USDA official says staffing cuts make constant human oversight of AI untenable

The Agriculture Department lacks the staff and contract support to keep a person involved in every AI task, a department IT official said. Speaking at FedInsider’s October 21, 2026 Innovation Summit in Washington, D.C., Rudolf Rojas called that expectation “untenable” and outlined an alternative: let tightly restricted agents work, then report back to employees.

Rojas, USDA’s IT program manager and security lead, tied his position to cutbacks rather than confidence that AI can safely operate without safeguards. “We just don’t have the resources anymore,” he said, according to FedScoop. He described a balance between tools doing work and employees retaining visibility through restricted permissions and continuous monitoring.

An employee’s access should not become an agent’s access

The first boundary Rojas drew was around what an agent can reach. Agents should receive only the permissions needed for their work, not automatically inherit the access rights of the employee using them. “I would not want to give an agent my privileges or roles,” he said, as MeriTalk reported.

That recommendation applies to systems that do more than answer questions. The summit discussion concerned agents that retrieve information, call tools and act on employees’ behalf. For workflows using several agents, Fortinet Federal chief technology officer Felipe Fernandez recommended dividing the work and limiting each agent’s privileges to its particular tasks.

Rojas also set conditions for allowing agents to work without a person: strong risk management, analysis of agency networks and a demilitarized zone—a separated network area. He pointed to USDA’s existing security work, including employee training, multifactor authentication, vetting AI tools, defined access privileges and continuous network monitoring.

The security check sits outside the model

Fernandez identified a threat that can enter through the material an agent reads, not just an employee’s instructions. Website content and tool responses become part of the context a model uses to complete a task. He warned that this creates an opening for prompt injection: outside material introducing instructions into the agent’s work.

His proposed response was to place security and governance controls outside the agent and its underlying model. An AI gateway would monitor incoming and outgoing information, make communications visible and provide a way to interrupt access to resources. Employees responsible for agents also need a way to stop them, he said.

You can’t expect the model to protect itself.

Felipe Fernandez, chief technology officer, Fortinet Federal

CBP keeps people in the loop

The discussion did not produce a shared position on how much human involvement agencies should retain. Ajay Phogat, senior adviser to the chief information officer at U.S. Customs and Border Protection, defended keeping people involved as AI pursues assigned goals. He said that keeping a human in the loop is the approach CBP is taking.

Phogat also emphasized approved models and preventing personally identifiable information and law-enforcement-sensitive information from leaving agency networks through AI use. Before agencies move forward, he said, they need clear policies and defined workflows. His position puts continuing human involvement alongside restrictions on the models and information an agency uses.

Rojas’s argument for less constant human participation was not an argument for leaving AI companies unconstrained. He separately called for national AI rules and regulations, saying he somewhat disagreed with what he characterized as the administration’s preference for giving developers “free rein.” He cited railroads and nuclear energy as examples of industries where regulation had served a useful role.