"AI agents will become the primary way we interact with computers in the future.” 
                                                                                     Satya Nadella, CEO, Microsoft

I have been asked to give several talks on agentic AI in healthcare as it is a popular topic these days in meetings of AI in healthcare. Here are some of my thoughts on agentic AI: 

Since early 2025, there has been considerable interest and activity in the domain of agentic AI in healthcare. Automation in healthcare was previously explored this past decade in the hyperautomation form of robotic process automation (RPA). While RPA automated repetitive, rule-based actions especially in back-office tasks, agentic AI makes decisions, plan, and act with autonomous processing that yields multi-step actions in the form of a personal assistant. In short, agentic AI systems extend beyond deterministic automation of RPA by deconstructing healthcare goals into steps. While agentic AI holds considerable promise for healthcare, especially given the amount of administrative and operational burden, there is ongoing understandable concern regarding its safety due to its autonomous characteristic. 

Potential uses of agentic AI are numerous, but include primary care workflow and care coordination (scheduling followup appointments, monitoring medication adherence, conducting post discharge outreach, and sharing patient education), generation of patient-friendly reports, ambient scribing to ambient clinical intelligence (with insights, care plans, and quality improvement), and even healthcare education and training (AI pedagogy with both teaching about AI and leveraging AI for education). Perhaps it is prudent for agentic AI to start with reversible, single-agent, nonclinical projects (such as patient education or hospital marketing) and then slowly progress to more nuanced multi-agent applications (such as care coordination or clinical pathways).  
 

Here are my three major takeaways on agentic AI in healthcare based on three AI dimensions (architecture, evaluation, and system): 

Architectural: Agentic AI with multiple agents in healthcare has an inherent design problem. A single patient may have a list of agents- primary care physician, cardiologist, payer, pharmacy benefit manager, and remote monitoring manager, all with distinct utility functions. In the real world, there are often misaligned objectives and preferences. The question is “Which agent loss function dominates when there is disagreement?”. 

Evaluative: Agents are autonomous policies that mark a phase change in clinical AI. While AUC for a sepsis classifier can be computed, the equivalent metric for even a relatively simple agentic system that titrates inotropic agents, calls for consults, and adjusts orders with some degree of autonomy does not currently exist. The FDA 510(k) paradigm strained to accommodate models but cannot similarly accommodate agentic AI in healthcare.  

Systemic: Tight coupling and interactive complexity of multiple AI agents in healthcare can yield emergent failure. The agentic AI systems need to be tested perhaps in the framework of non-clinical systems such as those seen in financial systems, not FDA or health systems. The Perrow normal accident theory reminds us that health systems are complex, not complicated, so serious accidents are structurally inevitable (hence “normal” or expected). 

 

The fascinating world of agentic AI in healthcare will be a major topic of discussion throughout the AIMed26 meeting this coming November 10-12 at the sublime Renaissance Resort at SeaWorld in Orlando, Florida. A new feature this year will be AIMedX, an immersive experience of AI in healthcare. See you there!