---
id: "workshop:navigating-tech-governance-in-2025-policy-procurement-and-ai"
type: "InnovateUS workshop"
title: "Navigating Tech Governance in 2025: Policy, Procurement & AI"
url: "https://innovate-us.org/navigating-tech-governance-in-2025-policy-procurement-and-ai"
date: "2025-04-30"
duration: "60 minutes"
language: "English"
instructors: ["Cassandra Madison","Josiah Raiche"]
recording: "https://www.youtube.com/watch?v=q0pxALf3AKU"
transcript: "automatic, from the recording"
source: "InnovateUS (innovate-us.org)"
licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org"
---

# Navigating Tech Governance in 2025: Policy, Procurement & AI

InnovateUS workshop · 30 April 2025 · 60 minutes

Technology governance is no longer a niche topic—it’s central to how governments deliver services, manage risks, and earn public trust. In this timely and practical InnovateUS workshop, Cassandra Madison and Josiah Raiche will guide public servants through the five most important shifts shaping technology governance in 2025. Drawing on real-world examples and policy insights, the session will unpack how public sector organizations are rethinking procurement, AI oversight, data infrastructure, and cross-agency coordination.

Designed as a practical map for public service professionals, this workshop offers clear direction through a fast-evolving landscape—helping participants make better decisions, communicate more effectively across teams, and lead with clarity in the face of technological change.

Whether you’re setting strategy, managing digital tools, or navigating emerging technologies on the ground, this session will give you the context and confidence to move forward.

## Learning goals

- Identify five emerging trends that are reshaping technology governance in the public sector in 2025.

- Understand how these trends impact areas such as policy development, procurement practices, AI implementation, and digital infrastructure.

- Engage in cross-functional conversations about responsible tech adoption and interagency collaboration.

## Speakers

- **Cassandra Madison**, Executive Director, Center for Civic Futures
- **Josiah Raiche**, Chief Data and AI Officer, State of Vermont

Recording: https://www.youtube.com/watch?v=q0pxALf3AKU

## Transcript

Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/navigating-tech-governance-in-2025-policy-procurement-and-ai.txt (plain text) and https://innovate-us.org/navigating-tech-governance-in-2025-policy-procurement-and-ai.json (every segment with its start and end second).

**Cassandra Madison**

[0:00](https://youtu.be/q0pxALf3AKU?t=0) You get to, which feels really special, Josiah and I are both from Vermont, although neither of us happen to be in Vermont today. We're both in Pennsylvania. But I'm really excited to be here with you. Let me just get my screen shared here. And I will get the presentation going. Give me one second. All right, thumbs up if you can see my screen. Excellent. Okay, so Josiah and I are really excited today to talk to you a little bit about the trends we see emerging on AI over the next year, and impacts for the public sector. So we'll introduce ourselves in a little bit more detail in a minute. But first, I just want to give you a little bit of background on Center for Public Sector AI. So we're a nonpartisan nonprofit organization that's really focused on helping state leaders make good decisions on emerging tech. We do that through education, convening, tools, best practices, things like that. And there's two main constituencies that we've been supporting for the past year. One is state health and human services leaders. So state leaders who are thinking about how might we apply AI to deal with snap backlogs, or Medicaid backlogs, or backlogs in child welfare, we are helping them understand what tools they're being presented with and what a good decision looks like. The other group that we support is state AI advisors. And so that's where Josiah and I comes in. We convene folks who kind of wear the hat of leading on AI in their state. We convene a community of practice of that group that meets twice a month and has some in-person convenings with the purpose of sharing best practices, and trying to break down some of the silos that come when you're working on your own on something new in a state. So when I just

[2:01](https://youtu.be/q0pxALf3AKU?t=121) boil down CPS AI into four key components, we educate leaders, we help states mitigate risk, we help states build capacity, and we accelerate learning. And a lot of that we do in partnership with Innovate US. So thank you, Agata and team for having us here today. So before we dive in, I thought it would be helpful for Josiah and I to just share a little bit about our stories with you, and some of our lessons learned from the work that we've both done in Vermont, because I think they'll help frame how we think about AI more broadly. So I'll go first, and then I'll hand it to you, Josiah. So for me, I like to say I started my career as a bright-eyed, bushy-tailed policy person that really believed if you just passed the perfect law or wrote the perfect policy, things would magically get better for people. And I ended up in state government when Vermont was going to be the first state in the country to do single-payer health care, and the health insurance exchange technology was going to be a path to doing that. And much like healthcare.gov, our system imploded on itself. And it really changed how I see the world. So I just ended up scrubbing in where I was needed. It turns out I like the operations and the tech, I think even more than the policy, but I tend to sit at the intersection of the three. But it was during that experience that I learned a whole lot about not just what, unfortunately, not enough about what good tech delivery looked like, but what it looks like when things go wrong. And so I've made this whole second part of my career focusing on helping government not make the same mistakes that we made at that time. Josiah, you want to share a little bit about your story?

