Problem
Electronic health record (EHR) adoption has improved access and communication but, for care practitioners, has also created an exponentially growing volume of patient portal messages. Within Duke Endocrinology, hundreds of MyChart messages arrive each week, many of which are administrative in nature but are initially sent directly to physicians or advanced practice providers. To address this, the division established a nurse-led triage team that intercepts all incoming messages, responds when appropriate, and forwards others to the correct destination. While this solution protects providers from unnecessary inbox volume, it has shifted a substantial workload onto the triage nurses.
Over time, the expanding scope of triage responsibilities on top of consistently high message volumes has led to backlogs and delays in responses to patient concerns. Nurses must review every message before determining where it belongs, even when it concerns scheduling, prior authorizations, forms, or other non-clinical requests. This manual sorting process consumes valuable time and limits the team’s ability to focus on urgent patient needs.
Solution
Our team partnered with Endocrine Specialty Engagement Center to develop an AI-assisted classification system for inbasket message routing. Working closely with triage nurses, we defined message categories that reflect real clinical workflows and built a flexible framework that predicts the most appropriate destination for each incoming patient message. The system uses large language models to interpret the intent of a message and assign it to the correct operational category.
A key strength of this approach is its flexibility. As clinical workflows evolve, categories can be easily refined or expanded, enabling continuous improvement as operational and patient care needs change. Ultimately, we aim to augment clinical insight by providing early insight into message intent and reducing the cognitive load of manual sorting.
February 2026
Over the past year, we have partnered with Duke Health Technology Solutions and Endocrine clinic triage leadership to integrate our Duke-built AI-assisted classification system for inbasket message routing directly into Epic. The system now runs in real time within the Endocrine Specialty Engagement Center InBasket pool, automatically labeling each message with a predicted category. Triage nurses can see this classification before opening the message, giving them immediate context about its likely purpose.
While we are actively working to enable automated routing functionality, real-time labeling has already changed how the team manages messages. Messages can be sorted by predicted category, allowing staff to batch similar requests and quickly surface time-sensitive concerns. The tool operates quietly in the background.
Impact
Since implementation, the triage team reports meaningful improvements in efficiency and responsiveness. The system helps them rapidly identify high-priority or potentially urgent messages, flag cases that require return calls, and forward administrative requests to the appropriate teams without delay. By previewing message intent before their review, nurses are able to focus their attention where it is most needed rather than spending time manually sorting routine requests.
The team has also observed a significant reduction in turnaround time for patient messages. What previously averaged approximately three days has decreased to roughly one day. This faster response improves patient experience, accelerates access to medications and services, and reduces the operational strain associated with message backlogs, particularly after weekends or high-volume periods.
Next Steps
In February 2026, we are working closely with the triage team to formally evaluate real-world performance and quantify the time and effort savings generated by the system. Ongoing feedback from nurses is helping us refine categories, address edge cases, and ensure consistently high accuracy in operationally important message types. This phase will provide a clearer understanding of the measurable impact on workload and patient care.
In parallel, we are collaborating with Epic to enable automated routing for low-clinical-complexity messages such as scheduling, prior authorizations, and forms. Achieving safe and reliable auto-triage would complete the original vision of the project. At the same time, we are exploring opportunities to partner with additional specialties to develop customized AI-assisted triage solutions tailored to their unique clinical workflows.
