U.S. state governments are steadily expanding AI use by prioritizing strong governance, security, training, and employee engagement, while learning that successful adoption depends on usability, trust, iterative implementation, and alignment with workforce needs and modernization efforts.
In October 2025, the IBM Center for the Business of Government published a report by Katherine Barrett and Richard Greene titled “AI in State Government.”
In the six months since then, the world of AI has been moving forward at a rapid clip, prompting the IBM Center to ask for a follow-up post on how the use and expansion of generative AI is progressing, signals of the process improvements achieved, and lessons learned. For this research update, information technology leaders and program managers were interviewed from California, Massachusetts, Nevada and New Jersey. A more extensive profile of state policy and governance, with new use cases from a wide variety of states, will be released in early 2027.
Watch our webinar as state leaders from across the country gathered for a fireside chat on how agencies are turning early success into scalable, sustainable programs. Leaders include: John Harrison, Director of Information Technology, New Jersey Department of Community Affairs; Molly Lennon, Administrator, State of Nevada Department of Motor Vehicles; and Dr. Dave Sloan, Chief Information Officer, Maryland Department of Human Services.
In the first half of 2026, a clear message from top level state sources was that expanding the use of AI needed to be handled with clear governance: care, extensive communication, continued policy refinement, training, and attention to security, privacy and data quality. “When we approach these types of projects in the public sector, we want to take our time and have a thoughtful approach to preserve the public trust,” says Jared Johnson, California Deputy State CIO and Chief Deputy Director in the California Department of Technology.
One of the top lessons learned was the critical need for residents and frontline staff to find AI advances accessible and easy to use and understand. When expanding AI technology, “it doesn’t matter if we have something shiny and bright that we think works well. If they can’t use it, it’s useless. We want to make sure, through design thinking, that we understand resident and staff pain points and help to resolve them,” says Laura Glenn, IBM program manager for utility assistance at the New Jersey Department of Community Affairs.
Some common themes emerged in these conversations, including the need for:
· A constant and continuous emphasis on training and the need to adjust training approaches depending on employee roles.
· Steady progress with an emphasis on security, attention to data accuracy and the protection of private information.
· Continuous modernization of non-AI technology, which doesn’t align well with new AI systems.
· Engaging the state workforce with a focus on change management, accompanying technological development.
· A focus on upskilling the workforce as new AI tools and approaches alter employee roles.
· Attention to prioritization of how and where to expand, with a view of limited resources, and a focus on problems that can most effectively be addressed with AI solutions.
· Interest in innovation and the ways that AI use cases can uncover solutions to persistent problems, increase productivity and encourage creative thinking.
AI Expansion and Employee Buy-In
By the end of calendar year 2025, many states had already developed significant policy statements and new AI governance practices, with the October 2025 report outlining a variety of approaches to the oversight of AI and the interplay between central executive offices and departments. In our interviews with a dozen state CIOs and other high-level technology officials during summer and fall 2025, many of our conversations focused on “use cases” that were generating excitement but were still in a pilot or early stage of development. As NASCIO executive director, Doug Robinson told us, “I see calendar year 2026 as the major year of a lot of implementations.”
Not surprisingly, as AI use has expanded, that implementation is requiring attention to state employees, how they adapt to new tools and risks, as well as a potential shift in their day-to-day work life.
In both California and Massachusetts, our conversations this spring focused not only on the importance of employee training, but on working with employees as job functions change and new skills are needed. For example, a product that’s soon to emerge in California is a digital productivity assistant that has been piloted in 67 different state departments with 2,800 users. This product is intended to assist with administrative tasks and general productivity. It was engineered with a focus on security and privacy, with a current plan to release it into full production in July. “It was built by state employees for state employees,” explains Johnson. “We brought in many employees across the state, doing different types of jobs, to provide real time feedback.”
