Assessing the Potential Impact of Agentic AI in Radiology: An Interview with Nina Kottler, MD

While convolutional neural networks, multimodal AI and large language models have dominated the discourse thus far with the evolution of AI in radiology, agentic AI represents more of an adaptive system-level approach to achieving goals, according to Nina Kottler, M.D., MS, FSIIM, FAIM.

“(Agentic AI) uses foundation models as its base but the whole goal is a system-level goal. … It's going to pursue a sort of defined goal. You give the AI a goal, and then it will select the actions that it needs to take based on the output that it's getting, data that it has access to and feedback on that system to see how well it's actually achieving that goal. Then, really importantly, it adjusts its behavior based on that feedback. That's what makes it agentic,” explained Dr. Kottler, the chief medical AI officer for Mosaic Clinical Technologies.

In a recent interview with Diagnostic Imaging, Dr. Kottler said she foresees an evolution with worklists moving away from the rules-based systems at the orchestration level to more agentic worklists with a degree of autonomy in triage decisions. That said, Dr, Kottler emphasized the importance of transparency with agentic AI to foster trust of the radiologist.

“I want some transparency at the point of care. I want to know why is it triaging this case to me? Why is it assigning it? Why did it move things around? So you need some kind of transparency and visualization into that black box. This is important because … if the human doesn't trust the (AI) system, they're not going to work well together, so we need to make sure that we're calibrating trust and reducing the inappropriate or unconscious biases,” emphasized Dr. Kottler.

That said, Dr. Kottler maintained that agentic AI provides an important active agility to assess various data and AI model outputs and make corresponding changes to achieve overarching goals at a system level.

“Frankly, in medicine, there's so much unknown. It's just (that) we live in a world where there (are) all kinds of broken links in the system, and you can't necessarily predict what's going to happen. So, you can't create rules to define what to do in those scenarios. That's where agents are great. You want them supervised to make sure they're making good decisions, but those agents can react more thoughtfully to things we can’t foresee,” posited Dr. Kottler.

(Editor’s note: For related content, see “What’s Coming Down the Pike in Radiology in 2026?,” “The Inflection Point for AI in Radiology: Emerging Insights for 2026” and “Emerging Directions with Advances in Enterprise Imaging in Radiology.”)