AI Drives Federal Efficiency Gains but Magnifies Existing Challenges

Federal health agencies are realizing measurable gains from artificial intelligence deployments, but officials said organizations must address underlying data, security and operational challenges before scaling the technology across their enterprises.

“AI is an absolute magnifying glass on any existing pressure points or challenges that you have,” said Ben Rogers, acting deputy chief AI officer with the Centers for Disease Control and Prevention during the AFCEA Health IT Summit last week.

AI can accelerate both the strengths and weaknesses of an organization, Rogers said. At the CDC, generative AI tools were made broadly available to employees beginning in February 2024, enabling the agency to assess the technology’s impact on productivity and operations.

An analysis conducted in June 2025 found the deployment saved approximately 41,000 staff hours and generated a 520% return on investment. Despite the impressive gains, Rogers said the assessment also revealed areas where AI can amplify existing challenges.

“If you do not have your data security system set up in place, AI is going to make that worse. If you are having a challenge from whatever process workload … if you have cybersecurity challenges, AI is going to make that worse,” Rogers said.

Keeping the Human in the Loop

Rogers stressed the importance of maintaining human oversight of AI-generated outputs. However, effective oversight requires personnel with both the expertise and time to validate the technology’s recommendations.

“Human in the loop only matters if you have a person with the right expertise and, most importantly, time to be able to form that verification,” he said.

Without adequate staffing and expertise, organizations risk creating a process where employees simply approve AI-generated information without meaningful review, potentially undermining public trust. Agencies must align human oversight responsibilities with both mission requirements and AI capabilities “to make sure that you’re not in a situation where you’re harming public trust in any way, shape or form,” Rogers said.

Speeding Research

AI is increasing productivity in clinical research, creating new opportunities and new challenges for health agencies, according to Chris Kinsinger, assistant director for Catalytic Data Resources at the National Institutes of Health.

The technology is helping researchers generate hypotheses and accelerate scientific discovery at a much faster pace than traditional methods.

“What does that mean for our system? Are we going to see a lot more [clinical] applications? We’re already seeing tons of applications. … So what do we have to do to our system to account for potential increase in hypothesis generation?” he said.

As AI produces more discoveries that could translate into treatments and clinical products, health systems must be prepared to evaluate, integrate and deliver those innovations to patients, he added.

“Do we have the systems [in place] to actually deploy them?” he asked.

Predictive Analytics at the VA

The Department of Veterans Affairs sees significant potential for AI to accelerate the development of predictive analytics tools used to improve patient outcomes, said John Scott, the VA’s interim chief data officer.

The VA has spent decades building predictive models, and AI could help speed the discovery and analysis phases of that work. However, Scott said success depends on having comprehensive, high-quality data.

“You have to have a complete data set on the population that you’re trying to care for. It has to be high enough quality data that you can allow the AI entity to figure out what the inferences are in prediction,” he said.

For example, Scott noted that the VA spent decades developing systems to identify and monitor diabetes-related health indicators. AI-powered predictive analytics can generate insights much faster, but clinicians must ultimately determine whether those insights are reliable enough for patient care.

Ultimately, it is up to clinicians “to decide whether that’s good enough to put into practice,” Scott said.

Building these systems also requires access to complete data sets. One challenge for the VA is applying predictive analytics to service members transitioning from active-duty military service to civilian life and eventually becoming VA patients.

“The data isn’t necessarily authorized to flow from the record system to our data platform on those people until they become our patient and then we have to work on that aspect in order to apply that [capability],” Scott said.