**Josiah Raiche**

[3:32](https://youtu.be/q0pxALf3AKU?t=212) Sure, yeah. So I joined the state kind of in the immediate aftermath of, of our health exchange. And I worked with our agency of transportation, they hired me as like, the most junior developer that you could ever be. And my job initially was to copy and paste information from one website to another. And I automated that in my first couple weeks, and then they didn't have any work for me to do. So. So they said, Hey, can you go learn how to do data analytics and data visualization, and whatever this new thing, artificial intelligence is. And that was, that was how I came to AI. So one of those early projects was a data visualization project that a bunch of it people had spent a bunch of time working on and built a set of dashboards and reports that were really useful to exactly one person in middle management in the agency of transportation. And learning from that, like, how do we build data driven tools that give people insights that let decision makers and policymakers really drill into the things that matter, and give them access to information to help them make better decisions upfront. So that's kind of what I've spent a lot of my time with the state doing in various ways.

[4:57](https://youtu.be/q0pxALf3AKU?t=297) And now I'm the chief data officer and chief data and AI officer, focusing on how we can do AI well within state government, education for Vermont's workforce broadly, as well as all the usual chief data officer, data quality, data governance, analytics, those types of things.

**Cassandra Madison**

[5:15](https://youtu.be/q0pxALf3AKU?t=315) Thanks, Josiah. And there's a couple things I love about that story. One is I know a lot of folks in the audience are thinking about how they can upskill and build a path into government and then within leadership and government and Josiah, I think is a great example of someone who came in and like, got his hands dirty and is now like understands how the mechanics work and is now in this really important leadership position in the state. The other thing I think is when we think about and we'll talk a little bit about data later, but AI and data are so intimately connected here. And so having someone join us today that is wearing both hats, I think is really powerful and will help with this discussion. So thank you for that, Josiah. Okay, so we wanted to share a little bit about our lessons learned. And the reason why we share before we get into the five trends and lessons learned, is that the best predictor of the future is the past. And historically, government has struggled to deliver technology and those challenges as fundamental challenges don't go away, just because we're talking about AI. And there's a lot of aspects of AI that are new. There's also a lot of things that are just the same around best practices for how you deploy technology well, but there's a little bit more risk here. So whenever you know, we're talking about the trends in policy and procurement, and governance, we want to make sure that we're also talking about these fundamental good practices and some of these lessons learned. So Josiah, do you want to share a little bit about this

**Josiah Raiche**

[6:47](https://youtu.be/q0pxALf3AKU?t=407) this first lesson? Yeah, sure. So in this moment where, you know, everybody is doing AI, and everybody has a splashy AI use case, I think it's a great opportunity to kind of reground ourselves in, in why we do tech, and really, we do tech in order to achieve outcomes. And while it's useful to do pilot projects, in order that our workforce can gain kind of skills and working with new tech, it's still not the end, right, the the AI itself is not is not the main goal. So the way we like to think about AI in Vermont is that AI is really a power tool for our information workers. And in the same way that power tools revolutionized our, our manual labor, again, my my first role in state government was with the Agency of Transportation. And when I started working there, I talked to some of the old timers who, who were there when they went from hand sides to weed whackers. And just the transition that that was and how that that use of a power tool really changed the way that they do the work that that work is still, you know, getting rid of clearing brush on the roadsides still very manual, and very labor intensive, but that kind of transition from doing it by hand with with simple tools to using power tools. And now, largely that work is is being done, you know, through tractors and mowers. Those were like pretty significant transformations for them. And so we I like to think about AI being a power tool for our information workers in that a power tool is used by a person to accomplish a particular task. So if you think about using a cordless drill versus using a table saw, right, those are very different, very different tools used for very different purposes. So power tools are used for particular tasks,

