One month after federal agencies were required to demonstrate meaningful human oversight of their highest-risk AI systems, a senior U.S. Department of Agriculture technology official confirmed on the record what government watchdogs had long feared: the workforce to perform that oversight no longer exists.

Rudolf Rojas, USDA's IT program manager and security lead, told attendees at FedInsider's Innovation Summit 2026 in Washington, D.C. on October 21, 2026, that years of DOGE-driven budget reductions and staffing cuts had stripped his agency of the capacity to maintain meaningful human sign-off on AI-generated outputs.

"I think that's untenable in our industry or in government — we just don't have the resources anymore," Rojas said. "So we're going to have to get smarter and utilize these tools so we can have them do the work, and then report back to us."

The remark arrived at a moment of acute accountability. September 22, 2026 marked the compliance deadline for OMB M-25-21 compliance requirements, "Accelerating Federal Use of AI through Innovation, Governance, and Public Trust," which required every federal agency to implement minimum risk management practices for high-impact AI — systems whose outputs serve as the principal basis for decisions with legal, material, or significant effects on individual rights and safety — or discontinue those systems. As of April 2026, agencies had reported 445 high-impact AI use cases out of 3,611 total AI use cases across 56 submitting agencies government-wide.

A USDA Inspector General inspection completed in May 2026 found that "almost none" of USDA's AI use cases lacked required authority to operate, that the agency lacked a generative AI policy, and that employees were using unauthorized AI tools management was unaware of. Now, five months later, the agency's senior IT official was saying the human oversight M-25-21 required was operationally impossible.

What OMB Required — and What It Found When It Looked

M-25-21 went beyond a general mandate to "be careful with AI." It imposed specific, measurable obligations for any AI system classified as high-impact: pre-deployment testing, completed AI impact assessments, adequate human oversight of outputs, notification and remedy processes for individuals adversely affected by AI decisions, and a duty to pause or discontinue systems that fail those standards. The M-25-21 high-impact AI safeguards applied to any system whose outputs could shape decisions about benefits eligibility, visa applications, procurement, and more.

The rule carried over core protections from Biden-era AI governance frameworks and reflected a bipartisan recognition that federal AI systems—making decisions about benefits eligibility, visa applications, procurement, and more—carry consequences that require institutional accountability, not just technical safeguards. The fact that M-25-21 carried over Biden-era protections reflected a shared recognition across administrations that government AI without adequate oversight threatens individual rights.

Whether agencies met the September 22 deadline in substance, rather than on paper, was already in doubt before Rojas spoke. The page hosting M-25-21 vanished by the compliance deadline — returning a 404 error as of the compliance deadline date — a disappearance that advocates noted made independent verification of compliance reporting more difficult. Kevin Walsh, director of the Government Accountability Office's IT and cybersecurity team, was assigned to determine whether substance matched paperwork.

DOGE Built the Gap It Now Proposes AI to Fill

Rojas's argument rests on a concrete institutional reality that predates the summit. Federal agencies absorbed significant workforce reductions over the past two years, with DOGE cost-cutting initiatives that, in many cases, specifically targeted technology and oversight roles. Review pipelines that once staffed human evaluation of AI-generated decisions — on procurement, records, benefits eligibility, border enforcement — have in many cases ceased to exist.

The feedback loop this created is, by design, self-reinforcing. The Department of Government Efficiency publicly framed AI automation as justification for further staff reductions, positioning autonomous AI as the structural replacement for the human oversight capacity it had just eliminated. The result is a compliance requirement — M-25-21's human oversight mandate — that is now operationally unfulfillable at resource-stripped agencies, not because of any technical limitation, but because the people required to fulfill it were let go.

Rojas was careful to distinguish between removing human oversight entirely and replacing it with a different control architecture. He said he wants AI agents to have "minimal privileges, minimal role," and would support autonomous operation only at agencies that have "a really good risk management framework, you've done analysis of the networks, and you've got a great demilitarized zone within your agency." He described zero-trust principles — continuous network monitoring, minimal-privilege access, network segmentation — as the replacement mechanism.

"By using zero-trust principles," Rojas said, "we can do continuous monitoring of those networks and the tools."

Why Zero Trust Cannot Do What Human Oversight Does

The technical adequacy of Rojas's proposed substitute deserves close scrutiny, because zero-trust architecture and human oversight of AI outputs solve different problems — and substituting one for the other leaves a structural gap M-25-21 was designed to close.

Zero-trust security, as defined by NIST SP 800-207 Zero Trust Architecture and implemented under the OMB M-22-09 federal zero-trust strategy, governs identity and access: which users or systems can connect to which resources, under which conditions, with continuous verification. Zero trust answers the question: who is authorized to connect? It does not — and was not designed to — answer the question: is the AI's output correct, fair, and legally compliant?