For Massachusetts, a focus on employees is centered around its plan to roll out AI tools to all 40,000 executive branch employees. According to a February 2026 press release from the Governor’s Office, this makes Massachusetts the “first state to adopt the technology for the entire executive branch.”[1]
As this ambitious plan emerges, each employee will have access to an array of self-paced voluntary training and will be asked to agree to terms and conditions prior to using ChatGPT in their protected personal space. They will be able to consider the possibilities for generative AI use in their work and that of their department. This use will be considered as personal notes, similar to the way other personal work notes, such as drafts or brainstorming, are treated.
When employees share elements of AI-assisted content and it becomes part of an official government record -- – for example, in an email or office document -- it becomes subject to external review to ensure no private or internal state information is improperly included.
By rolling out ChatGPT to every state employee, “we wanted to democratize innovation with AI,” says Jason Snyder, Massachusetts CIO and Secretary of the Executive Office of Technology and Security. He says that he has already begun to see gratifying new and different use cases that come from employees at all levels rather than through a more traditional “top down” approach. In that way, employees who are not necessarily tech savvy can consider innovative ways that AI can address issues within their own departments or areas of expertise.
For example, the Executive Office of Economic Development (EOED) launched an AI-powered Expression of Interest (EOI) tool in early 2026 to help streamline and strengthen the state’s Community One Stop for Growth application process.[2] The tool allows applicants to submit a short project description and answer targeted follow-up questions through an AI chatbot interface. It then generates immediate, project-specific guidance with applicants receiving real-time, on-screen feedback as well as a personalized feedback report via email. According to the department, this new AI launch is already helping communities and organizations navigate the application process more efficiently and with greater clarity.
One general challenge that Massachusetts and other states face is tangible employee resistance to AI. Snyder says the state has found that about 18 to 20 percent of employees have such strong concerns about AI that they don’t want anything to do with it. An additional similar percentage have “real concerns about the use of AI” but training can help to alleviate them.
Though the state had anticipated a faster rollout, by early May, 4.500 of the state’s employees were certified and licensed to begin using ChatGPT tools.
A Departmental Perspective
Not all AI efforts are conducted statewide. In fact, a good deal of action is happening at the departmental level, where expansion of AI tools and approaches can move forward in multiple stages.
In the New Jersey Department of Community Affairs, for example, the creation of a resident portal and case management system had already started on July 1, 2021, almost 18 months before news began to explode about the potential for generative AI. The foundation for a Data Lakehouse[3] was added the following year, but it wasn’t until Fiscal Year 2024 that the portal included document intelligence capabilities through generative AI.
That addition helped to dramatically reduce the time that case managers needed to examine and verify documents that confirm eligibility for the Low-Income Home Energy Assistance Program (LIHEAP), an important element of the state’s safety net.
AI tools not only increasingly saved caseworker time but also sped up the eligibility determination process. “The goal for us is for technology to reduce the administrative burden so the case worker can focus more on the residents,” says IBM program manager Glenn.
Glenn notes that the advent of AI use has been moving forward on “tippy toes” to make sure that the documents that are used to confirm eligibility can be easily read by the technology and are useful. Four types of documents were initially tested and approved to verify applicant address, household members, income and utility cost
As the use of AI tools expanded, it became clear that other documents should be added, including tax forms, court orders, child support letters, veteran benefit information, and many others. One concern raised by the expansion of documents has been cost and has presented the need for rigorous testing. An issue that can present a challenge is the inconsistency of the way information is presented by different sources. If there are too many variables, the Large Language Model used by AI, may not read the document correctly.
This has led to work by business analysts on what documents need to be run through document intelligence and potentially which documents would work better. “If we run too many documents through, the cost increases,” says John Harrison. Director of Information Technology at New Jersey’s Community Affairs department. “We’ve had to monitor the cost of the documents we are reviewing.”
Another departmental expansion in New Jersey is in its Department of Community Affairs’ program called “Ask Claudia.” It is a tool that case managers have the option to use that replicates the knowledge of a real department employee named Claudia, who is close to retirement from state service and has accumulated a wealth of knowledge that can be used by others in the department. This generative AI tool accomplishes this through knowledge transfer input she has provided and through its ability to keep up with regulations and policies that it can scan, providing help to LIHEAP employees in accurately answering questions that come from the public.