[8:47](https://youtu.be/q0pxALf3AKU?t=527) they're used by skilled operators, different types take different levels of skill. But you know, if you're giving somebody a power tool, there is an assumption that the person has a sense of how to use it. And then finally, it's in the context of some sort of human process. So if I'm doing trim on a door, if I'm doing that with my hammer, I'm thinking about the mechanics of because I'm not that handy, I'm thinking about the mechanics about how I'm swinging that thing about not hitting my thumb, about making sure I'm not denting my trim. When I switched to using something like a staple gun or a nail gun, I can think a little bit less about the mechanics of doing the task, the kind of drudgery of doing the task. And a good power tool helps me to zoom out and think about the outcome that I'm trying to achieve. And so I can be thinking about is that corner coming in tight? Is this is the reveal even all the way around things like that, that I just I don't have the cognitive capacity for when I'm doing this manually, when we switch to power tool, it kind of frees up cognitive load in a well designed power tool to help you focus more on the outcome. So those are all paradigms that I think are relevant to how we think about AI. And when we think about AI as a power tool used by a human operator for a particular task, it keeps us from stepping in all kinds of potential pitfalls in AI. And we'll get to some of those in a little bit. But back over to you, Cass.

**Cassandra Madison**

[10:14](https://youtu.be/q0pxALf3AKU?t=614) Thanks. I think this is especially as we get to later in the presentation, talk about things that are coming like autonomous agents, thinking about this power tool analogy, I think is really important grounding. Okay, second piece here is understanding the problem. I think there's lots of saw some great design and product technologists on the call. So this is a not a foreign concept. But sometimes in government, it is to say, look, AI exciting, it can do so many things. But what problem are you trying to solve? And do you really understand the dimensions of that problem. So falling in love with the problem first is really important and a critical best practice. Because something simpler may then AI might solve your problem. And the only way you're going to know that is if you can understand your problem and best practice would be understanding that both qualitatively and quantitatively from the perspective of your users, whether they are internal staff or the constituents you serve.

**Josiah Raiche**

[11:07](https://youtu.be/q0pxALf3AKU?t=667) One of my early projects in AI at the Agency of Transportation was we were trying to predict asset degradation over time. And we had a slack channel that we were using. And toward the end of the project, our slack channel became named how we're not using AI to do asset degradation. Because we found that we were able to use AI and get as good as the current state of the art. But we weren't any better than the current state of the art, right. And so this idea of how do we start small? How do we iterate? How do we fail fast? I think those are those are crucial. And for many of us, we've kind of been in the pilot phase of AI for a while where, yeah, we do focus on failing fast. But as we as we start to scale, we'll get to this a little bit later, like this idea of iterating and growing existing solutions, starting with a small one and continuing to grow is really important. The other guidance I like to give my team is think about what the biggest kind of mistake you can make, right? Like, can you make a mistake on something that's, you know, $10,000 or $50,000, depending on your level in the organization, that really should drive your iteration size, right? Like, what's your what's your organization's risk tolerance? What's your organization's tolerance for things to not go well, when you're in kind of this emerging tech space? You know, I talked to my boss about how near 50% of our pilots haven't worked out the way that we thought they were. Either, either we didn't get any value out of them, or the value that we got out of them was not the value we initially expected. And we need to kind of normalize that in this emerging tech space, we need to get better at making good bets. That's another thing we need to do. But making sure that we can we can learn from the the, the things that

[13:07](https://youtu.be/q0pxALf3AKU?t=787) don't go as well as we intend where where our theories were not validated, and grow.

**Cassandra Madison**

[13:14](https://youtu.be/q0pxALf3AKU?t=794) And then this, I think, is a goes nicely with the things we've already talked about, which is you need to be able to measure these things. And so both understanding the dimensions of your problem, and how, you know, if you have a backlog, how big is it? And what are the causes of those backlogs? So that quantitative piece, and from the outset, understanding like, what does it look like when we've solved this problem? And can we measure it is critical to understanding whether one you want to take, you know, apply an AI application? And is it better than what a human can do? Isn't it better than meaningfully better than what a more simpler system can do? And then you can understand the return on investment and your risk calculation associated with that. So you want to make sure you're doing that first.

