States Governments Seek to Leverage AI’s Promise While Mitigating Its Hazards

This is one in a series of five articles examining key debates that will unfold in the nation’s statehouses in the year ahead.

As states move from studying generative artificial intelligence to actively deploying it, the benefits and risks of integrating this rapidly evolving technology into government operations are starting to become clearer—with the potential for wide-ranging fiscal implications on both sides of the budget ledger.

Many government agencies are now actively experimenting with functions that could transform how they operate and interact with residents and businesses. Amanda Crawford, president of the National Association of State Chief Information Officers (NASCIO), compares this part of the process with the reality-sinking-in stage of a classic romantic comedy. The initial excitement has settled down—and the plot is thickening.

“Now we have to decide: How are we going to make this relationship work?” said Crawford, who also serves as the chief information officer (CIO) of Texas. “Have we set up the appropriate boundaries and guardrails? And we also realize that maybe we need to work on ourselves a little bit—cleaning up our own practices, whether it’s our data, or our governance, or guardrails.”

At this point, both the costs and benefits of generative AI are moving targets, and state leaders have been racing to harness its potential while bracing for plot twists. Texas agencies, for example, are already using cybersecurity tools, public-facing chatbots, and a whole host of other technologies that leverage generative AI, Crawford said. At the same time, the state is aggressively auditing and refining the tools on an ongoing basis, encouraging state and local employees to learn from one another’s experiences, and finessing its statewide strategy.

The long-term costs of AI are unpredictable because of rapidly evolving business models in the industry—many of the major players are taking losses during this period of explosive development but will need to recoup those costs at some point. And although many state officials expect the eventual benefits and cost savings to be widespread, plenty of trial and error will be required before they have clarity on which investments can be made with confidence and how to manage them effectively.

Building a foundation

States undertook a flurry of executive and legislative activities in 2024 to establish policies and governance aimed at ensuring the responsible and ethical use of AI and the data that powers it. According to the National Conference of State Legislatures, states considered more than 150 bills related to government use of AI last year, and at least 30 states have put guidance in place for state agency use, sometimes including requirements for inventories and impact assessments. Several have recently released—or are expected to release soon—task force or council reports intended to inform upcoming legislative and executive branch actions on these issues.

Data quality is likely to be an area of emphasis this year. AI tools can bring new relevance and usefulness to obscure government datasets—but this relevance could backfire if the tools are fed a steady diet of unreliable or biased data.

According to a NASCIO report released in November, only 24% of state CIOs say their state has data governance in place for generative AI—a shortcoming that is holding states back. Still, NASCIO anticipates that many states will make long-deferred investments in data quality and accessibility in the coming years, an essential foundation for the more advanced future AI uses that they are building toward.

“As we think about what fuels the AI—we have to recognize, even with the many exciting innovations—we still need ground truth data,” said Theresa Pardo, associate vice president for research at the University at Albany and one of the authors of a primer on AI for government managers. “It’s not very sexy but it’s absolutely necessary—and it’s going to be expensive.”

Despite these complex challenges, many states are moving forward with a growing urgency that reflects the evolving expectations of their workers and constituents. As AI becomes increasingly woven into commerce and daily life, state workers are sometimes unofficially introducing their own free AI tools into their workflows—a risky reality that leaves states without the option of sidestepping the technology’s impact altogether.

“In the next three to five years, states are expected to see significant growth and development in the use of generative AI, akin to the early days of the internet,” the NASCIO report said. “The potential applications are vast.”

States of adoption

States vary wildly in their progress with AI adoption—some are moving quickly, while others are proceeding with great caution. And 8% of state CIOs report that their states aren’t using generative AI tools at all, according to NASCIO. Among the adopters, the most common current uses are virtual meeting assistant transcription, cybersecurity, document generation and management, and software code generation.

Many states are focusing on improving operational efficiency and cleaning up their systems, data, and governance to pave the way for more complex and higher-risk uses in the future—such as AI tools that interact directly with residents.

Among early AI explorers, California stands out for its careful but ambitious approach. The state has been assessing generative AI in state government since 2023, following Democratic Governor Gavin Newsom’s executive order that established a comprehensive framework for its study, testing, training, and evaluation. Through these efforts, the state identified five promising use cases—related to translation, call center productivity, traffic analysis, road safety, and health care facility inspections—that have the potential to scale across government. For each use case, officials then issued a “request for innovative ideas.”

The technical specifications that are usually central to IT procurement were notably absent from these requests. “We really wanted to focus more on what’s out there, especially for a technology like this, where we’re not entirely sure of what the best solutions could be,” said Jonathan Porat, California’s chief technology officer. Vendors worked closely with the state during the second half of 2024 to prove their concepts in the state’s “sandbox” environment.

