Missouri’s AI leader focused staff training and early successes for state innovation

Missouri lawmakers are cranking up efforts to regulate AI in state government this year as the technology continues to drive innovation, efficiency and service delivery in the public sector. One state tech official says state agencies have also been exploring how artificial intelligence can enhance government operations and business processes. 

Gov. Mike Kehoe signed an executive order in January that directs state leaders to “develop a strategic framework for the safe and effective integration of artificial intelligence within state government operations.”

According to the order, the AI framework must prioritize expediting government efficiency and the quality of government services, data privacy and security, human oversight and intervention, among other considerations. 

A group of Missouri lawmakers announced in August that the Missouri House Future Caucus will introduce legislation to establish AI guardrails and protections for constituents. The group also released a framework of considerations to assist Missouri local governments as they engage with potential AI infrastructure development in their communities. 

Helping drive and support the state’s AI approach is Tim Marczewski, director of AI and innovation for Missouri. Marczewski sat down in a virtual interview with Route Fifty during the Google Public Sector Summit this week in Washington, D.C., to discuss how state agencies are building a foundational understanding of AI now to leverage the technology moving forward.  

This interview has been edited for brevity and clarity.

ROUTE FIFTY: In your role as the state’s director of AI and innovation, how have you been leveraging AI tools in your work currently? 

TIM MARCZEWSKI: I started this role about a year ago, like last June, so a little over a year. What we did is start with some pilots to determine where does AI even fit? Does it work with us? So with that, we did kind of a runoff with a few different things like Gemini Enterprise and Copilot to learn what else is out there.

We did small pilots to be able for people to get comfortable with AI, and that also helps us build our governance. So a lot of the stuff that we've been doing is building that foundation that we can have like a playground so that staff users can experiment with AI without the worry that they're going to get in trouble for putting this data in or that data in. We have guardrails because you don't want to do something wrong. You hear all these horror stories about AI and of it going awry, so a lot of it's been building the foundations with it.

ROUTE FIFTY: You mentioned doing early work — pilots, playgrounds, etc. — and dedicating time to figure out what tools are available and how they can be used. Could you provide some examples of what those early testing or piloting efforts look like?

MARCZEWSKI: Some of the early pilots involved starting with just 100 people across 17 different state agencies and showing them AI tools, whether it was Gemini or giving people access to Vertex AI. This was also guided with trainings from our partners with lunch-and-learns. We found out lunch-and-learns work really well because people could sit at their computers, click through AI tools and just learn what the layouts are. 

But to come to us to be part of the test, staff had to bring us a use case. So then we got hundreds of use cases from different agencies across all of state government as they proposed what they thought AI could be useful for. Some of them worked, and some of them didn't. But that created a ton of feedback where staff were saying, ‘Hey, this tool is great, and it works here for these purposes.’ It’s helpful to have small pilots that you can control, because you may not have all the guardrails in place yet, but a big enough group to get the feedback. 

ROUTE FIFTY:  In the last year or so, or since your state has been really exploring AI tools, I'd love to hear any success stories or anecdotes that really stick out to you in the progress you’ve made. 

MARCZEWSKI: An early success was with our Beef and Dairy Producer Grants — I know it sounds very fascinating. But the Missouri Agricultural and Small Business Development Authority put these grants out, and they usually get like 100 applicants. This time, they got 3,000. For three people to review all these dairy grants all at once and by hand, we were able to help automate that.

We used Vertex AI from Google to intake those application documents, which could read peoples’ handwriting better than we could, while also having a human review the process. They thought it was going to take from January to June to process applications, and we cut it down to one month. All that time where people were able to not be looking at individual grants, there's a dollar figure to it. We saved about $200,000, and MASBDA was able to get that money out there to farmers in time. That was one of the early success stories that we had, and it helped us really build an understanding of what it means to check AI at each step, have the humans look at it to make sure that it is working and how they could actually work together. 

Route Fifty: On the flipside, what have been some challenges or setbacks you’ve run into when it comes to adopting and implementing AI in state government that you didn’t expect? 

MARCZEWSKI: One big one right off that bat is that trainings with 700 people in them don’t work. Large trainings work well if you have 50 or 100 people at the same tech level and kind of the same job descriptions. Especially with enterprise tools, you have got to have smaller, targeted trainings around functional areas and business practices. You could even get one-on-one trainings at some point because you really have to tailor trainings to where people are because some people pick it up and they're like, ‘I'm done. I got five agents running already.’ And then other people are still looking at AI and they're like, ‘Okay, how does this fit into my work?’ You also can’t just send automated trainings to staff. You really have to show them what does AI do, and don’t force it on staff. Take your time with AI training and enablement — you cannot do it all at once. 

Having people get too excited to a point they may think AI can solve anything is also a challenge. There has to be an intake process when considering using AI. We worked a lot with our operational excellence team to look at process improvements and business processes for AI. 

Route Fifty: How are you measuring the state’s progress and impact of AI usage, particularly when it comes to metrics beyond the number of dollars or hours saved? 

MARCZEWSKI: With government, it's not always you know you save this much money, it's also how much did AI improve efficiency. So we collect feedback with surveys that we send out. We have workforce surveys that also collect free text so staff can just say whatever they want about a tool. We’re also using AI to gauge staff sentiment in free text responses. For example, quantitative metrics may say that efficiency is increasing, but staff feedback reveals more about whether AI feels like a burden or good to use in their work. 

Route Fifty: Based on your previous and current work, and what you've learned, where is your AI mission headed? What kind of projects do you have in mind for 2027? 

MARCZEWSKI: We've been playing a lot around with single- or double-use AI agents, but we want to get down to what that really looks like in practice. We’re thinking about how AI can improve our technical review process, which is quite cumbersome. It’s done by multiple committees across 17 agencies that all have different missions. When you’re trying to find a technical solution that works for all of them, they’ll want to submit a review, and that process used to take anywhere from six to 18 months for approval and takes up enormous amounts of staff time just looking through hundreds of pages on a new product. Creating an AI agent graph can help speed up that process. That’ll also help us because when we’re building enterprise systems, there are a lot of integrations that need to happen that individuals might not catch. I think this is going to be super cool streamlining on the backend. 

Using these agent graphs could apply to a bunch of other things, too. Maybe we can move it to something on the front end where you have a better customer experience or a better automated experience with government services. So if we can work out those kinks on the back end and also train up our workforce on how to use those AI tools, how to make them and everything, I think that would be great.