Federal AI adoption soars, yet real-world impact remains hidden in plain sight

Terry Gerton We’re discussing how federal government agencies are adopting artificial intelligence. While recent White House teams have all placed AI high on their priority lists, your new Brookings report suggests the on-the-ground picture is more varied. How would you describe where the government currently stands with AI? Are we seeing fast, widespread progress, or is it a mixed bag?

Valerie Wirtschafter There has been real consistency from one administration to the next. The Biden team continued several of the first Trump team’s approaches, such as cataloging AI implementations and putting governance steps in place. Now the second Trump team has resumed those efforts with some adjustments—mostly in wording and minor details—but largely keeping the same underlying framework. What we’re seeing is a real drive to weave AI more deeply into day-to-day operations. Some of that involves behind-the-scenes efficiency work, while other efforts are directly tied to core missions, like delivering health care or supporting law enforcement. Agencies publish their AI use case inventories each year, though the reporting instructions have been tweaked across the last three cycles—2023 through 2025—so some of the variation in the numbers is just due to clearer guidance. Still, much of the data points to genuine growth in adoption, which matches what federal technologists have told me. In 2023, roughly 700 use cases were reported; by 2025, that figure jumped past 3,600. The number of participating agencies also grew from about 20 to over 40. That said, the growth is misleading because it’s heavily concentrated. Five agencies alone account for roughly half of all reported use cases across the 2023–2025 inventories, and over the past year, large agencies—those with more than 15,000 employees—submitted over three quarters of the total. So while AI adoption is clearly expanding, it’s clustered among a small number of particularly large organizations.

Terry Gerton Let’s circle back to those use cases for a moment. When you look at the inventory data, what kinds of applications show up most frequently?

Valerie Wirtschafter Oh, it’s hard not to go down a rabbit hole with this because the detail is fascinating. We see a lot of behind-the-scenes improvements—efficiency gains, better document processing, smarter summarization. But DHS and DOJ each have a substantial share of law enforcement-related cases, around 36% for DHS and 54% for DOJ. Health and Human Services and the VA are dedicating a large portion of their AI efforts to improving health care delivery and medical services. So the use cases are clearly reaching into the heart of what these agencies do.

Terry Gerton One theme that stands out in your report is the distance between identifying a use case and seeing real, measurable outcomes once it’s been put into practice. Can you walk us through that?

Valerie Wirtschafter Sure. The data includes information on where each use case sits in the pipeline. What jumps out is that projects in the pilot phase—still being tested and evaluated—are roughly equally likely to be built in-house or with outside contractor help. But once a project reaches full deployment, the balance shifts significantly toward external vendors. Far more projects are stuck in pilot or pre-deployment stages than are actually running in production. This tells us a couple of things. It could mean agencies have the staff and know-how to prototype and experiment but run into trouble when it’s time to scale. Or it could reflect a newer crop of projects being built internally from the start. It might be a encouraging sign of growing in-house capability, or it could highlight real challenges with scaling and long-term adoption. Most likely, both forces are in play at the same time.

Terry Gerton Valerie Wirtschafter is a fellow in the Artificial Intelligence and Emerging Technology Initiative at the Brookings Institution. Valerie, when you talk about scaling and capacity—is there something unique about AI that makes the jump from pilot to full deployment harder than with other technologies?

Valerie Wirtschafter I think the defining feature of AI systems right now is that the technology is moving so fast that the frontier keeps shifting. New models with new capabilities come out on a regular schedule. A pilot that an agency spent months building and getting approvals for could be outdated before it ever launches, simply because a better solution appeared. That kind of rapid obsolescence is a distinctive challenge of AI adoption—it’s not something agencies really face with cloud computing or conventional software.

Terry Gerton You noted that real momentum is mostly limited to a handful of large agencies. What’s holding back the smaller ones or those still getting off the ground with AI?

Valerie Wirtschafter It mostly comes down to two things. First, the broader the agency’s budget, the more room it has to invest in experimentation, absorb possible funde failures, and accept that a project might become irrelevant within ten months. For smaller agencies with tighter resource constraints, that kind of risk tolerance is much harder to sustain. Second, there’s a talent challenge across the board, but it’s especially acute where the pool of skilled people is already thin. Culture also plays a role. Larger agencies sometimes have more latitude to try new things. Even within a workforce that tends to avoid risk, bigger agencies may have a dedicated budget for piloting AI solutions that support their mission, even if those solutions end up falling short of expectations. So limited funding, a narrow talent pipeline, and organizational culture all weigh more heavily on smaller agencies.

Terry Gerton You’ve mentioned culture and risk aversion earlier. Given how rapidly the AI landscape evolves, I’m curious—could there be hesitation to dive in now, reasoning that by the time implementation takes two years, the technology will have already transformed?

Valerie Wirtschafter Absolutely, that’s a significant challenge—not just within federal government, but across sectors. The key is figuring out how to stay current with rapid advancements, iterate continuously, and adopt new tools now—while accepting that whatever you implement today will require ongoing updates down the line. Part of the solution lies in fostering continuous learning, such as integrating skill development into performance reviews. This keeps both technical experts cutting-edge and the broader workforce equipped to integrate AI effectively—even as it keeps changing.

Terry Gerton In your research, what specific approaches have proven most effective in helping agencies move from being merely AI-curious to actively using AI?

Valerie Wirtschafter Great question. At its core, the barrier is often that people lack room to experiment and explore. AI use cases vary widely—not just by agency, mission, or organization, but also by individual workflow preferences. So, creating space to experiment—and rewarding that exploration—is crucial. Equally important, though, is resisting the urge to force AI where it doesn’t belong. There’s a common narrative that to appear forward-thinking, you must embed AI everywhere. But sometimes, what you really need is basic data cleanup—no AI required. Thoughtfully choosing where to deploy AI makes a big difference. For instance, AI shines when dealing with massive data volumes or processing backlogs—like the mountains of paperwork piled up across federal agencies—that humans simply can’t tackle efficiently. Those are ideal spots for AI to add real value. Being strategic about placement also reassures employees who don’t feel pressured to use AI unnecessarily—and that kind of pressure can backfire. The most effective strategy, in my view, is encouraging adoption where it genuinely helps people do their jobs better.