Federal agencies explore agentic AI but governance lags behind adoption
More than half of federal technology executives are exploring agentic AI or planning pilots, according to a March survey of over 200 government IT leaders. Another 15% have already implemented the systems, while just 6% haven't started considering them.
The gap between interest and readiness is significant. Seventy-seven percent of federal leaders say oversight frameworks are essential, yet fewer than one-third have actually built them. Only 20% of agencies have defined policies for testing agentic AI before deployment, and just 8% have incident response frameworks.
Agentic AI systems perform complex tasks with minimal human oversight - reasoning through problems and taking independent actions across software systems. The technology appeals to agencies trying to do more with shrinking workforces, but the infrastructure to manage it safely hasn't caught up.
The oversight problem
Agencies recognize what they need to do. Nearly 90% of respondents require logging and audit trails for all AI actions. More than 80% demand automated policy checks and guardrails. Yet fewer than half include liability clauses in vendor contracts, and only 29% have documented "kill switch" procedures to stop systems if they malfunction.
The mismatch is stark: federal leaders say they want human control over high-risk AI, but lack the mechanisms to enforce it.
Data readiness slows pilots
Agencies struggle moving agentic AI from test environments to production. The barrier isn't just policy - it's data. Government systems often run on fragmented, disconnected data sources that aren't prepared for AI consumption.
One unnamed IT director flagged the core issue: "How am I getting my data ready for AI consumption? That governance piece becomes critical to making sure that your data within your organization and your AI are working together."
Risk levels determine oversight intensity
The survey shows agencies calibrate human oversight based on data sensitivity. For national security, critical infrastructure, and emergency response data, 79% require human approval for every AI action. For high-risk data like benefits claims or financial records, 78% require formal approval before high-risk actions, but not every action.
For low and moderate risk data, more than 90% of agencies favor reduced direct involvement, requiring only periodic check-ins.
Context
Federal AI use more than doubled in 2024, according to the Office of Management and Budget's annual inventory. Agencies reported over 3,000 AI use cases, with significant increases at NASA and the departments of Health and Human Services, Veterans Affairs, Justice, and Energy.
This expansion happened despite a decrease in the total number of federal employees, underscoring why agencies see agentic systems as a way to maintain output with fewer staff.
Learn more about AI Agents & Automation and AI for Government implementations.
Your membership also unlocks:
