How Vermont is leveraging AI across government, one ‘power tool’ at a time

After years of developing a statewide ethical framework for artificial intelligence, Vermont has entered what its chief data and AI officer calls an enablement phase, in which state employees are using and building their own AI “power tools” to automate routine work and free themselves from “drudgery.”

Vermont was an early bird in developing a state-level approach to AI as the first state to establish a task force on AI in 2018. In 2020, the body recommended the development of a statewide AI code of ethics in a report that it published. The report said guardrails should be focused on human oversight, transparency and individual rights, but it also cautioned against regulating the underlying technology rather than its applications.

In September 2022, Vermont named Josiah Raiche the state’s first director of AI, and then in 2024, the state merged its AI and data teams under Raiche, elevating him to the chief data and AI officer role. This further integrated AI governance with the state’s broader data strategy, Raiche said.  

Now, several years later, the fruits of that work are being realized, Raiche told StateScoop. The state has begun training developers to use AI tools to identify and harden vulnerabilities across its roughly 1,100 applications. Raiche said the process started with cybersecurity, and the AI tools can be used to scan source-code segments, model attack paths, generate Python checks, identify vulnerabilities and then support remediation.

So far, Raiche said about 100 developers have gone through the training, with another 50 expected to participate. He hoped the training would free up some of the developers’ time doing these processes manually, allowing them to harden their systems. He said it was important as AI gives both government and potential attackers new capabilities along with other imminent threats, such as post-quantum cryptography. 

“Long term, the goal was that the devs would be able to take the outputs of those of those findings and deploy them, and that there would be enough [outputs] that would incentivize [the developers] to adopt a bunch of the other best practices, like continuous integration and deployment and other modern dev practices in general,” Raiche told StateScoop, adding that some of the applications were decades-old. 

Raiche said the effort has already resulted in patches to several dozen applications, with developers using AI now to group similar vulnerabilities and generate fixes, while testing instructions and steps to verify that the changes didn’t cause any disruptions on the frontend. He added that he hoped the approach would lead to improving developers’ ability to maintain and eventually build their own applications. 

This is what he said makes AI a “power tool” for not just developers, but all state employees. And, it makes the tools “a lot less scary.” 

“The approach we’re taking in Vermont overall is this idea that AI should be used as a power tool, which is an inherently kind of limiting opinion about how to use AI,” he said. “A lot of people … are talking a lot about agentic workflows and like kind of autonomous agents doing stuff. And in a power tool mindset, you actually can’t do that, right? Because a power tool is operated by a human, there’s a human operator — it’s in the context of a process that we’ve created. The operator is well trained, and then a good power tool lets you zoom out from mechanics and think about outcomes.”

It’s not just the IT shop that’s building its own power tools. Vermont is also using a “pathfinder” model to encourage AI adoption within state agencies, he said. Each agency is asked to identify employees who are trusted by both their peers and leadership — the people he said are often not “early technology adopters” — and they are given early access to training on building the AI-powered tools for their teams.

Those employees get roughly a three-month head start to develop practical applications before the state trains their broader teams, he said. This approach allows agencies to identify use cases from the ground-up, rather than having leadership dictate how AI should be used, he pointed out. 

Raiche said the use cases emerging from the program are “boring,” but they’re the ones that have high returns and make employees’ work more efficient, freeing up more time to serve constituents.

“I call it drudgery, right? Just do the drudgery. Let people work on creative, engaging stuff because people don’t become public servants to do drudgery,” he said. “They come because they’re mission oriented and they want to help.”

Recently, Raiche said selected officials participated in a half-day workshop using ChatGPT-style tools, an internal chatbot, prompt frameworks and custom GPT skills.The participants researched policy approaches across all 50 states and built tools in cross-agency pairs.

“I got to watch the commissioner of Corrections helping someone from the Department of Financial Regulation, build a tool that would evaluate new securities that were going to be offered in Vermont, and like the two of them having a pair session, building this,” he said. 

“What was a surprise to me was that we had 50 really busy executives in a room, all having a good day. Like, how often does that group of people all have a good day?” he added. 

Raiche also described a recent moment when the Secretary of State’s office reached out to his office for help with ballot proofing under a tight deadline. The SOS team had hundreds of ballots to check against a spreadsheet of roughly 3,000 candidates, and the second round of proofs still contained errors on a Friday when the ballot was set to print the following Tuesday. Raiche joined the elections director and staff to craft an AI “skill” to speed up ballot proofing under the tight deadline.

The tool checked the proof files and flagged errors it was confident about, along with items it wasn’t sure about. Instead of two full rounds of manual review, staff used the AI’s findings as a first pass, then reviewed every ballot once themselves — confirming the flags and checking for anything the AI had missed. The team finished the task in just three hours with the skill, Raiche said, avoiding the overtime they feared would be needed before the ballots went to the printer.

“I got there at 9 in the morning, and we got it done by noon,” he continued. “And they were just like, ‘This is amazing!’ Like they were excited, so this was an example of the power tool approach again.”