**Josiah Raiche**

[13:55](https://youtu.be/q0pxALf3AKU?t=835) And then finally, thinking about the change management, power tools are used by skilled operators, you need to make sure you're you're training your operators to use the tools effectively. And also that as you're evolving your processes to take advantage of AI, because there are a lot of opportunities to do that. We are doing that in ways that remains human centered. And we're, we're, we're bringing the humans along, not merely making the humans kind of the extension of the machine, or the the, you know, living people who execute the orders of the robots. Those are both anti patterns we want to avoid, we want to make sure that we're staying human centered, and that we're doing change management well, and that our staff are supported all throughout the transition, which hopefully will be making work better for them all along the way.

**Cassandra Madison**

[14:48](https://youtu.be/q0pxALf3AKU?t=888) And then the last thing, and I think Josiah already spoke to this is just safe spaces for experimentation and psychological safety, government. It's not a place that is always easy for innovation, and making sure people feel comfortable with something like, hey, 50% of the things we're experimenting with aren't going to go well, we're not going to move those things to production, is okay, that's part of the point, is a very unusual paradigm in government. And so people need that, not only the mechanics of creating a safe space where it's okay to fail, but the psychological safety that goes along with that. And so we wanted to ground with those things, just because I think, as we're talking about what's coming next, the way Josiah and I both think about this is that everything needs to be grounded in those principles that we just talked about, as you think about what are the right decisions, and where are we going, and what are the implications of how fast the technology is moving. These are the principles that we always ground the decisions and frameworks and thinking in. So we're going to talk about the five things on our mind. These are not the only five things, I will say. I think as all of you in the audience are probably trying to keep up with AI and understand the advances and the implications and the executive orders and the pilots, there's a lot to keep up with, and the word I keep hearing in all the conversations we're in is dizzying.

[16:11](https://youtu.be/q0pxALf3AKU?t=971) And so what we wanted to do today was just share the things that are top of mind from our perspective, and have a good discussion around it, but know that they're not everything. I will say, as I'm sharing my screen, I can't see the chat, but please, if you want to start posting questions or thoughts in the chat, once we get to the Q&A, I'll be able to see it. And then, Josiah, if you happen to see something interesting and want to comment on it and pause on it, let me know. Okay. But Josiah, you're up on autonomous agents. Tell us about them. Yeah, you want to

**Josiah Raiche**

[16:42](https://youtu.be/q0pxALf3AKU?t=1002) go to the next slide? Yeah. So autonomous agents are the beginning of the shift from kind of power tools to AI machines, where a power tool is used by a single operator. When you shift to the concept of machine, the process kind of defines a series of power tools that are all working in tandem to accomplish something, and that's kind of the transition from tool to machine. So what we're seeing with autonomous agents is a jump from using AI to do automation to really having some autonomy there. And this is a kind of a complicated world to live in for government, because there's a reason we use form letters. There's a reason mail merge works really well for government, and that is that we want to give the same information out in the same context 100% of the time. So there are places where autonomous agents are going to be super useful for government, and kind of building smarter power tools is going to be helpful. But one thing that I'll say is that your government gets to think about how much autonomy do you want to have in your processes, in your AI tools. So if you're in a context where the majority of your residents aren't very tech savvy or they're mostly older and they really actually prefer doing work on paper, shifting to kind of very automated autonomous types of processes that are heavily AI focused is maybe not going to be a great brand for you. It's not going to be highly effective.

[18:33](https://youtu.be/q0pxALf3AKU?t=1113) It's not going to be that successful as a way of doing government. So thinking about when is it appropriate to have autonomous agents plugged into your processes is I think really crucial. Why don't you go over to the next slide. So a couple of things there that we're thinking about in Vermont as we're prepping for autonomous agents is really thinking about what's on brand for us. A question I ask in the beginning of most of our AI projects is, when was the last time you picked up the phone and called a call center and went, I'm really glad I got to talk to a robot today and I didn't have to talk to a person. That is not an experience most people have for anything but the very simplest types of questions. And so making sure that we are being human centered and that we are thinking about outcomes for constituents, kind of thinking about that in parallel with thinking about the efficiency gains or cost savings that we could achieve by using more AI I think is very important.

