Pentagon AI Adoption Outpaces Accountability Frameworks

The Speed-Governance Gap

A senior Pentagon official recently celebrated a milestone: his depleted staff used GenAI.mil to produce what he called "the best report they had written in five years" to meet a congressional deadline. No review process or accuracy check was mentioned. In the five weeks since the Agent Designer tool launched, Defense Department personnel built more than 100,000 user-created AI agents.

The enthusiasm is understandable, but the episode exposes a fundamental problem. The military's AI adoption sprint has dramatically outpaced the development of accountability mechanisms that should accompany it.

Why it matters

Without clear ownership and audit trails for AI-generated work products, commanders cannot reliably identify the risks they're underwriting. The Defense Department has essentially created hundreds of thousands of autonomous agents performing organizational actions with no centralized registry, no standard for attribution, and no clear answer to the question: who is accountable when an AI agent produces flawed output that informs military or congressional decisions?

Three Converging Pressures

The GenAI.mil rollout occurred under conditions that created what Colonel Josh Goodrich calls a "governance bypass." First, the Deferred Resignation Program hollowed out institutional knowledge—the tacit judgment about which outputs to distrust and which approvals to route. That expertise departed before AI tools arrived, with no requirement to transfer governance functions explicitly.

Second, operational urgency drove adoption. Deputy Assistant Secretary of Defense Jacob Glassman framed the agent milestone against active military operations, noting the department was operating while at war. When urgency logic migrates from kinetic operations into information governance, process safeguards become invisible.

Third, mandate compliance created adoption pressure from the top. Since December 2025, Secretary Pete Hegseth has pushed aggressive AI adoption across the force. GenAI.mil now serves 1.7 million users—roughly half the department's workforce—and expanded from one model to three, adding ChatGPT and Grok alongside Gemini. When senior leadership signals that adoption is the performance metric, practitioners closest to implementation problems have the least power to slow the pace.

Organized Irresponsibility

The congressional report case illustrates what Goodrich calls "organized irresponsibility." Who owns that product? The official who directed its production? The staff who wrote the prompts? The office that deployed the platform? The vendor whose model generated the content? Every actor holds a defensible argument that accountability belongs elsewhere.

The FY2026 National Defense Authorization Act mandates governance frameworks, but timelines defer the reckoning: the Section 1533 assessment framework isn't due until January 2028. The problem is operating now.

No centralized registry tracks the hundreds of thousands of deployed agents. No standard requires AI-assisted products delivered to Congress to identify the model used, the prompting authority, or the human review step. The National Geospatial-Intelligence Agency adopted such a disclosure template for machine-generated intelligence products more than a year ago, but no equivalent exists for AI-generated staff work.

Four Immediate Actions

Goodrich, a Massachusetts Army National Guard colonel with a PhD in management, proposes four steps that don't require congressional action. First, establish an AI provenance standard for congressional and executive products. Second, require governance transfer before workforce reductions eliminate positions exercising judgment over information quality. Third, maintain command-level agent registries so commanders can identify agents operating in their formations. Fourth, pair every adoption metric with an accountability metric—a named reviewer, an error correction pathway, an accountable official.

The mission command principle that Secretary Hegseth correctly applies to warfighting doesn't transfer automatically to AI-enabled staff work. Commanders need not know every action their subordinate elements take, but they must be able to establish intent, assign responsibility, and assess consistency. With hundreds of thousands of unregistered agents, that's currently impossible.

These details were first reported by the Modern War Institute at West Point, where Goodrich published his analysis of the Pentagon's AI governance challenge.