State governments are starting to pursue agentic AI

Instances of agentic AI, in which groups of digital agents are granted permission to automate multistep tasks without human approval, remain rare in state government, but a report published by an industry association Tuesday proposes that adoption could be on the rise.

The report, published by the National Association of State Chief Information Officers, notes that while “this shift won’t happen overnight,” state governments appear to be granting their artificial intelligence agents increasing levels of autonomy. “I think it is still really new,” said Amy Glasscock, director of NASCIO’s innovation and emerging issues program and the report’s author. “But there are some states that are starting to use it.”

“It seems like agentic is going to be the next step,” she continued. “[Generative] AI started off creating and generating content and summarizing and things like that, and now states are finding ways for it to actually do some things, do some of the work, of course with a lot of oversight still.” Glasscock said a recent survey of the group’s membership, which includes most of the nation’s state and territorial CIOs, found eight states that are using agentic AI, though she declined to name them.

But by drawing on public documents and news reports, the report does name several early state efforts to use agentic AI to automate low-risk tasks. Among these are Alaska’s Department of Administration, which last November published a request for information, noting imminent plans to add agentic capabilities to the state’s mobile app. The document cites interest in “proactive notifications, reminders and personal recommendations for relevant state services,” “context-aware assistance, including dynamic form-filling, document retrieval and eligibility checks,” along with proposed back-office uses of AI, such as “facilitating autonomous data exchange, workflow orchestration and service fulfillment across disparate platforms.”

Former Virginia Gov. Glenn Youngkin last summer issued an executive order, kickstarting “a first-of-its-kind pilot program” that used agentic AI to scan all of the commonwealth’s regulations and guidance documents, and then spot redundancies and suggest “streamlined language.” His office dubbed it the “Virginia Model” of regulatory modernization, claiming that AI would “supercharge the regulatory reduction process.” Youngkin’s order is no longer hosted on the state’s website, however, and Abigail Spanberger’s administration hasn’t mentioned whether it plans to continue using agentic AI.

In her report, Glasscock imagined five phases of AI maturity that states might pass through, starting with the “assistive” generative AI tools, in which humans are the “primary doers and decision makers.” A majority of states have reported such basic uses of AI. Phase two is the “more sophisticated” “context-aware” generative AI. Phase three is “task-level automation,” in which agents are permitted “limited actions inside defined boundaries,” like pre-filling a form for a resident using data that had already been collected.

Phase four is “stateful, multi-step workflows,” in which agents retain context about users, not just the most recent conversation, and progress is managed over time. This might be used for employee onboarding or guiding residents’ applications for benefits. In phase five, the report says, “AI initiates work before anyone asks.”

Glasscock pointed out that government tends to lag behind the private sector’s adoption of new technologies, and that it could be years before fully autonomous AI agents are allowed to roam state agency networks. One barrier to adoption, cited in Tuesday’s report, is that state governments maintain a lot of outdated computer systems that aren’t always compatible with other systems, to say nothing of newfangled AI. The other barrier, of course, is that states are highly risk averse, and for at least some CIOs, agentic AI could be a hard sell.

Glasscock said that within the next couple of months, she thinks states will begin to talk about their early uses of agentic AI. (The association will host its midyear conference, in Philadelphia, in late April.) She imagined that early uses of agentic AI may include small, discrete tasks carrying minimal risk. In 2026, phase-five agentic AI is harder to picture operating inside a state government: “I think for automation, there’s some real possibility for that,” she said, “but not just doing all the jobs and making all the decisions any time soon.”