[19:52](https://youtu.be/q0pxALf3AKU?t=1192) Another item I like to flag for our teams is AI agents are going to be like a very eager intern. They want to get the right answer. They might cut some corners to get the right answer. You want to treat your AI agents as very, very junior members of your staff. And does your organization have good muscles for dealing with a very junior person who makes a wrong call? Because AI tools are going to hit 99, maybe 99.9% accuracy. But can you deal with that one out of every 100 or one out of every 1,000 failure rate? And how do you deal with that? Is your solution, we let the person go? If so, adding AI to the mix is not really going to work out that well. So making sure that you have a culture that has built in quality checks, that has escalation paths where people can learn from mistakes and iterate on them and the process can evolve to changing, to adapt to changing conditions. Those are all like really crucial in general, but especially when you're going to try to plug in AI. As you think about for your brand, right, where do you want an autonomous agent? Where do you want human review? Where do you really want a human touchpoint? The way that we are thinking about AI in Vermont is freeing up staff to spend more time with Vermonters rather than Vermonters main touchpoint being with a robot and then using our staff for back office tasks. We actually want to shift our staff away from back office tasks and allow them to focus more on direct support for Vermonters. So those are things that you'll want to look at. Explainability, super critical. And often there are ways to accomplish that by using

[21:52](https://youtu.be/q0pxALf3AKU?t=1312) AI kind of in very narrow key points in your process and having your process be very well defined and very explainable so that the AI step is just kind of a sorting type of a step or some sort of a categorization or a flagging process that leaves your overall process and the outcomes being very explainable but still allows some of the efficiency gains that you're looking for. And then, you know, do think about ethics. Vermont has a robust ethics policy that's been adopted by a number of states. Feel free to go grab that. There are a number of other models out there from various cities and states. But thinking about those ethical frameworks and how you're going to accomplish things early on is important so that you can avoid a reactive governance mode. You want your governance to help you to move quickly and in a streamlined way not to be kind of a roadblock you throw up at the last minute or even worse responding to bad things

**Cassandra Madison**

[22:53](https://youtu.be/q0pxALf3AKU?t=1373) that happen. Okay, the second theme is one of acceleration. And I think, look, we all know that the technology is moving really fast and every day that goes by it gets better, faster, more expansive. And over the past year, we've seen a lot of states, cities, federal government also accelerate what it's doing. But I think in the past, you know, transition to new administration also, the federal posture on this has changed a bit to be one that is really promoting acceleration. And so there's going to be continued impacts of that. So technology moving fast and then lots of encouragement for adoption starting from the federal government all the way down. So there's actually been like a fair amount of guidance that's come out over the past few weeks, if you haven't seen it, that includes requirements around responsible adoption, around procurement standards, around testing and oversight and monitoring, all those things are really strong. But it's within that frame of move faster, and try to do everything you can to take away unnecessary bureaucracy and push for that adoption. So the feds are really focused on that adoption, acceleration and decentralized oversight. And what this means for states is that there's a lot of discretion, and there's a lot of responsibility. And, you know, we could end up with 50 different approaches to AI, and there's some risk to that. When you have lots of different experiments running, it can leave lots of room for creativity, and shared learning. But if there's not connective tissue between those things, it can also result in fragmentation, duplication and trust gaps. And so there's, we see this moment as a real opportunity for states and cities to lead with clarity and with courage and to really work with each other in ways that maybe they haven't up until now.

[24:57](https://youtu.be/q0pxALf3AKU?t=1497) And so with when thinking about responsible adoption, we really see this as an opportunity to not necessarily wait for, I mean, take what model guidance you can from what the feds are putting out, but don't wait for it. And really build your own North Star for ethical, effective AI in government. And to promote coordination across states. This is one thing that CPSAI is really promoting, but there's lots of other organizations who are doing that. And so to the extent that it's possible to have governments be sharing frameworks and assessment and guardrails, I think it's really helpful. And so the more also the governments are doing these things out in the open, the better. And Josiah, your example of, hey, come take our framework over here is a really good one. These things are accessible. Investing in shared infrastructure. I think this is a little bit nascent still across states and cities, but what we hear out there is a lot of interest in doing this. And so I think this is relatively unexplored, but hopefully we'll see more of that over the next year. I think it varies widely, the kind of talent and resources and risk tolerance that different governments have. And so I think there's an opportunity for states and cities that are a little bit more out front to share their learning with others who are moving a little bit more slowly. And there's an opportunity for clusters of states who are working on similar things or cities who are working on similar things to get together and support each other. Josiah, you can talk about pilots.

