Federal AI adoption metrics miss the human factor

The State Department’s chief information officer has told the story of how the agency placed StateChat, its internal AI chatbot, on an employee homepage only to have one employee still ask whether its use was permitted. The problem was not necessarily unwillingness to try new technology. In this case, an employee remained unsure about authorization despite having access to an agency tool. An AI adoption dashboard counting available licenses or logins would miss that distinction.

A change to federal employee surveys gives agencies an opportunity to investigate such obstacles systematically. Under a July directive from the Office of Personnel Management, agencies must administer their own annual employee surveys starting in 2026, rather than participate in an OPM-run Federal Employee Viewpoint Survey. OPM’s proposed rule would reduce the required common questions from 16 to 10, emphasizing performance, feedback and organizational goals. None of those proposed questions asks directly about AI.

That matters to technology leaders as much as human capital officers. The Government Accountability Office found that reported generative AI use cases at 11 selected agencies increased from 32 in 2023 to 282 in 2024. Those inventories establish that agencies are experimenting. They cannot establish why an authorized tool is unused, whether employees trust its answers or whether managers are inadvertently discouraging responsible experimentation. OMB’s federal AI guidance calls for understandable, implementable risk-management policies. Agencies need evidence that their workers actually understand those policies.

A Federal Reserve survey released in May found that 25% of employed Americans used generative AI at work during the previous month, while 20% worried it could replace their jobs. The Federal Reserve Bank of Boston found that concern about personal job loss from AI roughly doubled between its late-2024 and late-2025 survey waves. Neither sample describes the federal workforce specifically. That limitation is precisely why agencies should collect their own evidence rather than extrapolate from national attitudes.

Measure the cause, not just the adoption rate

In my work helping organizations adopt AI, I distinguish among three recurring responses. AI alarmists worry about economic security or losing the professional expertise that gives their work meaning. Reluctant adopters can see potential benefits but hesitate to experiment openly, sometimes because they are unsure what managers approve. Pragmatic resistors question accuracy, data protection or whether the technology suits their mission. These concerns overlap, and a single employee may experience all three.

The distinction determines what an agency should do. A person afraid of replacement needs credible information about how duties could change. Someone uncertain about whether a chatbot may handle internal documents needs clear rules and examples. Someone who has caught an AI-generated error in an eligibility determination may be providing a valuable warning, rather than obstructing modernization. Agency leaders have acknowledged job-replacement fears as a barrier to adoption. Treating every objection as a need for another training course would ignore these differences.

A short, properly authorized set of supplementary survey items could diagnose them. Agencies might ask employees to rate agreement with five statements: “I worry that AI could reduce the need for my role or skills.” “I understand which AI tools and tasks are approved for my work.” “I have had a chance to test an approved AI tool on a useful, low-risk task.” “I know how to verify AI-generated information before using it in my work.” “I can raise concerns about AI errors or inappropriate use without negative consequences.” An open-ended question could ask: “What change would make approved AI more useful or safer for your work?”

These questions should be piloted for clarity, with responses analyzed by broad mission function where sample sizes permit. A blanket AI adoption score would conceal whether acquisition employees distrust outputs, if benefits staff lack approved use cases, or field offices lack access to tools. Agencies should preserve confidentiality and avoid using the survey to identify individuals suspected of resistance or unauthorized AI use. OPM’s 2024 FEVS technical report describes suppressing small demographic response groups to protect respondents. Any agency-run instrument needs similarly explicit privacy and reporting safeguards.

Survey designers must also address a real regulatory obstacle. A September notice from OPM said agencies adding questions to its common form would need their own OMB control numbers for those separate information collections. The notice requests comments through Oct. 19. Chief human capital officers and chief AI officers should involve their information-collection specialists early, rather than assume they can append questions without additional approval.

Connect employee feedback to deployment decisions

The results should change agency behavior. Where job security dominates, leaders need to explain which tasks AI is intended to support, where human judgment remains essential and how employees can develop relevant skills. They should acknowledge uncertainty about future staffing instead of making promises they cannot guarantee. Where permission is the issue, CIOs and program owners should publish task-specific examples of what is authorized, what information must stay out of AI systems and when human review is mandatory.

Where accuracy or sensitive information is the concern, employees should help design tests using realistic, low-risk work. Agencies can compare AI-assisted and existing processes on errors caught, time spent verifying outputs and the quality of final decisions. The National Institute of Standards and Technology’s generative AI risk profile provides a framework for examining these risks. A skeptical employee who spots a recurring failure mode can improve an AI deployment more than one who quietly complies with a mandate.

This requires shared ownership. The chief human capital officer should manage survey validity and confidentiality; the chief AI officer should connect findings to approved use cases and governance; and CIOs and mission leaders should address access, workflow design and technical controls. They should review aggregated responses alongside authorized-tool usage, help-desk questions and independently measured task quality. Correlations between attitudes and outcomes can identify where to investigate, but cannot by themselves prove what caused an adoption problem.

The GAO’s July 2026 federal training guide stresses aligning training with mission outcomes and evaluating actual performance. That offers a better response than assigning the same AI-literacy module to every employee. A unit struggling with verification needs practice and review standards; one whose employees are uncertain about permission needs better policy communication; one confronting job anxiety may need an honest management conversation before another workshop.

Agencies can then report back what they heard, what they changed and what improved. Follow-up assessments, obtained under appropriate approvals, should test whether employees now understand permitted uses, report errors more confidently and complete work accurately. This is the practical value of studying AI adoption at work: leaders can address the human barriers that licenses, training counts and AI inventories fail to reveal.

The federal survey overhaul could become more than a change in who administers the questionnaire. If agencies use it to diagnose the reasons behind hesitation, they can make smarter AI investment and deployment decisions while preserving the scrutiny that public service requires.

Gleb Tsipursky, PhD, is a behavioral scientist, CEO of Disaster Avoidance Experts and the author of The Psychology of AI Adoption at Work: From Resistance to Results.