The Office of Management and Budget (OMB) issued two memos last week aimed at accelerating AI adoption across the federal government. The first (M-25-21) outlines how agencies should adopt AI to improve public services, while the second (M-25-22) focuses on how to procure AI systems more efficiently. Together, these documents lay out the administration’s vision for AI in government: Accelerate innovation, reduce red tape, and scale what works. But that vision won’t materialize unless other parts of the administration stop pulling in opposite directions or failing to act altogether.

Telling agencies to adopt AI effectively isn’t enough to make adoption happen. Before agencies can procure and use AI effectively, they need to know what kind of performance they want from an AI system, which technical features are likely to produce that performance, and how to verify whether a system actually delivers it. Right now, none of that foundation is fully in place. There are three steps the White House should take to change that.

1. Direct Federal Agencies to Identify Their Desired AI Performance Outcomes

The first step to effective AI adoption is ensuring agencies define the type of AI performance that matters in their specific domains. To see why, think of Olympic cycling, which includes road racing, track cycling, mountain biking, and BMX. All these disciplines use bikes, but the type of performance each one demands from the bike itself is entirely different. Road racers need bikes to be aerodynamic to maintain high speeds over long distances, mountain bikers require bikes with effective shock absorption to handle rough, uneven terrain, and BMX riders demand bikes that are agile and durable enough to withstand sharp turns, jumps, and impacts. To pick the right bike, each team must clearly understand what performance outcome it’s aiming for.

It’s the same for federal agencies with AI systems. For example, the Department of Justice might require AI systems that ensure fairness and minimize bias, the Department of Energy might need secure, resilient systems for critical infrastructure, and the Department of Health and Human Services might need AI tools that are clinically validated for reliability. But unlike Olympic teams that know their performance needs, many federal agencies haven’t pinned down the specific outcomes they want AI to deliver. Meanwhile, OMB’s guidance effectively tells them to acquire AI that “works well” for their missions—akin to telling a cycling team “just buy the best bike” without having them identify the event they’re competing in or the performance they need to optimize for. Agencies can’t make good AI decisions until they define what “working well” actually means for their context.

The forthcoming White House AI Action Plan should direct every federal agency to identify and articulate the precise performance outcomes they need from AI systems. That clarity is a prerequisite for any effective procurement and will also provide valuable feedback to the private sector on what it should prioritize if it wants to sell to these agencies.

2. Direct NITRD to Prioritize R&D That Links Technical Features to Those Outcomes

Even if agencies identify the outcomes they care about, such as fairness, reliability, or security, they still need to understand which technical features produce those results. In cycling, it took years of engineering research to connect specific design choices with improved speed, enhanced durability, or better handling. Engineers and riders studied how frame geometry influences aerodynamics, how certain alloys reduce weight without sacrificing strength, and how suspension elements distribute force on uneven terrain. The cycling world developed this knowledge over time through concerted R&D and testing in real conditions.

In AI, efforts to map technical features to desired outcomes remain nascent. Researchers are exploring methods for enhancing security (e.g., fault-tolerant architectures and encrypted data flows) and achieving privacy (e.g., federated learning and differential privacy), among others. But there is still not yet an established body of knowledge that reliably connects which features matter most to improved real-world performance.

The White House should direct the Networking and Information Technology Research and Development (NITRD) subcommittee to update the National AI R&D Strategic Plan with a specific focus on identifying how different technical parameters map to measurable AI performance outcomes. The National AI R&D Strategic Plan was first launched under Obama, updated under Trump, and expanded under Biden, and it has always reflected the priorities of the administration. A second Trump term should now refocus it on the core technical challenge hindering AI adoption: understanding which system design choices reliably produce the kinds of AI performance organizations want to see.

3. Support NIST’s Ability to Develop Evaluation Protocols That Fuel Adoption

Understanding which performance outcomes matter (step 1) and which technical features drive them (step 2) is crucial but not enough for effective adoption. Agencies also need reliable methods to test whether an AI system actually performs as intended. In cycling, teams do not stop at good design. They collect performance data to confirm that certain frame shapes reduce drag or that specific materials enhance strength. A promising design still has to prove itself on the track.

Federal agencies face the same challenge. They need the ability to verify whether technical choices such as encrypted data flows, differential privacy, or fault-tolerant architectures actually deliver the outcomes they are meant to support. The National Institute of Standards and Technology (NIST) has led efforts to develop testing methods for exactly this purpose, and that work is essential to translating technical progress into concrete, measurable benchmarks that agencies can use to evaluate and select AI systems. Unfortunately, just as this work becomes more central to effective AI adoption, NIST’s capacity is at risk. Budget cuts threaten to undermine the agency best positioned to turn emerging design knowledge into the evaluation infrastructure needed to scale adoption across government.

The Trump administration should preserve NIST’s technical capacity to develop credible, outcome-based evaluations for AI. Indeed, this is precisely what NIST is uniquely well-placed to do: Develop robust evaluation protocols that test systems against clearly defined, pre-established performance criteria. What they are not, however, is a substitute for the fundamental policy work of guiding agencies to first define what outcomes matter in their domains—a fact that congressional bill after congressional bill continues to misunderstand.

Conclusion

Agencies aren’t starting from zero when it comes to AI adoption; more than 1,700 use cases of AI have already been reported, but most are still experimenting without the technical foundation needed to scale AI effectively. If the administration is serious about accelerating AI adoption, it should not leave these gaps unaddressed. OMB may set the direction, but unless the administration equips agencies to act, AI adoption will be stalled. These three steps would make a big difference in turning vision into action.

Image credit for social media preview: Executive Office of the President of the United States