**Josiah Raiche**

[26:30](https://youtu.be/q0pxALf3AKU?t=1590) Yeah. So many of us have been kind of in pilot mode for a year or two, depending on how early you jumped on the generator AI bandwagon. And yeah, thank you. So there are a number of challenges we've run into in the past when we move things from pilot to doing them at scale. And a lot of those are applicable lessons learned for AI. So things that I would encourage folks to think about as you go from kind of your one or two production use cases to 10 to 20 production use cases to eventually, I'd like to see AI become ubiquitous across our processes. Things we need to think about are when we do pilots, we build these kind of ideal environments to do the pilot in. They're low risk. They're relatively high reward. If they go down for an afternoon, no one cares because we've got fallbacks in place. It's just a pilot. Even with our first production use case, it's usually okay if it goes offline for a couple hours. Some of those kind of ideal environment things are just not feasible once we get to-

**Audience**

[27:45](https://youtu.be/q0pxALf3AKU?t=1665) Are we ready? All right. Ladies, let's call this to order. We'll call to order the work session for the Cochise County Board of Supervisors for Wednesday,

**Josiah Raiche**

[27:52](https://youtu.be/q0pxALf3AKU?t=1672) April 30th. They're muted. Go for it, Josiah. Okay. Thank you. All right. I think I'm unmuted now. Thank you. So anyway, moving from those ideal environments to actual real life has a set of challenges. So think about data. Think about how does your AI tool deal with unexpected data that flows into it? We've got a lot of experience in the data. Those of us who work with data, people writing, I'm John Smith III, but they use three pipes instead of capitalized, for example. Those kinds of scenarios. How does your AI handle those unexpected inputs? How does your AI deal with very large or very small household sizes? Things like that. Thinking about the edges of your data being much more robust in your integration testing and making sure that you have monitoring upstream and downstream of wherever your AI step in the process is. Thinking about how once you identify places where the AI tool does not perform that well, how do you divert and go around it, either using a manual process or flag for review? We know that it gets, I don't know, household sizes greater than six. The AI tool tends to have issues with those about a quarter of the time. So how do you flag those for additional review? Things like that. And then making sure that you have staff to support these things. We tend to put very heavy emphasis on making sure we have a lot of technologists or change management folks when we're doing a pilot. Can you scale that? Can you scale that to doing real life? So,

[29:54](https://youtu.be/q0pxALf3AKU?t=1794) yeah, we put a number of things on there. I've already talked about auditing your data. But think about, you now have this new set of tech that has both new capabilities and constraints. And in the same way that when we transitioned from, oh I don't know, doing mainframe technology systems to web apps, right? That brought a different set of constraints and capabilities and we redesigned our processes to take advantage of those. We're going to do that again with AI. So we've now got AI tools that are good at some things, not so good at other things. Let's design our processes to take advantage of that and really think about that process engineering. Failing fast is super critical and I talked about that at the beginning, but just like want to push on that a little bit more here as we think about scaling. Every time you do an expansion of an existing AI tool, you need to review your risk assessment because your risk profile can change quite a bit as soon as you add maybe a user base that isn't as highly trained or you're handling a slightly higher risk population because you've expanded from your initial kind of easy cohort to a more complex cohort of folks that you're serving. All those things can change your risk profile pretty dramatically and it's worth going back and checking that risk assessment every time you expand a pilot. And then thinking about the change management piece and remaining human-centered, stay grounded in the outcomes. Cass talked about that at the beginning. Really stay grounded in what are you trying to accomplish and support the people who own the process. For us in Vermont, that's usually a director or a deputy director is the person who's responsible for the overall

[31:56](https://youtu.be/q0pxALf3AKU?t=1916) process outcomes. So staying well aligned with them, with their teams, with their staff, and then with the constituents that they're serving is really important. And I think one other piece that is very easy for us technologists to overlook is normal computer programs when they fail, throw a blue screen of death or some sort of an error box or something. There's a red something that appears that says, hey, something's going wrong. AI doesn't do that. Instead, AI gives you really good looking content that is nonsensical or incorrect or not useful. And those failure modes, we need to train our teams to be able to know and to look for. So one of the things we've done in Vermont is tell all of our employees, go ahead and use generative AI tools. Talk to your supervisor. And we put out a code of ethics and our guidelines, but we said, go use them. And the reason for that wasn't so much for the productivity gains. It was more for we really want staff to know how these tools fail and what types of failure modes and just build some literacy with the AI tools. And then finally, just share, share to the community.