For example, customer service agents from the California Department of Tax and Fee Administration handled more than 800,000 calls, emails, and chat inquiries during the 2023 fiscal year—more than they have the capacity to manage effectively during the height of tax season. “Inaccurate information or delays in assistance adversely impact our taxpayers trying to comply voluntarily with their tax and fee obligations and can delay or diminish revenue for the state,” the agency’s problem statement said.

Agents conduct detailed research about the agency’s 41 complex tax and fee programs on the fly—consulting voluminous manuals and reference guides—and sometimes need to ask customers to wait while they ask for help from colleagues with more experience. Now, the state is working with vendors to test approaches that use AI to quickly comb through reference materials and assist agents with responses. The goal is to improve the experience of both customers and agents by reducing call lengths, wait times, and abandoned calls—while ensuring accuracy. Successful experiments will soon move into more formal development, clearing the way for new ideas that are entering the testing pipeline.

Colorado’s Department of Local Affairs is using AI to streamline its approach to process improvement. Traditionally, the laborious first step of improving government processes involves gathering information from all of the players who touch a particular process and mapping that out, often by huddling in conference rooms for hours on end with Post-it notes, whiteboards, and sometimes strong emotions.

But the state’s recent efforts to streamline administration of its housing voucher programs sidestepped all of that. Instead, a “computerized consultant” named Coco, trained on videos of a process improvement expert, conducted one-on-one interviews with dozens of stakeholders and created detailed process maps.

“Staff were pleasantly surprised by how natural the conversation with Coco was and how effective Coco was in asking follow-up questions,” said Maulid Miskell, the department’s deputy executive director. Armed with impressive process maps and guiding information from Coco, staff members are now working to comprehensively redesign these processes so they are more efficient and user-friendly, he said.

Workplace pilots

The most powerful current uses of AI in state governments may simply involve state employees finding ways to speed up their work by outsourcing some day-to-day drudgery.

Rhode Island’s Office of Management and Budget has begun using AI tools as a sort of well-rounded graduate student intern, said Director Brian Daniels. The agency has used the technology to conduct background research, improve the readability of documents, and even to help draft a request for proposal (RFP) by relying on examples as inputs.

“That was very exciting because writing an RFP is one of the most annoying things you can do in government,” Daniels said.

Pennsylvania, Colorado, and Utah have deployed AI tools like Gemini for Google Workspace and ChatGPT Enterprise to a cross section of employees through carefully controlled pilot projects. By equipping employees across the far-flung reaches of state government with these tools, governments hope to identify opportunities and pathways to making them more widely available.

In Pennsylvania, employees piloting ChatGPT initially experienced steep learning curves but now they’re racking up time savings—more than an hour a day on average, according to the governor’s office.

Meanwhile, Colorado is rolling out Gemini statewide after a promising pilot program that reported nearly universal increases in worker productivity, ranging from basic research to spreadsheet analysis. More than one-third of users said they saved at least six hours each week.

But most striking to David Edinger, the state’s CIO, was the human impact. The experience was especially meaningful to employees with disabilities or neurodivergence, who said AI helped create transformational changes in their workflows that better align with their skills and the way their minds work.

“I discovered something that allowed me to use my unique way of thinking and processing in a way I had not experienced before,” one respondent wrote in response to a state survey aimed at assessing the pilot’s impact. “Further, it elevated the same traits and skills that have traditionally isolated me, caused misunderstandings, and sometimes reputational/professional harm.”

Still, not everyone is convinced that productivity savings will be worth it in the long run. Alan Fuller, Utah’s CIO, said the state pays more than twice as much per user for the AI component of Gemini as it does for the entire Google productivity suite—a significant cost that only pays off if the workforce embraces and consistently uses the tools.

“It’s sort of like a gym membership,” he said. “If you use it, you’re probably getting value out of it, right? But if you’re not using it, why are we paying for it?”

Investing in the long haul

Although AI offers the promise of efficiency and savings, the path to understanding and implementing it in government continues to be a tricky one. It’s exciting, full of peril and promise, and the terrain is constantly changing. The self-work and training that states are doing now will take time and money to get right.

“We have this mistaken premise that AI systems are end-to-end systems that can be procured off the shelf and just dropped into a setting,” said Daniel Ho, senior fellow at the Stanford Institute for Human-Centered Artificial Intelligence and a professor at Stanford University. “The reality is way more complicated than that.”

Melissa Maynard oversees The Pew Charitable Trusts’ Fiscal 50 project.