Clinical decision support adoption is fragmented across healthcare, and varies widely among different health systems, new research from Duke University Health System shows.End-users often struggle to see artificial intelligence tools' impact directly, often leading them to stop using AI-enabled CDS. But hospitals that move away from a static list of implementation rules toward a more dynamic implementation process can improve tool adoption over the long term, Duke researchers found.They developed models that describe two forces at work: "balancing loops," where success reduces urgency and effort, and "reinforcing loops," where perceived visible benefits encourage continued tool usage."If you change a parameter, it leads to very different outcomes," said Scott Rockart, associate professor of practice at Duke University, and co-author of the report, "Learning From the Adoption of a Readmission Clinical Decision Support Tool: Group Model Building Approach." "The mathematical model can help test the logic of your explanations and identify more effective interventions," he explained to Healthcare IT News.Mapping use of predictive toolsThe researchers examined the Unplanned Readmission Model, version 1, developed by Epic Systems at Duke Health. It analyzes electronic health records – everything from diagnoses to medications – to generate a risk score that indicates whether a patient is likely to be readmitted after hospital discharge.They looked at training, workload, incentives and team dynamics, and studied how these variables interact over time to understand how adoption evolves under different conditions and which strategies can improve adoption."CDS adoption is not static and changes according to dynamic systems of behaviors and workflows, requiring a deeper understanding of how evolving conditions affect implementation and outcomes," the researchers said in their report, published recently in the JMIR Human Factors journal.They spoke with case managers, physical and occupational therapists, hospitalist faculty physicians, and resident physicians who participate in decisions about discharging patients."We were trying to move beyond 'it's a good tool and it should be used' to 'why isn’t it being used?'" Rockart said.They guided study participants "to identify and connect variables in causal loop diagrams.""We coded workshop transcripts in software designed for system dynamics analysis to identify themes, aggregated them into a causal loop diagram, and reviewed them with participants to converge on a common model," the researchers said. Then, they applied equations and tested data to simulate conditions leading to full, limited or no adoption of the tool.Factors diminishing AI useIn the face of conflicting demands, CDS tools may fall short of sustained long-term use. "When you have something like the Centers for Medicaid & Medicare Services putting pressure on you to achieve certain outputs, and you start to achieve those, obviously, the pressure falls, and there are so many things we'd like to do well, attention shifts elsewhere," Rockart said.CMS penalties have pushed hospitals to reduce readmissions, and institutional incentives lead to tool training and high early use, the researchers noted.Their simulation models show how pressure and training led to initial success, and then interest waned as workflows improved and readmission rates approached goals. "When conflicting priorities [such as COVID-19] were introduced, adoption stalled earlier, and fewer staff were trained," they said. Staff turnover and workflow changes can also lead to reduced usage.These "balancing loops," driven by external pressure, like the CMS readmission penalties, might motivate initial adoption, but regulatory momentum doesn't always last, Rockart explained.During an initial rollout, staff go through intense training and may understand the tool deeply and know how to use advanced features.But newer staff who were not part of the initial rollout may only have a surface-level understanding of a tool.For example, "the tool gives a rating – yellow, orange, green or a red type of rating – they might have been aware of the tool's existence, but then unaware of the fact that if you get a red rating, you could click on it and it would explain the main factors driving it," Rockart said.Of note, users can fully customize their Unplanned Readmission Model tool dashboards in Epic, and often they marginalize, minimize or hide the AI tool if they don't immediately see its value, he said. While forcing the tool into a bold, centralized position on the screen could prevent it from being ignored, he stressed that health systems run the risk of staff blindly clicking through it without actual engagement."One of the things I'd obviously love to have is data on who was using it, how often they were using it, how deep did they go – did they drill down in the tool rather than just see that it was red or green or orange or yellow, and the difference between that, and did they click on it and figure out why it was that?" Rockart said. 'It's not always obvious'When the researchers modeled "reinforcing loops," they found that predictive tools are more likely to have extensive and sustained usage when those using them perceive benefits to care delivery and quality."Reinforcing loops emerged when staff described clinical utility, such as improved discharge planning and team communication," the researchers noted. "However, staff also suggested that these loops were often weak due to difficulty linking the use of the tool to outcomes in real time."Reinforcing loops may be weak for rare events, like unplanned hospital readmissions, making it difficult for frontline staff to see the immediate value of using the AI tool. When a tool prevents a rare event, success might be invisible in the moment. "It's not always obvious, as you're using a tool, how much value it's actually created for a rare event that itself is somewhat complicated," Rockart said. Ultimately, the success of a technology depends on whether people experience it as visibly useful and worth their continued attention over time.AI adoption simulator could helpRockart emphasized that the simulation model does not predetermine a single outcome for AI tool adoption, but it can provide healthcare leaders with an opportunity to restructure workflows. During the study, participating staff were prompted to shift to a more collaborative, team-based, consultative discharge process. But over time, as the collaborative behavior became institutionalized, the team-based communication essentially replaced the need to look at the tool itself, as the staff began gathering that critical data organically, he said.The researchers are now translating their findings into a working mathematical model that healthcare leaders could use to experiment with timing, training reinforcement and maintaining organizational attention over the long term.By simulating the dynamics of tool usage, healthcare organizations may be able to design better implementation strategies from the start, Rockart said.The model would enable organizations to test how different scenarios might play out over time, such as what happens if training drops off, if competing priorities increase, or if users need clearer evidence to believe the tool works.When leaders invest in new technology, they should pay at least as much attention to the human and institutional context as they do to metrics of tool performance, he said."We can imagine sort of a technologist's dream that both made sure people use the information and let us know how they'd use it and if they'd use it effectively," Rockart said. "But of course, in a setting where there are so many other goals, as well as limited resources to put that sort of weight on people, it is probably a mistake." "The goal is to get the old piece so well institutionalized into practice that when the next leader has to come in, it doesn't need that attention. It's already integrated into your workflows."HIMSS is hosting the one-day AI Executive Leadership Summit in Boston on June 24, 2026, followed by its AI in Healthcare Forum June 25–26. Register separately for the two events here and here. Andrea Fox is senior editor of Healthcare IT News.Email: [email protected]Healthcare IT News is a HIMSS Media publication.