Dennis G Perry, PhD, MBA  |  October 26, 2026

The disruptive question is no longer whether artificial intelligence can perform work without people. It is whether an autonomous system should be permitted to take an action that no person has specifically authorized.

The Federal Government Has Exposed the Contradiction

At a federal technology summit on October 21, USDA IT program manager Rudolf Rojas argued that routine human-in-the-loop oversight is becoming untenable as agencies lose personnel and contractor support. He proposed constrained privileges, risk management, network segmentation, and continuous zero-trust monitoring as conditions for unattended agents. The pressure he describes is real. The proposed safeguards, however, do not by themselves settle whether a particular permitted action is appropriate at the moment it is attempted [1].

There is a revealing contrast. USDA’s Inspector General reported in May that 73 of 82 operational AI use cases in the inventory reviewed lacked an authorization to operate before deployment. The report also found incomplete AI policy updates and failure to implement minimum high-impact AI risk practices by the April 2026 deadline [2]. These are historical audit findings, not proof that every deficiency remains unresolved today. They nevertheless expose the danger of treating more autonomy as a substitute for established governance.

Permission Is Not Judgment

Least privilege answers a necessary question: what resources may this agent access? It does not answer another: should the agent execute this particular instruction now? An agent authorized to change a configuration may still make a dangerous change. A system allowed to issue payments may still issue the wrong payment. A model permitted to coordinate several devices may still produce a hazardous aggregate outcome.

Zero-trust architecture is foundational, not sufficient. NIST SP 800-207 addresses access decisions and trust assumptions for resources [3]. Authentication, segmentation, and monitoring can constrain or reveal activity. They do not inherently provide an independent, context-sensitive authorization decision for each consequential act. Nor does a report delivered after execution reverse a physical shutdown, disclosure, or transfer.

Replace Universal Human Review With Consequence-Aware Control

The answer is neither a human approving every routine operation nor an agent operating with blanket discretion. It is a governance layer that evaluates proposed actions before execution, under policies grounded in consequences, operating context, and available evidence. Low-consequence, well-understood actions can proceed automatically. Prohibited actions must be blocked. Actions with serious irreversible consequences and insufficient trustworthy information should require an accountable human decision.

The governing principle is straightforward: govern the act, not merely the identity or apparent intelligence of the model.

Why This Matters Most in Critical Infrastructure

Consider an AI assistant optimizing a battery energy storage system or a water-treatment process. It might have valid credentials and be communicating over an encrypted, segmented network. Those facts do not establish that its next control command is safe under current equipment conditions. Nor does verifying each individual agent guarantee that simultaneous actions from several agents will remain within system-wide constraints.

An effective design therefore needs system-level limits, reliable operational-state information, pre-action policy enforcement, auditable decisions, and a safe response when critical facts cannot be established. The details vary by application; no generic AI guardrail can replace process-safety engineering, deterministic interlocks or accountable operational authority.

The Regulatory Test Is More Demanding Than a Dashboard

OMB Memorandum M-25-21 establishes minimum risk-management practices for high-impact federal AI, including predeployment and ongoing controls [5]. Agency implementation plans describe human oversight, intervention, and accountability as part of that framework [6]. The correct conclusion is not that every AI action must be manually approved. It is that agencies must define and document what oversight is appropriate to the impact and risk of the use case, including any authorized exceptions.

A reviewer who receives an opaque recommendation and must click “approve” under time pressure is not meaningful governance. Human intervention should be reserved for decisions where it changes the outcome, supported by the action proposed, its consequences, the applicable rule, and the reason escalation occurred.

The Hard Constraint: Some Actions Must Wait

There is an uncomfortable limit to responsible autonomy. If an agent proposes an irreversible, high-consequence action and the system cannot establish the conditions needed for safe authorization, a shortage of reviewers is not a reason to proceed. The action must be deferred, blocked, or placed in a defined safe state until competent authority can resolve it. This principle is harder to market than frictionless autonomy. It also distinguishes automation from governance.

A Practical Agenda for Federal AI Leaders

Before scaling unattended agents, inventory operational use cases; complete applicable security authorizations; classify high-impact actions; distinguish reversible from irreversible outcomes; enforce policies before execution; test coordinated multi-agent behavior; and measure false approvals, unnecessary escalations, intervention latency, and audit completeness. Require vendors to demonstrate failure behavior, not simply model accuracy or a polished monitoring dashboard.

The Disruption

The next competitive advantage in agentic AI will not belong solely to whoever deploys the most agents or eliminates the most human checkpoints. It will belong to whoever can prove that autonomous work remains within enforceable authority, including when the model is wrong, the environment changes, or no human is immediately available.

The question is not “How do we remove the human from the loop?” It is “How do we ensure that no consequential act escapes governance?”

References

[1] K. S. Will, “Keeping a human in the AI loop is ‘untenable’ due to cuts, USDA IT official says,” FedScoop, Oct. 22, 2026. https://fedscoop.com/ai-human-in-the-loop-agency-workforces/

[2] U.S. Department of Agriculture, Office of Inspector General, “Cybersecurity of Artificial Intelligence Technology at USDA,” Inspection Report 50801-0018-12, May 12, 2026. https://www.oig.usda.gov/sites/default/files/reports/2026-05/50801-0018-12_FR_508.pdf

[3] S. Rose et al., “Zero Trust Architecture,” NIST SP 800-207, Aug. 2020. https://doi.org/10.6028/NIST.SP.800-207

[4] Removed

[5] Office of Management and Budget, “Accelerating Federal Use of AI through Innovation, Governance, and Public Trust,” Memorandum M-25-21, Apr. 3, 2025. https://www.whitehouse.gov/omb/information-resources/guidance/memoranda/

[6] U.S. Department of Homeland Security, “Compliance Plan for OMB Memorandum M-25-21,” Sept. 2025. https://www.dhs.gov/sites/default/files/2025-09/25_0926_cio_dhs_compliance_plan_for_omb_m-25-21_508.pdf

LinkedIn Hashtags

#AgenticAI #AIGovernance #Cybersecurity #ZeroTrust #FederalAI #CriticalInfrastructure #HumanOversight #ResponsibleAI #AITrust #OperationalTechnology