Human oversight of AI decisions, by contrast, validates output quality and preserves legal accountability. M-25-21 requires it specifically because AI systems can be technically functional and substantively wrong—producing decisions consistent with their training data while violating individual rights or causing harmful outcomes at scale. Zero trust cannot detect that failure because zero trust never examines the content of what an AI produces; it only governs whether the AI has the credentials to connect and act.

The engineering evidence for this limitation is not theoretical. A 2026 analysis by HHS ICAM Program Manager Adam McBride and ATARC Agentic IAM Task Group lead Jim St. Clair found that AI agents overwhelm zero-trust authentication tools — specifically, 1,000 AI agents operating at machine speed generate approximately 7.4 million authentication events per day, 148 times the load generated by a comparable human population. The tools federal agencies use for zero-trust identity management — PIV cards, SAML single sign-on, OAuth — were architected for human login sessions, not for the continuous, parallel, high-frequency access patterns of autonomous AI agents. Delegation chains between agents spread permissions without traceable audit trails. Service accounts used as the default onboarding mechanism for AI agents frequently violate the very least-privilege principles zero trust is meant to enforce.

A federal agency that governs its AI agents through zero trust while removing human review of their outputs has achieved secure connectivity to unaccountable decision-making. That is not a governance solution; it is a compliance gap wearing a security label.

How DHS Is Already Redefining High-Impact AI Out of Existence

Even agencies that have not formally announced abandonment of human oversight have found a quieter route to the same destination. The Department of Homeland Security has reclassified AI systems using M-25-21 loopholes — arguing that when a human agent makes the final call, the AI output is not the "principal basis" for the decision, and therefore M-25-21's oversight requirements do not apply.

Two systems illustrate the mechanism. The Anduril surveillance towers avoid oversight classification — Anduril's Autonomous Surveillance Tower is deployed in more than 300 units along the southern border, using AI to flag, track, and direct border agents toward detected individuals. DHS classifies it as non-high-impact because agents receive an "alert" rather than a directive — but critics note that the AI controls what gets flagged, which in practice determines where armed personnel are sent. The Palantir ELITE ICE targeting system, used by Immigration and Customs Enforcement to generate address confidence scores for deportation targeting, received its last privacy impact assessment before the agency acquired the tool — and assigns risk scores agents cannot inspect algorithmically.

A 2025 meta-analysis of 35 studies covering approximately 20,000 participants confirmed that humans systematically over-rely on automated recommendations even when those recommendations are demonstrably wrong—a phenomenon called automation bias. In a system where a human technically "makes the final call" on an AI-generated recommendation but cannot audit the algorithm, automation bias means the AI is effectively making the decision. Emily Froude Democracy Forward assessment characterized this dynamic directly: agents exhibit "an over-reliance on algorithmic output without being positioned to exercise meaningful scrutiny."

If DHS's "principal basis" argument is upheld and imitated, it provides a template other agencies can use to reclassify virtually any AI system out of M-25-21's scope. The rule's compliance architecture would remain intact while its substance evaporated.

"The Barbarians Are at the Gate Right Now"

Not every summit panelist agreed with Rojas's framework. Ajay Phogat, senior advisor to the chief information officer at U.S. Customs and Border Protection, offered the sharpest rebuttal — framing the removal of human oversight not as an efficiency trade-off but as a civilizational risk.

"All the pillars that we have, things that we have thought about as being extremely secure, become meaningless," Phogat said. "If the barbarians are through the gate, you're done for, right? That's what it is. They are at the gate right now."

Phogat drew on CBP's operational experience integrating AI into border enforcement and national security workflows — contexts where errors are not correctable after the fact and automated monitoring cannot substitute for judgment applied to individual human circumstances. He drew a direct analogy: "I don't know if AI is going to destroy us all, but if you're going to give a gun to a child, you are still responsible for the consequences."

CBP's caution is notable given the agency's own exposure. A coalition of immigrant rights organizations — Pangea Legal Services, Mijente Support Committee, and Just Futures Law — filed suit against USCIS and DHS components in October 2024 in the U.S. District Court for the District of Columbia, alleging that agencies had adopted AI for immigration benefit decisions including deferred action grants, visa fraud detection, and naturalization without disclosing what safeguards existed against discriminatory outcomes, and without providing applicants any process to challenge adverse AI decisions. That litigation is ongoing.

A Patchwork of Approaches Across Agencies

Tuesday's summit made visible how federal AI governance has become an agency-by-agency improvisation, with each institution developing practices calibrated to its mission profile and remaining workforce.

