, this length of stay, this disposition, this pharmacy, this whatever," he said. "They would read it, and they would handwrite their notes going into the conversation."

"Now they can read this very high level summary and then they can document the things that they know they need to do going into the round, versus trying to do that live."

Improving pharmacy and infusion experiences

Advocate Health has developed autonomous agents for two key areas – inpatient pharmacy and infusion chart prep – with Epic.

With hundreds of pharmacists verifying medication orders each day across the health system, new inpatient pharmacy agents perform as intelligent investigative assistants to lower pharmacists' cognitive burdens.

A medication order verification AI agent combs through a patient's medical history, home medications, current clinical activities and progress notes on their care journeys to generate summaries.

It gathers up all the vital cross-referenced information so the pharmacist can quickly verify that it is "the right medicine at the right time," greatly reducing the time spent manually hunting for background data, Crowder explained in a follow-up conversation about Advocate's specific Agent Factory use cases.

"At Advocate Health, we have hundreds of pharmacists every day that verify every medication order that gets ordered by a provider in the inpatient setting to make sure that it's the right medicine at the right time," he said.

Infusion chart prep and automated patient outreach agents are tackling notorious administrative overhead in oncology and other therapeutic infusion services at the Midwestern health system.

By automating backend triage and outreach, agents are drastically lowering nursing staff tasks, but it's also improving the patient experience, he added.

"Anything that you can do to streamline that, and simplify that for the caregiver who has multiple patients to see to get ready, and for the patient, can be significant improvements in the patient experience."

Advocate's medication order verification agent is also helping the infusion experience, condensing the timeline needed to safely mix and prepare treatments like chemotherapy based on patient's real-time blood work.

"Those can be stressful situations for patients and their families when they come in for those, especially depending on the diagnosis of the treatment plan," said Crowder.

Through a MyChart portal integration, an agent communicates directly with patients.

Before a visit, an autonomous agent standardizes clinical needs weeks in advance, ensuring patients are ready for their treatments when they arrive. They receive instructions regarding pre-appointment protocols, such as whether to fast and which home medications to take or hold off on.

On Tuesday, Part 2 of this two-part series will focus on architectural strategies for monitoring autonomous AI agents and the challenges of token budgeting.

Andrea Fox is senior editor of Healthcare IT News.Email: [email protected]Healthcare IT News is a HIMSS Media publication.