The pilot has proved the value of this effort, and is now a full program, that is also being used for the Universal Service Fund, a state bucket of money from the Board of Public Utility, a sister agency that is a partner to the Department of Community Affairs, says Harrison. He adds that “Ask Claudia” has been so popular with caseworkers in the LIHEAP program that it has also generated requests for similar AI assistance from the department’s weatherization and lead grantors.
Progress in Nevada (and lessons learned there)
The development of AI services in Nevada is occurring in conjunction with an eight-year modernization project that is now in year five. That effort, which has moved the state from an antiquated mainframe environment, received the NASCIO “Technology Champion Award” on April 28, 2026 at its midyear conference.
AI has been a focal point at the Nevada Department of Motor Vehicles, given the heavy customer contact, typical of DMVs, the fast-growing Nevada population and employee numbers that haven’t grown since the early 2000s.
In the Nevada DMV’s development of AI, a goal was replacing its pre-AI chatbot, which operated on a set of answers that responded to most common questions and often resulted from the plea from DMV clients to speak with an actual live representative. “We learned very quickly, both from our constituents and from staff, that there were concerns with the canned answers,” says Molly Lennon, DMV administrator.
The DMV also learned that its initial effort to move to an AI-enabled chatbot without adequate preparation was a mistake. After three months of use in summer 2024, that chatbot version was pulled down when department leaders realized “we hadn’t done enough work on our knowledge base to make sure that what it was generating was truly accurate,” Lennon says.
In September 2025, after focus on the knowledge base, and a shift in contract, an AI-enabled text chatbot was launched again, providing helpful service to DMV clients at a time when the state was recovering from a major ransomware attack and when school was starting (a period when DMV requests tend to increase).
The impact of the AI chatbot has been a decrease in people asking to escalate to a staffer. More AI capabilities are still being added to the chatbot with the goal of cutting down dramatically on the emails and calls that have consumed staff time in the past. The chatbot is still all text, with a shift to voice about a year away.
This is only one area in which Nevada has been adding AI components as part of the state’s technology modernization efforts. Others include training and software testing and AI Assisted/Intelligent Automation for workflow and queue management. This all has happened in conjunction with a refresh in January, 2026 to Nevada AI policy, which had originally been published a year before, and the creation of DMV’s own AI policy document.
Despite its excitement about AI opportunities, the effort is to move at a moderate pace to stay in line with policy and build a strong security posture. AI governance and structured data checkpoints focus on “feasibility, privacy, security, accessibility and return on investment,” according to Suzie Pollard, DMV CIO and IT administrator.
In adding AI features in DMV, leaders took an iterative approach. “We delivered in increments and then we measured and then we adjusted based on what constituents and staff were experiencing. So, we didn’t try to bite off more than we could chew,” says Pollard.
Looking at the implementation and expansion of AI initiatives, one of the most challenging aspects has not been the technology itself but dealing with team culture and the high degree of employee resistance to change. To deal with that issue, and the fear that people had of ultimately losing their jobs, the DMV has focused on “explaining the why behind every shift”, by emphasizing the importance of learning new AI skills, and by inviting employees into “solution thinking and by showing them how they could help move the work forward.”
Added to that, listening to employees was critical. “When people felt heard, acknowledged, included, even if the final outcome wasn’t what they hope for, they remain more open and connected to the process,” says Pollard. “Building trust and creating space for input has been just as important as the technical work itself.”
[1] “Governor Healey Announces Massachusetts to Become First State to Deploy ChatGPT Across Executive Branch,” Executive Office of Technology Services and Security and Executive Office of Economic Development, February 13, https://www.mass.gov/news/governor-healey-announces-massachusetts-to-become-first-state-to-deploy-chatgpt-across-executive-branch
[2] “Community One Stop for Growth is the “main vehicle for public economic development investment in the Commonwealth,” according to the description on this Massachusetts state website, https://www.mass.gov/info-details/community-one-stop-for-growth
[3] A Data Lakehouse pulls from and connects with multiple technology sources. For example, it can include individual data warehouses.