AI agents are already being used in healthcare. But what are they, and how do they differ from other forms of healthcare AI?
AI agents perform tasks autonomously, rather than a one-and-done response to a prompt, such as large language models like ChatGPT. “Agents differ in that they take some action,” said Nigam Shah, MD, PhD, chief data scientist of Palo Alto, Calif.-based Stanford Health Care. “It could be as simple as making a call, teeing up an order or creating a summary to review.” Their complexity varies widely, he noted.
Like humans, AI agents follow “execution loops,” said Anthony Chang, MD, chief intelligence and innovation officer of Orange, Calif.-based CHOC Children’s: “think, plan, act, observe (outcome), reflect, adjust, repeat.”
They are emerging in healthcare for tasks such as clinical documentation and workflow automation, operations optimization, patient communication and research support, Dr. Chang said. He pointed to some companies in the space: Hippocratic AI (for low-level healthcare tasks), Aidoc (for radiology workflows), Qventus (for hospital operations) and Insilico Medicine and Benevolent AI (for scientific discoveries).
Epic and Oracle Health are both building AI agents into their EHRs for tasks including clinical documentation, prior authorizations and care gap closures. “That agent finds that a patient is behind on their vaccines or immunizations, and that agent can send a notification to the patient via MyChart that they might want to get a vaccination appointment scheduled, because it’s gone through their chart and identified that care gap,” said Epic Director of AI Sean McGunigal.
Charleston, S.C.-based MUSC Health has deployed AI agents from health tech company Notable to complete 40% of prior authorizations without human involvement. What often took about 30 minutes of manual work is now executed by AI in about one. The health system is also using the company’s AI agents for appointment reminders and check-ins and care gap notifications, scheduling and completing over 2,000 mammograms in the process.
In a recent staff meeting, new CMS Administrator Mehmet Oz, MD, reportedly promoted the use of AI avatars over human physicians in some instances, saying they could save money while providing equal or better care.
The future evolution of AI agents in healthcare will necessitate both comprehensive evaluation and technological advancement. AI agents require “some of the most advanced AI technologies to date,” Dr. Chang said.
“In the long term — assuming we figure out evaluation — it is possible to imagine agents doing all sorts of tasks on behalf of both patients and doctors: making appointments, summarizing multimodal records, making treatment suggestions, etc.,” Dr. Shah said.
