Agentic AI at Work Inside Health Systems: Lessons from UTMB

How UTMB turned more than a hundred clinical ideas into goals its AI agent sees through, and what its leaders learned about scaling, governance, and trust.

In a recent webinar with the Scottsdale Institute, leaders from the University of Texas Medical Branch (UTMB), one of Bunkerhill’s deepest deployments, sat down to talk about what happens when a health system stops buying a point solution for every problem and instead gives one long-horizon AI agent a growing list of goals to pursue. Bunkerhill is live at more than two dozen health systems. UTMB shared how it got from a handful of pilots to an agent working alongside its staff every day.

Representing UTMB:

• Dr. Peter McCaffrey, Chief Digital and AI Officer at UTMB, a pathologist by training who oversees the system’s AI intake and deployment.
• Dr. Salim Hayek, Chair of the Department of Internal Medicine at UTMB, a cardiologist and physician-scientist.

A live poll set the stage. Asked where their organization stands with agentic AI today, roughly 23% of attendees said they were still exploring, 50% were piloting one or two use cases, and 27% were already scaling across the enterprise. It is a snapshot of an industry that has moved from curiosity to real adoption in under a year.

How does a health system scale from one AI pilot to twenty use cases?

The short answer: treat the agent as an extension of your workforce, and build a repeatable path from idea to production instead of starting a new project every time. At UTMB, that path turned more than 100 submitted ideas into more than 20 live use cases, with Dr. McCaffrey expecting that number to reach 40 or more within a year.

His central point is that most work in a health system does not fail because it is hard. It fails because there is simply too much of it for people to get to. When a patient “falls through the cracks,” that is usually a polite way of saying a job needed doing and no one had the hours to do it.

Rather than reinvent a technology stack for each use case, UTMB builds on what is already in place: the way the agent reasons, how it looks and feels, how it reads an HL7 feed or writes back into the EHR. Every new connection becomes a permanent part of what the agent can do, so a new use case is an iteration, not a fresh procurement cycle. That is what lets aortic-aneurysm detection, infectious-disease triage, and case-mix-index optimization all run on the same agent.

One example made the stakes concrete. UTMB’s first deployments included screening imaging for coronary artery calcification, a finding often noticed on scans ordered for something else. In a case later published in Texas Monthly, the agent surfaced a patient with a heavy calcification burden and little prior cardiology contact. The follow-up that resulted led to a CT angiogram and, ultimately, a three-vessel bypass. Not every case ends that way, but without the agent paying attention, no one was paying attention to that patient at all.

Can AI fix patient access without hiring more physicians?

Yes: by routing the right patient to the right physician through the right modality, instead of scheduling first-come, first-served. Dr. Hayek framed patient access as the most common problem shared across health systems, and one that hiring alone cannot solve.

He had seen the old playbook up close. At the health system where he previously worked, wait times to see a cardiologist stretched to six months. The fixes were familiar ones: Saturday clinics, after-hours slots, telehealth, and physician incentives. They worked, but they were expensive, and they still funneled low-complexity cases into specialty clinics. By some estimates, roughly 30% of specialty care is low-value and could be handled elsewhere.

The deeper flaw was triage, or the lack of it. When scheduling runs first-come, first-served, urgency is ignored entirely.

UTMB’s answer was to give the agent a clear goal defined by its clinicians: get every referred patient to the right care at the right time. For each incoming referral, the agent assesses three things: complexity, urgency, and the best modality to see the patient, whether in person, by telehealth, or by e-consult. Crucially, this was not bought off the shelf. Every specialty built its own playbook, each division assigned an AI champion, and those champions review a random sample of charts every month to validate the agent’s work and refine the rules.

The outcomes, across several subspecialties in internal medicine:

• Nephrology: high-complexity patients were seen roughly twice as early as before, while 10 to 15 e-consults a week resolved lower-complexity questions in 48 to 72 hours instead of a months-long wait.
• Endocrinology: a service that looked like it needed three to four more endocrinologists began meeting its full demand with the existing team, simply by shuttling the right patients to the right modality.
• Colonoscopy: evidence-based risk scoring prioritized patients at higher risk of colorectal cancer, surfacing a meaningful number of adenomas and interventions earlier than the queue would have.

The principle underneath all three: when you pull a simple question out of the queue, you move a high-urgency, high-complexity patient to the front.

What did UTMB learn about deploying agentic AI?

Beyond the individual programs, UTMB’s leaders shared a set of hard-won principles for making an AI agent stick inside a health system.

Let demand pull, don’t push

UTMB did not recruit champions so much as watch them volunteer. Once clinicians saw the agent solve a colleague’s problem, they arrived with their own. About 90% of the people driving use cases nominated themselves.

Governance is one front door, not a maze

UTMB runs a single enterprise-level AI governance process. The agent is vetted once; each new use case gets its own approval, with an expedited path when it closely resembles one already live. There are no shadow deployments. Every use case is inventoried, tied to a champion, and held to a time-bound KPI it must meet to stay in production.

Clinicians own judgment; the agent owns execution

Clinical criteria always come from the physicians and staff who live the work. The forward-deployed team builds those criteria into how the agent pursues each goal and how it acts, whether through HL7, FHIR, SMS, or a write-back into the record. Dr. Hayek was blunt that a boxed product a technical team designs alone will not survive contact with patient care, and that a leader drafting the first playbook is what gets everyone moving.

Match the rigor of validation to the risk

UTMB validates the agent continuously: each division’s AI champion reviews a random sample of charts every month to check the agent’s work and refine the rules. For lower-risk decisions like referral triage, Dr. Hayek noted, two physicians can reasonably disagree, so the bar is not perfection. It is beating how the system schedules today. “And we’re beating it by far.”

Treat the vendor as an extension of the team

UTMB’s leaders described Bunkerhill’s forward-deployed model as an extension of their own staff: the same people, embedded for the long haul, who understand the culture and sit in on the recurring work of tuning the agent. As Dr. McCaffrey put it, one central contract is fine, but without many touchpoints into the organization, momentum rarely lasts.

Key takeaways

1. One agent beats a dozen point solutions. Giving one agent a growing set of goals turns each new use case into an iteration, not a new procurement cycle.
2. The agent does the work no one had time for. Most “falling through the cracks” is simply undone work; the agent extends the workforce to reach it.
3. Access is a triage problem, not only a staffing problem. Routing by urgency, complexity, and modality let UTMB meet endocrinology demand with its existing team and see high-complexity nephrology patients twice as early.
4. Clinicians lead, governance is centralized, and rigor scales with risk. Physician-owned criteria, one front-door approval, time-bound KPIs, and validation matched to the stakes are what make the agent durable.