Urban Institute releases guidance for state and local agentic AI adoption

Agentic AI is among the newest frontiers for state and local government innovation as leaders deliberate how to capitalize the tech to find efficiencies, drive down costs and provide residents with innovative services. As the public sector begins its venture into agentic AI solutions, a new playbook offers guidance to state and local officials to govern and implement the technology responsibly. 

Agentic AI differs from traditional and generative versions because it not only generates content based on training data, but it “often has some autonomy to make future decisions or to think about future decisions and plan then take action,” Graham MacDonald, chief information officer and vice president of technology and data science at the Urban Institute, told Route Fifty.

Despite agentic AI’s early days in public sector adoption, states and localities have begun introducing the technology into their government operations in recent years in areas like public health, housing and public assistance, according to an agentic AI playbook authored by MacDonald last month.

Former Virginia Gov. Glenn Youngkin last summer announced a pilot program during which the state would use agentic AI to modernize regulatory and guidance documents. The agentic AI scanned and identified where state officials could streamline redundant or burdensome language in such documents. The initiative helped uncover more than $1.4 billion in annual savings and reduce permit and license processing times by nearly 80%, according to a January report from the Virginia Office of Regulatory Management. 

The Utah Office of Artificial Intelligence Policy launched a pilot program earlier this year to evaluate how an autonomous AI system could assist with routine prescription renewals. The pilot aims to address gaps in prescription access, reduce procedural delays that lead to medication lapses and broadly improve health outcomes for people with chronic conditions. The office partnered with health platform Doctronic in January, and preliminary results released in May showed that the tool suggested a prescription renewal for 72% of cases, and of those cases, a physician agreed with 91% of the AI’s decisions. 

In Boston, Chief Information Officer Santiago Garces late last year started integrating agentic AI into the city’s open data portal in an effort to uncover potential actionable insights. The city is using a model context protocol server from Anthropic to analyze city data and inform queries related to policymaking and broader data-driven decisions, Garces told StateScoop in April. 

“What’s different about the agentic generative AI is that the IT person is not the only person who is empowered to build it,” MacDonald said. While that expands access to innovative and advanced tech like agentic AI, it also introduces opportunities for errors in using such tools, workflow disruptions and other pain points that can negatively impact people, particularly when AI agents are used in public problems like benefits administration, he explained. 

The agentic AI playbook is particularly valuable for closing knowledge and governance gaps between staff who are familiar with the technology or specific AI products they are building and the people using or reviewing the end product, MacDonald said. For instance, if organizations require a human in the loop to review content produced by an AI agent, such reviews may be misinformed if the human reviewer doesn’t understand the system and how it works in the first place, he explained. 

To help prevent such lapses in responsible agentic AI adoption, the playbook recommends that AI system builders clearly document baseline components, such as the tool’s intended purpose, target population, measurable outcomes and estimated costs. 

The playbook also suggests that users consider more closely how to define accuracy and fairness in how their agentic AI system performs. For instance, an important part of assessing an AI’s tool accuracy is its quality, which includes broader factors like the completeness, clarity and reliability under variation of an AI agent’s output. 

State and local officials should also determine if a prospective AI agent is designed for internal government use or for a constituent-facing purpose because such use cases could change how agencies approach the development phase, according to the playbook. Internal AI agents can be piloted and launched at much smaller scales that could require less stringent guardrails when it comes to fairness and bias, MacDonald said. 

If the tool is being designed for external use, officials should consider how they are defining and implementing fairness metrics, which are one way states and localities can further build responsibility into their agentic AI systems. 

Governments should keep in mind that such metrics are “one of the hardest to think about and implement … so this is something deeply unique to every problem,” MacDonald said. As an example, he pointed to an agency that plans to roll out a public-facing AI agent to help constituents navigate applying for disability claims. This presents an opportunity for officials to rigorously test whether the product generates different or better recommendations from one population to another. Another fairness metric could include testing the percentage of time an AI agent offers a correct answer across diverse users, he explained. 

Government leaders could also establish dedicated roles among government or agency staff to oversee and monitor the responsible use and governance of agentic AI, according to the playbook. MacDonald identifies seven core roles — accountable owner, evaluation lead, security lead, transparency lead, responsible AI lead, contributor and independent review — to assume a range of tasks aimed at ensuring agentic AI systems continue to operate adequately. 

Such tasks include maintaining audit trails of an agentic AI system, assessing the tool against evaluation templates or rubrics, collecting stakeholder feedback and monitoring harms imposed by the AI agent, among others, according to the guidance. MacDonald also noted that many jurisdictions are practicing this principle through AI governance committees and similar organizations. 

Indeed, “I think it’s incumbent on everyone in their organizations to establish a good process,” and part of that means “bringing more people into the process” to ensure a government or agency can develop agreed-upon goals, metrics and standards that drive agentic AI adoption, MacDonald said.