Durham, N.C.-based Duke Health spent nearly a decade waiting for the technology to catch up to an idea, and when large language models finally arrived, the health system built the tool itself rather than wait for a vendor to do it.
The tool is Scout, an AI platform developed by the Duke Institute for Health Innovation that reads through a patient’s entire electronic health record, including notes going back years, labs, imaging reports and pathology reports, and synthesizes that information into a summary in one to three minutes. It lives inside Epic, Duke’s EHR system, so physicians access it from within their existing workflow.
“We saw a problem in healthcare where we have information overload,” Jeffrey Ferranti, MD, chief digital officer of Duke Health, told Becker’s. “Our doctors have electronic health records, but those records can be thousands of pages long, and synthesizing and organizing information in a way that makes sense to physicians, so they can take the very best care of patients, has always been a challenge.”
Scout has been an idea at Duke for awhile, but the health system lacked the technology to build it until generative and agentic AI matured enough to make it possible.
Every claim Scout generates links back to its source in the chart, letting a physician verify the underlying record rather than take the summary on faith.
“If it says the patient had an X-ray or an MRI with a certain finding, there’s a link back to the source record in the EHR, so the doctor can review that,” Dr. Ferranti said.
Physicians reach Scout through a tab inside Epic, and access is limited to clinicians who already have a treatment relationship with the patient. For many specialties, Duke has built predefined prompts tailored to specific tasks, an oncologist preparing for tumor board or a transplant physician compiling a summary can pull up a prompt built for that purpose, and Duke tracks how well those prompts perform in practice.
Physicians can also write their own prompts, asking Scout to pull a patient’s history on a specific medication class, for example. Duke has identified roughly 50 use cases it is testing across pre-visit planning, discharge summaries and tumor board preparation.
Before any physician gets access, they have to complete training on the tool, a requirement Dr. Ferranti described as central to how Duke deploys Scout responsibly.
“There always needs to be a physician in the loop,” he said. “At the end of the day, Scout is helping you synthesize data, but you, as the treating provider, need to make the ultimate decision on how that patient is best treated and validate the information you’re getting from Scout.”
Duke tested Scout with about 100 physicians before expanding, then took it enterprisewide in early July. Any physician or advanced practice provider at Duke who completes training and is credentialed in the organization can now use it.
That expansion followed a randomized controlled trial, still in preprint and not yet peer reviewed, that Dr. Ferranti said found clinicians completed tasks roughly 40% faster using Scout, with cognitive workload down 40% to 50% compared with manual chart review.
“We found that the summaries that Scout is creating are as good or better than our senior clinicians,” he said. “That was done in a randomized controlled trial across multiple service lines.”
Building Scout to live inside Epic, rather than as a standalone tool, was a deliberate choice. Duke has used Epic’s SMART on FHIR framework for years to build custom panels into the EHR, and Dr. Ferranti said he wanted to avoid a setup where physicians had to toggle between separate systems.
“I didn’t want to have Scout in a separate window where people swivel chairing between the electronic health record and Scout,” Dr. Ferranti said. “We wanted it to be in the place where physicians are spending their days, and that’s why we decided to integrate it into Epic. It also takes care of a lot of the security and governance concerns because the only people who can use it are people who have access to Epic.”
Duke chose to build Scout in house only after concluding no vendor or partner offered anything comparable. The health system generally defaults to buying an existing tool if one meets its needs, Dr. Ferranti said, and turns to co-development with an outside partner as a second option. Building from scratch through the Duke Institute for Health Innovation is the last resort, reserved for cases where nothing else exists.
“There were no tools in the market that could do this. There were no partners that were working on this, and we decided to build it ourselves,” Dr. Ferranti said. “It is based on a secure way that we can access data in the electronic health record that we’ve been working on for the better part of a decade. Layering a large language model on top of it and making it secure and safe for our providers to use was not as heavy a lift as you might imagine.”
Duke plans to extend Scout beyond physicians into disciplines including clinical psychology and social work, and to expand from clinical use cases into operational ones. Dr. Ferranti said the health system has learned that results improve when a use case is narrow enough to let Duke refine the prompt behind it.
He described Scout as the first in what he expects to become a broader system of specialized AI agents at Duke, covering ambient documentation, chart summarization, medication reconciliation and interpretation of outside records, with the harder work ahead being how to govern and coordinate them.
“We’re going to see providers fortified by an army of AI agents, and we want those agents to be trustworthy and unbiased and orchestrated in a thoughtful way across our health system,” Dr. Ferranti said. “Scout has caused us to really start thinking about how we do that safely, and how we oversee that, and how we deploy governance and orchestration at scale.”
At the Becker's 11th Annual IT + Revenue Cycle Conference: The Future of AI & Digital Health, taking place September 14–17 in Chicago, healthcare executives and digital leaders from across the country will come together to explore how AI, interoperability, cybersecurity, and revenue cycle innovation are transforming care delivery, strengthening financial performance, and driving the next era of digital health. Apply for complimentary registration now.