This is the 15th article in our IT lifecycle management series, Delivering the tech that delivers for government.

Federal agencies aren’t struggling to experiment with artificial intelligence. They’re struggling to scale it.

That disconnect is becoming the defining challenge for federal systems integrators, points out Scott Stapp, chief technology officer and chief revenue officer at DEFCON AI. And Stapp has inside knowledge. A retired Air Force brigadier general, he’s spent the better part of three decades helping the government solve data management and decision-making challenges both inside the government but also leading industry technology teams.

In a conversation for our series Delivering the tech that delivers for government, he made a direct point about what’s really holding government back.

Agencies do not have the “data fabric and ontology that allows things to cross. If you’re going to develop an AI tool for anything — whether it’s the Department of Homeland Security, Energy or War — you need to share that information,” Stapp said. “The whole idea of using an AI tool is the more data you have, the better. The more people who can participate in it, the better. But right now, that ontology is not connected.”

That reality is keeping agencies stuck in pilot mode, Stapp said.

The integration gap is the real bottleneck

In the commercial world, scale comes from connectivity, he said. Data flows. Systems talk to one another. Models build on shared inputs.

Government doesn’t work that way. “It is not well connected,” Stapp said. “It doesn’t talk. The [military] services don’t necessarily connect and talk to each other.”

The result? AI solutions deliver value in isolated pockets, but they can’t easily expand. “What you find out is that fabric doesn’t exist to actually do that,” he said.

For FSIs, this is the mission. The answer does not lie in creating more or better algorithms. It’s also about building the underlying architecture that lets the algorithm operate across systems, services and environments, Stapp said.

Cloud strategy is shaping everything

Agencies are moving deliberately when it comes to cloud and AI, not because they’re slow but because they’re cautious about long-term constraints.

The government “is really trying to be careful to not get vendor locked into [any specific] cloud environment,” Stapp said.

That concern is reshaping how AI systems are designed and procured. Portability matters. Interoperability matters more.

“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 said.

For integrators, that raises the bar, Stapp said. Their solutions will all need to be able to move seamlessly too.

The return of the prime integrator

One of the clearest signals is the re-emergence of integration as a central function, he said, and that weaves together software, data and AI ecosystems.

Historically, the federal government leaned on prime contractors to integrate complex physical systems. That model is now reappearing in the AI era.

Right now, there are limited prime contractors able to do this work, Stapp said, but increasingly agencies will be saying, “Hey, we do need people and companies who can come in and tell us how we start fitting these things together.”

He pointed to his own company’s recent deal with the Marine Corps. The corps in June awarded DEFCON AI a five-year, $115 million prototype agreement focused on logistics modernization and AI integration. It involves deploying tools plus ensuring they work together within a common data framework.

“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 and open architected capability,” Stapp said. The corps aims to be an AI-first force.

For FSIs, that signals that the market is shifting toward orchestration by aligning vendors, tools, data and environments into something that works, Stapp said.

Edge constraints force smarter design

Scaling AI also creates a physics problem. At the edge, compute and power can be limited. Latency can disrupt mission effectiveness. That changes how systems are built.

“You want minimal amounts of data needed in that edge for the specific decisions,” Stapp said.

That’s a departure from cloud-native thinking, where more data and more compute typically lead to better outcomes. But that’s neither efficient nor effective in real-time work and decision-making that might take place far from an agency’s headquarters.

For integrators, it means designing systems that are:

Distributed
Efficient
Purpose-built for specific missions

The days of simply extending centralized architectures are over, he said.

Government data is fundamentally different

Another reality is that AI in government is more complex than in commercial environments, and not just because of scale but because of other conditions too.

“You have an adversary whose goal in life is to show you information that is not accurate and is deceptive,” Stapp said.

That introduces challenges many commercial systems don’t face:

Deceptive or manipulated data
Incomplete datasets
Classified data silos

In the defense realm, Stapp noted, “the only people who have all the data are the people at the highest level, and those aren’t the people who are on the battlefield, right? The people on the battlefield are typically at a collateral level.”

That leads to data becoming increasingly less useful for driving decision-making at the edge. The dynamic forces a different approach to AI design, Stapp said. Models must operate with less information but still produce reliable outcomes.

Modularity isn’t optional

If integration is hard, then modularity becomes essential. Each agency or service, system and mission environment operates differently.

Even when the problem looks similar, the implementation rarely is, Stapp said.

“You must make it modular because really it doesn’t work with anything else,” he said. That means the same application often needs to be adapted across environments.

“You have to do tweaks and changes to make the same application work with a different service,” he said.

Therefore, modular design isn’t just good architecture, it’s the only viable approach, Stapp said.

Start small to scale at all

Despite the push for enterprise AI, large-scale transformation isn’t the starting point in government, he said. “You start with smaller problem sets. And then try to grow it out.”

An incremental approach reflects the realities of data access, classification and infrastructure maturity. It also aligns with where AI is gaining traction now, Stapp said.

The path to AI at scale in government runs through integration by aligning data, architectures, clouds and mission needs in a fragmented environment, he reiterated.

AI won’t scale until integration does, Stapp said. The onus is on the government’s industry partners to help agencies address the challenge.

“The number one thing with any customer is you’re there to solve their problem.”

What it takes to scale AI in government
DEFCON AI’s Scott Stapp believes helping the government scale artificial intelligence is about better integration, not better models. He suggests focusing on a few critical capabilities:
Build the data fabric first: AI doesn’t scale without shared data. Interoperability and ontology alignment are foundational.
Design for multicloud from day one: Agencies want flexibility. Vendor lock-in is a nonstarter.
Lead as the integrator: The opportunity is in connecting tools, vendors and data, not delivering point solutions.
Go modular by default: Every environment is different. Solutions must be adaptable across services and systems.
Optimize for the edge: Limited compute and power demand smaller, mission-specific AI models.
Start small, then expand: Narrow use cases are the fastest path to adoption and scale.

Discover more smart tips and tactic for FSIs shared by leading technologists for our Delivering the Tech That Delivers for Government series.

Copyright
© 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.