At the National Institutes of Health, Associate Director of Informatics Joe Croghan said the agency is building "AI deployment infrastructure where you can deploy policies, data governance, financial governance, and so forth." NIH's research culture tolerates higher risk by design, and Croghan noted that AI had compressed time-to-productive failure from months to days — a meaningful efficiency gain in a research context. A year earlier, NIH's AI use was concentrated in personal productivity; it has since shifted toward purpose-built applications with formal governance structures.

The Library of Congress took a different approach, treating AI as infrastructure rather than capital investment. Director of Digital Strategy Natalie Buda Smith pushed back on long procurement cycles in a domain where vendor tooling was evolving faster than procurement could track. LOC's AI sandbox program had grown more popular than its capacity could accommodate — a governance challenge of enthusiasm rather than resistance. Smith framed governance as an active, not passive, discipline: "Sometimes when we are deploying or rolling out a new tool, it's really exposing the ways that we need to communicate better and work better together."

The State Department embedded AI governance into its deepest institutional process: training new foreign service officers. Gray Barrett, acting director of SI Workforce Enablement at State's Bureau of Diplomatic Technology and Center for Analytics, said incoming officers receive an AI training module from their first day covering permitted tools, prohibited uses, and governance frameworks for a foreign policy context. The department's AI assistant, StateChat, reportedly saves employees roughly two hours per week. A separate media monitoring tool, North Star, extended that productivity gain further.

The National Nuclear Security Administration chose the most conservative posture of any panelist. Director of Federal IT Management Bill Watkins said NNSA deliberately restricts operations to a vetted, bounded set of AI models. "We can trust exactly what these models are going to produce," Watkins said — a posture of known quantities rather than frontier capability. Even so, NNSA is grappling with infrastructure constraints: data center space, power, chips, and hardware lead times of up to six months are creating deployment bottlenecks at an agency that cannot afford an unvetted surprise.

GAO: Federal Government Is "Over Its Skis"

The most institutionally significant assessment came from Kevin Walsh GAO IT director — a career government official who has led the agency's evaluations of federal IT acquisitions, legacy system modernization, and the implementation of the Federal IT Acquisition Reform Act since joining GAO in 2006. The GAO is the government's principal independent accountability auditor; Walsh's portfolio covers the same federal AI landscape he was assessing from the podium.

"I think the government is over its skis," Walsh said. "We don't have enough oversight or ability to truly understand the technology and what it's doing, and not just what it's doing today or a month from now."

Walsh's concern extended beyond governance philosophy into threat asymmetry. "The attackers are using it, and you really need to use it on defense if they're using it on offense," he said. "We always used to say in cybersecurity: it's not a matter of if, it's when you're going to get breached. That's absolutely still true, but now you're going to get breached, and you're not even going to know it until months later if you're not using AI to be on that spectrum."

He also raised a structural problem rarely surfaced in AI governance debates: the federal government's long-standing pattern of layering new technology on aging infrastructure, rather than replacing it. Without a clear definition of "legacy" and proactive update cadences, Walsh said, agencies risk deploying AI on foundations that cannot support the continuous monitoring those systems require to remain compliant with M-25-21 or safe under adversarial conditions.

The GAO 94 federal AI requirements report identified 94 AI-related requirements applying government-wide and noted that 10 separate executive branch groups have roles in AI oversight implementation — a fragmentation that, without coordination, risks becoming accountability without accountability to anyone in particular.

The Litigation and Legislation Closing In

While agency officials debated governance philosophy at the summit, the legal and legislative landscape around federal AI was narrowing from multiple directions.

A federal judge ruled in May 2026 that the Department of Government Efficiency violated the law when it used ChatGPT in federal grant cancellations to review and cancel federal humanities grants — establishing a precedent that AI-driven federal decisions without adequate process carry legal exposure. That ruling has not been lost on agencies now contemplating broader AI deployment in higher-stakes contexts.

On Capitol Hill, several legislative proposals are advancing that would impose stricter requirements than M-25-21's current framework. The FRONTIER Act H.R. 9925 introduced, introduced July 23, 2026 by Representatives Lori Trahan and Jay Obernolte, would establish a tiered risk-based AI governance framework with legally enforceable safeguards for high-risk applications.

The AI LEAD Act (S.2937), introduced by Senators Hawley and Durbin, would classify AI as a product subject to civil liability and strip AI companies of Section 230 immunity for AI-generated outputs that cause harm. Senator Bernie Sanders and Representative Greg Casar introduced legislation in September 2026 to establish a cabinet-level AI safety agency and impose a development moratorium on the most powerful AI systems. Representatives Ted Lieu and Brian Moran introduced a separate bill authorizing the DHS Secretary to order a slowdown or shutdown of AI systems determined to pose a risk of catastrophic harm.

