Peter McCaffrey, M.D., M.S., vice president and chief digital and AI officer at UT Medical Branch (UTMB), an academic health science center and health system based in Galveston, Texas, recently spoke with Healthcare Innovation about his perspective on agentic AI. Rather than looking for agents for individual use cases, he thinks it’s important to take a platform view. “It's about articulating the workflow and allowing intelligence to be placed there efficiently alongside people,” he said. 

UTMB now has 22 agents in place in clinical use cases such as triaging patients for access to specialist visits and improving lung nodule detection, as well as in use cases such as pharmacy benefits prior authorization, clinical documentation integrity and supply chain. For all these efforts, UTMB is using Carebricks, Bunkerhill Health’s agentic AI platform. Besides UTMB, Bunkerhill is working with several major health systems, including Intermountain Health, Endeavor Health, Cleveland Clinic and Mayo Clinic. 

Joining the conversation was Azita Hamedani, M.D., M.P.H., M.B.A., the new chief clinical officer at Bunkerhill Health. Most recently, she served as president and CEO of UNC Faculty Physicians and vice dean for clinical affairs at the University of North Carolina School of Medicine. Previously Hamedani spent nearly two decades at the University of Wisconsin School of Medicine and Public Health and UW Health, where she founded the Department of Emergency Medicine and chaired it for over a decade while also leading system initiatives related to quality and innovation.

HCI: Dr. McCaffrey, could you talk about your role and philosophy around agentic AI? 

McCaffrey: I have a very broad mandate around AI in the enterprise. It covers everything from HR and parking to diagnostic tools to research grants and predictive success. I would say the guiding light that I see for AI is if we look at the field of healthcare overall, there are a lot of jobs to be done that are just frankly not being done…There’s a tremendous amount of context management, synthesis, packaging, and surfacing that permeates every workflow in healthcare. The failure to realize that brings patient harm, and managing that is an expectation of good care. We see agents playing that role, and that is not a decision-making role. Decisions can work well, though, if that role is handled ahead of time correctly. 

HCI: I understand that one of UTMB’s first agentic AI deployments involved screening imaging for coronary artery calcification. Could you talk about that use case?

McCaffrey: That was our first population-level AI use case that we deployed outside of things that are clerical like ambient listening. This involves the idea that coronary calcification is a biomarker of risk. It's not an emergency. It's not a thing you have to deal with right now, like an aneurysm. It is a meaningful context point for the chart. But it is classically a context point that you only know about if you look for it — because a physician thought you might have this risk based upon talking to you and then ordered a calcium study. By definition you're only looking where you expect to find something, so it doesn't really serve the role of a population health tool. With this different workflow, which is only possible with the efficiency of something like AI, it doesn’t have to be contingent upon someone thinking of it. It’s a systematic way to do this. So if you have a chest CT done and there's a heart in the image, that can be evaluated for coronary calcification, even if the reason for the CT was lung cancer screening or vertebral trauma. This can give the clinician more context to decide around.

Hamedani: Can I expand on that for one second? It may or may not be standard of care for that calcium score to be noticed, reported, and acted upon. Right now there is this floor of negligent care, and then there's the level of care you and I would want for our loved ones. The difference is what agentic AI can fill, so that everyone gets the same care.

HCI: Dr. McCaffrey, how did UTMB start working with Bunkerhill?

McCaffrey: We were interested in this idea of AI in the screening segment, and looking at incidental coronary artery consultation (ICAC) as an example of screening. So we first came to them through that as probably the preeminent example of the FDA-cleared algorithm for ICAC. We deployed that, and that naturally led to discussions around navigation and how we handle this….That was the same time that Carebricks was coming to life, so we said let's see if agents can help do the work of care navigators, for example. And from there it has just sort of grown by horizontal grassroots use inside the enterprise.

HCI: Dr. Hamedani, from your bio I saw that you'd worked a long time in emergency medicine. In your work, what did you see about the promise of AI, and particularly with this company Bunkerhill?

Hamedani: Before I chose emergency medicine, I developed an interest in quality and operations and safety. When I was a medical student, Don Berwick came and spoke to my medical school, and this was before “To Err is Human” or the Quality Chasm or IHI. He gave some poignant examples of how bad systems can lead to bad care. He is famous for the saying: “Make it easy to do the right thing, hard to do the wrong thing.” 

So when I when I was thinking about what specialty to pick, I picked emergency medicine because it's so systems-based, right? We're constantly thinking about how to build systems for such diverse patient needs, diverse acuity, and trying to do right by all of them. Honestly, when the rest of the system fails, we're the ones who see it as well. I had talked to a lot of different AI companies. But what I love about Bunkerhill's agent is it closes the care gaps, both in obvious and nuanced ways, by not just highlighting, there's a dropped ball or there's some process that agentic AI can now help with, but even acting end to end. As our CEO likes to say, it's not just about saying here's the list of all the patients that need follow-up, but they go ahead and take that next step at a time when labor shortage is a real issue for most health systems. What I uniquely like about Bunkerhill is they're almost like that back office scaffolding, not necessarily at the point of patient care, but making sure all the things the patient needs done happen.

HCI: Could you give an example of how the agents could be applied in the emergency department?

