State and local governments are beginning to test agentic AI in public health, housing, regulatory work, public assistance, and data analysis. Unlike conventional generative AI, which produces content in response to a prompt, an AI agent can plan a series of steps, use tools, make decisions, and take actions with some degree of autonomy. That expanded capability creates potential operational benefits, but it also requires agencies to reconsider how AI systems are designed, approved, and monitored.

A new Urban Institute playbook provides a framework for this. It recommends that agencies document the intended purpose of an AI agent, the population it will affect, the outcomes it should produce, and its expected cost before development begins. These details give technical teams, agency leaders, and reviewers a common basis for deciding whether the system is appropriate and for measuring its performance.

The guidance also distinguishes between internal and public-facing applications. An internal agent used to summarize operational data can often be tested on a limited scale. A resident-facing agent that supports disability claims, housing applications, or public benefits requires tighter controls because an incorrect recommendation may affect someone’s access to essential services. Agencies therefore need to define accuracy in practical terms, including whether outputs are complete, clear, consistent, and reliable when the underlying conditions change.

A system may achieve a high overall percentage of correct answers while performing poorly for a particular group. Agencies deploying public-facing agents should test whether recommendations vary across different populations. Human review alone does not resolve the problem if reviewers do not understand how the agent works or what information shaped its output.

The playbook proposes seven roles for governing these systems: an accountable owner, an evaluation lead, a security lead, a transparency lead, a responsible AI lead, a contributor, and an independent reviewer. Smaller governments may assign several responsibilities to the same person, but the underlying functions must still be covered. Agencies should know who approves the system, who tests it, who examines security risks, who documents its operation, and who can independently challenge its performance.

Early government projects show the range of possible applications. Virginia used agentic AI to review regulatory and guidance documents, reporting more than $1.4 billion in annual savings and an almost 80% reduction in permit and license processing times. Utah is evaluating an autonomous system for routine prescription renewals. Preliminary results indicated that renewal was recommended in 72% of cases, with physicians agreeing with 91% of those recommendations.

These results show why agencies are interested, but deployment cannot end with a successful pilot. Governments need audit trails, evaluation criteria, stakeholder feedback, and continuing monitoring for errors and harms. Agentic AI introduces more autonomy into public systems. Its value will therefore depend on whether agencies can establish clear responsibility for every decision the technology helps make.