From Pilot to Playbook: How Missouri Is Scaling AI Across State Agencies

How Google Won Missouri’s Informal “Gen AI Competition”

Tim Marczewski came to Missouri from the federal government, where he worked at the Pentagon and the National Reconnaissance Office. He is now the state's director of AI and innovation and serves as its chief AI officer. Soon after he arrived, the state ran a competition among enterprise AI chat tools. About 200 employees across 17 agencies tested them against use cases they submitted themselves.

Google's tool “won out on the objective and for qualitative and quantitative reasons,” Marczewski said. Missouri then connected the tool to state systems, including the state's Microsoft 365 environment, and began rolling out agents and skills.

Employees reach those tools through GenAI.mo.gov, a single portal modeled on the federal government's GenAI.mil. The portal also delivers compliance training on the state's AI use policy and its risk management framework, which is based on the National Institute of Standards and Technology AI RMF, and it provisions licenses automatically.

“It's a one-stop shop,” Marczewski said. “So you don’t have to go to this address and bookmark this and bookmark that.”

Same Standards, Every Time

Missouri’s biggest gains so far have come from review processes: work that passes through several people and several steps, and usually generates a lot of paperwork.

“Now you can do prechecks,” Marczewski said. “You already have standards built. People have access to a common agent that grades it the same, looks at it the same and is able to output the information close to the same.”

The technical review for new software used to take six to eight months. Missouri now has it down to about 30 days.

“I know if you were in Silicon Valley, that would feel like an eternity,” Marczewski said. “But it's pretty quick, especially having to hit all the different things we have to do for an enterprise.”

READ MORE: Local AI agents are the next frontier for state and local government.

In another case, the state faced 3,000 handwritten grant applications. Staff expected the review to take months of full-time work by several people. But the AI tool read the handwriting “better than I could, better than a lot of people could,” Marczewski said. It flagged applications that were missing basics, such as an address or a requested dollar amount, so reviewers could focus on the grants that were already properly filled out and quickly send back the ones that needed minor corrections. This saved time for people on both sides of the grant submission-and-review process.

That consistency is the point of the platform, Hein said. Consumer AI tools can produce different results every time. Google instead benchmarks each skill against expected inputs and outputs, and refines it until it meets the bar its launch customers set.

“We still believe that there's going to be human review time. We don't want people to just blindly trust what Gemini puts out there,” Hein said. “But we believe that the amount of review time should be drastically minimized.”

Capturing Institutional Knowledge Before Employees Retire

The standardization matters most for a less visible problem. According to Marczewski, about half of Missouri's state workforce is currently eligible for retirement.

Missouri is using AI agents to record how veteran employees do their jobs and why. His advice to them is simple: Press the microphone button and talk.

“You have someone who's been here for 30 years and knows the process backwards and forwards,” he said. “Just sit down and explain it to an agent. So now it's got a little bit of that background about why we do it.” That helps bring along junior employees who are still learning the ins and outs of state government.

The same approach applies to legacy code. Some state systems date to the 1990s, and some are written in COBOL by people who retired long ago, Marczewski said. AI helps the state understand those systems as it modernizes.

Scaling AI Internally Before Rolling Out to Citizens

So far, most of Missouri's AI work is behind the scenes. That's on purpose. Taking paperwork off specialists’ plates gives them more time with residents, especially in social services, Marczewski said. It also gives the state’s developers time to learn how to secure the tools: sandboxing agents, labeling data and locking down Model Context Protocol servers “instead of just rushing to throw stuff out there.”

Resident-facing work is coming. The state's chatbot began by answering frequently asked questions, such as pointing anglers to the Department of Conservation rather than the Department of Natural Resources for a fishing license. Its next version, “Ask Mo,” is meant to fill out forms for residents and reach them by text message or other channels they already use.

Hein sees a similar path for plain-language notices. A Gemini skill can rewrite a regulation at a sixth-grade reading level while staying accurate to the original, then translate it into the languages a community speaks.

Many AI Questions Remain Open

Missouri has also set up an AI legal task force and an AI steering committee, because many legal questions about AI remain unsettled, Marczewski said. Some state laws date to the early 1900s, and nobody yet knows exactly how they apply to prompts.

Records retention is one example.

“Are you going to retain every prompt? How long do you retain it for? What is the schedule for that?” he asked. “It's a lot of storage, especially with images and movies and things that people are creating.”

Marczewski’s team is committed to answering those questions and exploring what AI can do for state government. They’re hoping it can help both state employees as well as Missouri citizens.

“It's been a learning experience for everyone,” he said.