None of these bills has passed. None of them would resolve the workforce gap Rojas described.

What Rojas Conceded

In a notable aside, Rojas — despite arguing that human-in-the-loop oversight was no longer viable — declined to defend the current administration's posture toward AI regulation.

"I think we've had good regulation in the past — railroads, nuclear energy and so forth," he said. "We gotta come together as a nation and think about regulating so that we can use these tools in a safe way and protect against these bad actors" — a reference to China, Russia, and other nation-state adversaries seeking to exploit gaps in U.S. AI governance.

He also said he "kind of" disagrees with what he characterized as an approach that gives AI companies "free rein" to develop without constraint.

The acknowledgment crystallizes the paradox at the center of Tuesday's summit. Rojas was not arguing for unaccountable AI — he was arguing for a different accountability architecture, one built on automated safeguards and zero-trust constraints rather than human reviewers. But the regulatory frameworks that would make that architecture trustworthy across vendors, across agencies, and against adversarial exploitation do not yet exist.

The technical substitute for human oversight he described requires the very kind of binding interoperability standards, vendor accountability requirements, and enforcement mechanisms that current federal AI policy does not provide and that the legislation he endorsed has not yet supplied.

The federal government is running 445 high-impact AI use cases. The deadline to bring them under meaningful governance has passed. The principal auditor says oversight is inadequate. An agency insider says human review is impossible. And the proposed replacement — zero-trust architecture — is a solution to a different problem.

The gap between mandate and capacity is not a policy debate. It is operational fact, confirmed by officials on all sides of the room.

Frequently Asked Questions

What is OMB M-25-21, and what does it require federal agencies to do?

OMB Memorandum M-25-21, "Accelerating Federal Use of AI through Innovation, Governance, and Public Trust," was published April 3, 2025, with a compliance deadline of September 22, 2026. It requires agencies to implement minimum risk management practices for "high-impact" AI systems — those whose outputs serve as the principal basis for decisions affecting individual rights or safety. Mandatory safeguards include pre-deployment testing, AI impact assessments, adequate human oversight of AI outputs, notification and remedy processes for affected individuals, and a duty to pause or discontinue systems that fail those standards. As of April 2026, agencies had reported 445 high-impact AI use cases government-wide. See the full M-25-21 high-impact AI requirements for details.

Can zero-trust security architecture replace human oversight of AI decisions?

No — not because zero trust is ineffective at what it does, but because it addresses a different problem. Zero-trust security governs identity and access: who can connect to what, under what conditions, verified continuously. Human oversight of AI decisions governs output validity and legal accountability: whether the AI's decision was correct, fair, and legally compliant for the individual it affected. Zero trust cannot detect an AI that has the credentials to act but is producing discriminatory or erroneous outputs. A 2026 analysis found that zero trust fails at agent scale — specifically, 1,000 AI agents at machine speed generate 7.4 million authentication events per day, a volume that overwhelms existing zero-trust tools built for human login sessions — compounding rather than solving the oversight gap.

What rights do Americans have when a federal AI system makes a decision that affects them?

Under M-25-21, agencies are required to notify individuals when high-impact AI is used in decisions affecting them and to provide a remedy process to challenge adverse outcomes. In practice, that requirement has been inconsistently implemented: a lawsuit filed in October 2024 against DHS and USCIS components alleged that multiple immigration AI systems were operating without any disclosed safeguards or challenge process for affected applicants. The USDA OIG AI inspection findings from May 2026 found that "almost none" of USDA's AI use cases had authority to operate. If an AI system affected your benefits, permit, or immigration status, you have the right to request a human review — but whether that review process exists in a staffed, functional form at a given agency is, based on the summit record, an open question.

What is the "principal basis" loophole, and how is it being used to avoid M-25-21 requirements?

M-25-21's high-impact definition covers AI systems whose outputs serve as the "principal basis" for significant decisions. DHS has argued that when a human agent technically makes the final call — even on an AI-generated recommendation the agent cannot audit or override in practice — the AI output is not the "principal basis," placing the system outside M-25-21's scope. Critics note that this interpretation, applied to Anduril border surveillance towers and Palantir's ELITE ICE targeting system, effectively removes some of the highest-stakes federal AI deployments from any oversight requirement. Research confirms that DHS AI loophole and automation bias affect how humans process AI recommendations — specifically, that people systematically over-rely on automated outputs even when those outputs are wrong, meaning "human in the loop" is not a meaningful safeguard when the human cannot inspect the underlying algorithm.