Why Federal Agencies Can't Deploy Agentic AI at Scale Yet

The U.S. government cannot yet deploy agentic AI systems at scale because today's models are not reliable enough for high-stakes operational environments, according to a former Pentagon AI official writing in The Hill.

Mark Beall, who served as the inaugural director of AI Strategy and Policy at the Pentagon's Joint Artificial Intelligence Center, argues that while agentic AI promises massive productivity gains—faster citizen services, smarter military logistics, lower costs—current systems fail unpredictably when moved from demonstrations to real-world operations. Model hallucinations, prompt injection attacks, and misaligned objectives could prove catastrophic in defense, financial, or infrastructure contexts.

The reliability gap

Beall describes a fundamental engineering challenge: unlike traditional software, AI systems write their own code and interpret instructions autonomously. When they fail, there is no stack trace to examine. "A chatbot that gets a fact wrong is an inconvenience," he writes. "A superintelligent agentic system with access to classified military data, financial controls or critical infrastructure is a national security incident waiting to happen."

The problem intensifies as AI capabilities grow. Systems that perform well in testing may behave differently in production. Adversaries can exploit these vulnerabilities. And the higher the stakes of deployment, the less acceptable these failures become.

Why it matters

Reliability is not a constraint on American AI dominance—it is the precondition for it. China faces identical technical challenges in deploying agentic AI within its state apparatus. The nation that solves reliability first will field AI that actually works at scale in critical systems, not just in marketing demonstrations or stalled pilot programs. Technical safety becomes a competitive advantage, not a regulatory burden.

The policy solution

Beall has worked with lawmakers on what he calls "the central AI policy challenge of this Congress": building the science, testing frameworks, and institutions needed for trustworthy agentic AI. A bipartisan coalition of national security and AI experts has urged Congress to fund the National AI Reliability and Control Initiative (NAIRCI) at $2 billion in the fiscal 2027 National Defense Authorization Act.

NAIRCI would provide targeted research funding for unsolved problems in AI reliability: making systems behave predictably, verifying they execute intended tasks, maintaining alignment with human intent as capabilities expand, and preserving meaningful human oversight. This research would accelerate American AI companies, defense contractors, and agencies in delivering products the government can actually deploy.

Two races at once

Beall frames the challenge as two simultaneous competitions. The first is the commercial and military AI race with China, where market share is the goal. The second is a longer race against time to ensure increasingly powerful systems remain under human stewardship as their capabilities grow.

Both races are won the same way: by investing in reliability science, building evaluation infrastructure, and creating policy frameworks that let American organizations accelerate deployment of trustworthy AI. Poll data shows the American public wants both AI benefits and responsible deployment assurances.

The details were first reported by Mark Beall in an opinion piece for The Hill. Beall now serves as president of the AI Policy Network and has testified before Congress on AI and national security.