Hamedani: Sure. One of the banes of my existence in the ED was radiology incidentals. Whether it was a lab or an imaging study, when I started the saying was, "You order it, you own it.”  That meant that you, as an individual physician, would need to wait for the final read, even if it's three days later. If there's anything positive, call the patient yourself. Arrange the follow-up. As you can imagine, with the exponential increase of imaging in the emergency department, that became prohibitive. So I ended up having to create a nine-hour PA shift, 365 days a year, to comb through radiology reports. Fast forward to today, Bunkerhill's agent could just take care of all of that. The PA can go back to patient care, and they would welcome that. That's one of their most common use cases.

HCI: Dr. McCaffrey, I understand that improving access to specialists is another area where Carebricks is making a difference. Could you describe that? 

McCaffrey: We’ve seen Carebricks access agents as a really transformative thing here at UTMB. Think about referrals to medical specialties such as nephrology. As a health system, you are often in a spot where you have a backlog of referrals. In many cases, it is just a first-in, first-out queue. The person who typically schedules that isn't the physician who's thinking about your clinical state, and they may not totally understand the details of your chart. 

Who is there to advocate for you when that click happens? Who is going to say, ‘Hold on, this person has real concerns that have been voiced about them and documented in their chart. They need to be at the top of the list’? Typically nobody. But the agents actually comb this cue of everyone sitting in this backlog for nephrology, and provide that access technician insights in their workflow as they're scheduling about this referral. How acute is this? How complex is this? Should this be prioritized? Should it not be? And we see the impact. For instance, a colonoscopy referral with anemia. That’s concerning, and not something to sit on. Now these folks have a lower time in the queue, sometimes as low as half of what it was before. Those slots are going to people with higher-urgency findings. It's not like it's a clinical decision that wasn't made or should have been made and made at a different place by an AI. It's that the context that the referring doctor probably knew just doesn't appear in this workflow. And without AI, there's not bandwidth to surface that.

Hamedani: Peter mentioned first in/first out. There is even a little bit more of an uncomfortable reality in terms of what it takes to get that person with anemia a colonoscopy. It could be heroic effort from a primary care doctor — so much of our system's built on this heroic effort — or a VIP. So if the system isn't built to do right by who needs it, it requires either heroic effort or it requires VIP access sometimes to make it work.

HCI: Dr. McCaffrey, if the potential for these agents is so high, how do you prioritize which things to go after first?

McCaffrey: A lot of it boils down to a cultural way of viewing AI use. For example, if the use case has to be heavily evangelized by some top leader down to the clinical team — it’s not that it can't happen, but it's not a great sign that that will be successful. If instead you can involve the service line owner, that is likely to be more successful. For example, one of the things we do a lot with the aid of agents is our lung nodule navigation. Is there a nodule? How big? Is it growing? Is it not? Is it concerning? How do I book them? Carebricks is a platform that helps us do that. The lung nodule service director is an active champion. So part of it is engaging not necessarily the executive at the top of whatever chain, but the service line owner, getting their buy-in and getting their participation in the design and development of an agentic workflow. That's a key piece. 

Beyond that, it involves having a set of values in AI. Is it basically raising the minimum bar for care? Making it less likely that you will fall through the cracks? Bringing visibility and consistency to your workflow? Improving the safety of what you experience? Then it's in the basket of relevant things. Within the basket of relevant things, is it feasible and do we have that buy-in? That's our paradigm. If there's further bandwidth beyond that, we just say in quarter one we do this, quarter two we do that, quarter three we do that, and we sequence it.

HCI: Does any of the AI governance work happen at the UT Health System level or does it all happen at the local level?

McCaffrey: It happens at both. If I can be a little bit biased with my Longhorn hat here, I think UT does this in a really unique way that's powerful. Every health institution — UTMB is one, UT Dell Medical Center is another — has our own leadership and governance structures. We also have a collaborative effort called UT Real Health AI, which spans all of our institutions and it involves executive leadership and technical leadership in AI. It's actually appropriated by our board of regents in our state, and one of the mandates of that group is to create common methods, playbooks, inventories, and approaches for governance and monitoring, and to facilitate discussion, but get to precision. 

We have legislation that says you need to evaluate before you deploy and detect bias. But this goes beyond that to say let's actually commit to a way of doing this. Let's all do it this way. Let's share with each other what we do. We've even gotten to a point now where we have a common maturity framework that we share with each other. So we have gelled a lot. Governance happens at the site level, but a lot of the way we do things, the orthodoxy that has evolved around it, has become systematically consistent, and that's really nice, because it creates a common language and a mutual de-risking of the path forward.

HCI: Anything else you would like to add about working with AI agents? 

McCaffrey: What I have found about agents that's so interesting, and I think the biggest argument for a platform in general, is that agents become a segment of the workforce at a certain point. That's kind of how we view them, and the workforce need is growing. Because they are agentic, they can handle instructions and do workflows. They are very polymorphic, as you might imagine. They can be used in many areas, so we view this less as there's a tool for this and a tool for that and more as it's about articulating the workflow and allowing intelligence to be placed there efficiently alongside people. And the many use cases are just flavors of doing this.