Running AI Inside a Top Clinical Environment: What Duke Taught Us

How do you earn the right to run AI inside one of the country's most regulated clinical environments? Here's what Duke Health taught us.

In cardiology, there’s no room to get AI wrong. Every decision touches a real patient, every system is bound by HIPAA, and every process has been shaped by clinicians with years of hard-earned experience. When Manesh Patel, MD, Chief of Cardiology at Duke University Health System, decided to bring agentic AI into his division, we needed to prove they could be trusted to perform within a system that has zero tolerance for guesswork.

Earning that right is a tall order; we started with fax routing.

The Fix

Our team started by sitting down with Duke's clinical team and asking dozens of questions: Which systems do you use? How long does each step take? What drives you crazy?

From there, we collaborated with Duke to build a structured framework to evaluate every AI use case before deployment. Designed to be repeatable across any regulated environment, the framework established an impact card defining what to measure and what baseline looks like; a model card documenting the data and the interface; a validation card tracking whether early metrics matched expectations; a monitoring card watching for drift over time; a pricing card making the cost case; and an evidence card mapping the path to publication.

Before deploying, we established two things. First, agents built on Trase Origin, our agentic operating system, could access systems the same way employees would, and would ask for help when uncertain of what action to take. Every decision gets logged in an immutable audit trail built to meet HIPAA and other clinical compliance standards. Second, agents could use Duke's existing workflows. Thus, all Trase agents are trained on Duke's own clinical data and institutional knowledge, informed by staff who have been doing this work for years.

The First Proof Point

As our first use case, we picked fax routing, a notoriously time-consuming workflow that places immense administrative burden on clinical staff. Duke’s Division of Cardiology receives more than 5,000 faxes a month, and for years, every one of them landed in a queue that nurses and medical assistants spent nearly 1,400 hours a month sorting by hand.

On a Monday morning in April 2026, with more than 800 faxes waiting in the queue, Trase's fax routing agent went live. By the end of the day, it had cleared the overwhelming majority of them.

Deployment did surface one real challenge: Faxes for patients with left ventricular assist devices (LVAD), essentially mechanical heart pumps, look very different from standard cardiology referrals. The training data didn't cover that edge case well, and the agent initially misclassified them. The Trase team caught the error, built targeted training for LVAD cases, and closed the gap. That feedback loop is the same process that will enable this system to take on other high-stakes clinical work, demonstrating that this is not a one-time fix for one edge case, but a scalable solution.

What that first use case proved out:

• Fax triage ran 7.1x faster
• 1,395 staff hours returned per month
• Accuracy rose from 70% to 98.1%
• $285,450 in annual staff capacity freed up and reinvested in patient care
• 250%+ ROI achieved on deployed agents

Those hours moved so that nurses who used to spend Monday mornings on paperwork could now see patients.

"When we first turned on the Trase fax routing agent, it processed the lion's share of them by the end of the day. It liberated our team to focus on the work they actually trained for and enjoy doing."

— Manesh Patel, MD, Chief of Cardiology, Duke Health

What Makes This Hold Up Under Pressure

Underneath any single agent is a micro-agent architecture — specialist agents for classification, patient matching, and routing that feed into the larger workflow. The system is modular, so any piece can be improved or replaced on its own. Above that sits a two-track learning system. When staff confirms or corrects a decision, that feedback directly trains the model. Agents running in the same environment also learn from each other, much like a cardiologist and an endocrinologist covering the same patients who start to pick up each other's clinical instincts over time.

Cost and accuracy are controlled through small, specialized language models trained specifically on Duke's data, not general purpose models running up a token bill on routine work. Drift gets tracked continuously through evals run against synthetic test cases, watching both how often the system flags items for human review and how often it produces a questionable result. Trase and Duke review the data together and use it to tune the model, catching regressions before they reach production.

The framework that matters most going forward is simple: separate low-risk, high-certainty actions from high-risk, high-uncertainty ones. Routing a fax to the right place is a low- risk task. Flagging a patient's worsening heart failure is a different story entirely. That use case demands the same evidentiary bar as a new therapeutic intervention, not just a good accuracy score. That is the long-term vision behind the Trase-Duke partnership: creating a pathway that enables AI to operate in high-stakes environments in a way that is not only effective, but also demonstrably trustworthy, predictable, and safe.

What We're Building Next

The fax agent was our entry point. Prescription refill automation, appointment scheduling, and in-basket triage are already moving through the same evaluation pipeline. After that, the work turns to chronic condition management, heart failure, hypertension, and high cholesterol, where agents follow patients beyond a single visit to check in, coordinate care, and surface what matters before it becomes an emergency.

Because the underlying governance layer is already in place, a new health system doesn't start from scratch. The compliance and audit infrastructure, the learning fabric, the framework for proving trust before granting access — all of it transfers. That's how we can take a new deployment and go from initial scoping to production in 30 days.

This is what it actually takes to get autonomous systems working in an environment where mistakes can have devastating consequences, and to build a track record of discipline and governance to earn the access they’re given.