Why Federal AI Projects Stall at Pilot Stage: Integration, Not Models

Federal agencies have moved past the question of whether artificial intelligence works. The challenge now is making it work everywhere.

Scott Stapp, chief technology officer at DEFCON AI and a retired Air Force brigadier general, identifies integration architecture as the primary obstacle preventing government AI from scaling beyond isolated pilots. The problem isn't the sophistication of machine learning models—it's the absence of connective tissue between systems.

The missing data fabric

Agencies lack the "data fabric and ontology that allows things to cross," Stapp explained in an interview first reported by Federal News Network. Without shared data frameworks, AI tools built for one department or service remain trapped there, unable to leverage information from adjacent systems.

This fragmentation contradicts a core principle of effective AI: more data produces better outcomes. "The whole idea of using an AI tool is the more data you have, the better," Stapp said. "But right now, that ontology is not connected."

The military services exemplify this disconnect. "It is not well connected. It doesn't talk. The services don't necessarily connect and talk to each other," he noted. AI solutions deliver value in pockets but cannot expand across organizational boundaries.

Multicloud strategy reshapes procurement

Agencies are deliberately avoiding vendor lock-in as they build AI infrastructure. "The government is really trying to be careful to not get vendor locked into [any specific] cloud environment," Stapp said.

This caution is driving demand for portable, interoperable AI systems that can operate across multiple cloud platforms. "Once you start to look at those ontologies, where data and information can be passed seamlessly across multiple cloud environments, you're going to see the use of AI tools grow drastically," he explained.

For systems integrators, this requirement raises technical complexity. Solutions must function equally well regardless of underlying cloud infrastructure.

Prime integrators return

The federal market is witnessing the re-emergence of prime integration contractors—companies capable of orchestrating disparate software, data sources, and AI tools into coherent systems.

DEFCON AI's recent five-year, $115 million prototype agreement with the Marine Corps illustrates this shift. The contract focuses on ensuring new AI tools fit within a common data framework and open architecture as the Corps pursues its goal of becoming an AI-first force.

"They want to bring all these new AI tools in, and they have us looking at all those tools to ensure that they fit within a data framework," Stapp said. The work centers on orchestration rather than point solutions.

Edge computing demands efficiency

Scaling AI to tactical edge environments introduces physical constraints absent in data center deployments. Limited compute power, restricted bandwidth, and latency concerns force different design approaches.

"You want minimal amounts of data needed in that edge for the specific decisions," Stapp said. This contrasts sharply with cloud-native AI, where more data typically improves performance.

Edge AI systems must be distributed, efficient, and purpose-built for specific missions rather than general-purpose tools extended from centralized architectures.

Adversarial data complicates models

Government AI faces unique challenges beyond commercial applications. "You have an adversary whose goal in life is to show you information that is not accurate and is deceptive," Stapp noted.

Defense and intelligence AI must operate with incomplete datasets, navigate classification barriers, and detect manipulation. Battlefield personnel typically work at lower classification levels than where complete data resides, forcing AI models to produce reliable outputs with partial information.

Why it matters

The federal government's AI scaling problem represents a market shift from algorithm development to systems integration. Agencies need partners who can connect fragmented data environments, ensure multicloud portability, and design modular solutions adaptable across diverse mission contexts. As Stapp emphasized, agencies increasingly need "companies who can come in and tell us how we start fitting these things together." For technology providers, the opportunity lies in orchestration capabilities rather than standalone AI products.

Stapp advocates starting with narrow use cases before expanding. "You start with smaller problem sets. And then try to grow it out," he said—an approach reflecting current realities around data access, classification levels, and infrastructure maturity.

These insights were shared in an interview conducted by Vanessa Roberts for Federal News Network's series on delivering technology for government.