[33:16](https://youtu.be/q0pxALf3AKU?t=1996) CPSAI has a great community where we get to share pretty often, but take advantage of conferences, take advantage of forums that are online and share what you're learning and what's working well and

**Cassandra Madison**

[33:28](https://youtu.be/q0pxALf3AKU?t=2008) what isn't. I love that point, that last point you made, Josiah, that I think as I think about some of the ways that we got trapped on early government tech delivery that might change with AI was this idea of governments used to doing these, what we call big bang technology projects where you sign some mega contract and you write all of these requirements down in the spreadsheet, but it's all conceptual and theoretical and you hand it over to a vendor and then you get to test it and see it right before production happens. And it's often in that limited testing or in production that you find out what's gone wrong, both from a practical perspective and a usability perspective. And I think this concept of getting AI tools in the hands of day-to-day staff so that they deeply understand how it works and what it can be used for is a really powerful way to then hopefully change the paradigm and how we're used to tech delivery in government as a whole, where fingers are on keyboards the whole time and it's not something that everybody's holding until the end right before you go from zero to like a hundred. You can do this incrementally and get more people involved, so that feels really exciting. I actually can see the chat now, so I just want to ask Josiah as we're talking about scaling, is Vermont thinking at all about energy use in AI or is that something that's just, I do think we didn't include that in this presentation, but I think it's a great point and something that we talked a lot more about over the

**Josiah Raiche**

[35:00](https://youtu.be/q0pxALf3AKU?t=2100) next year plus. Yeah, it's something that we're talking about because like both from, you know, Vermont's known for being a green place, as you can see from my background, and also from like we have a grid and the grid didn't grow a whole lot when we started turning on AI workloads, right? So do we have capacity to handle some of this stuff? And it's currently pretty opaque what energy usage looks like in the context of AI, so I've done a few talks and articles on it, but that's something where we need to continue pushing the industry to build more transparency on. The differences between like the more efficient models and the less efficient models to do the same task can be an order of magnitude or more in energy usage and that's crazy. So Hugging Face has released some benchmarking that they did. There were a couple of recent studies, I think out of MIT maybe, in this space. It's an area of ongoing research, but I think continuing to use our kind of collective power on the procurement side to get visibility into energy usage will be

**Cassandra Madison**

[36:18](https://youtu.be/q0pxALf3AKU?t=2178) important. Awesome, thanks Josiah. Okay, next theme is on workforce transformation. I know this is probably top of mind for a lot of folks. So I think it's clear that AI is starting to shift tasks in the private sector. You're starting to hear it more with coding and automation, and the changes will inevitably ripple into public work as well. So the things that are on our mind as we think about workforce transformation, on the government side, I think a lot of agencies, most agencies are operating in backlog status already and don't have enough humans to do the work that needs to be done. And I mean, my lens, I have a human services background, so that's the lens through which I see a lot of government activity. But there's always more tasks than people can accomplish. And so there's a real opportunity here for AI to get rid of some of this back office drudgery and improve the customer experience, make things more timely, and so that staff can, to Josiah's point, focus on the people and not the paperwork. But this also means that people's day-to-day work and the tools that they use are going to change, and it's important to recognize that, understand that, and try to help people navigate that change. The other thing we've been thinking about is just the economic uncertainty that we're in right now. AI aside, if we start to see more private sector layoffs and then you add AI on top of that, there could be increasing demand for government services over the next couple of years. And so at a moment when there's some federal cuts and some states are having to lay off and government's having to do more with less, increasing need makes that additionally difficult. And if you look at some government programs like Medicaid work requirements, for example, which will actually add additional administrative tasks to folks who are doing

[38:18](https://youtu.be/q0pxALf3AKU?t=2298) frontline work, you kind of have a perfect storm of operational pressure. And so there's both the how do we support government staff as their roles continue to change and evolve and leverage AI and get through some of the fear, and how do we leverage AI to serve people better? And I think even if within a government context there is reassurance going on that folks are not going to lose their jobs and everything's going to be okay, just it's important to acknowledge that that's not necessarily the chatter that's happening out in the world. And so there's not a magical division between how you experience your work life and all of the things you hear outside of your work life. And so it's likely that people are going to be bringing that generalized fear, anxiety questions into their day-to-day work as well. So what can states do about this? I think one is just acknowledging and owning that shift in that dialogue, open communication, being clear with people that we don't have all the answers, but we're going to figure this out together. And this is how we're going to try to focus on enabling and making the humans day-to-day better, I think is a really powerful piece.

[39:37](https://youtu.be/q0pxALf3AKU?t=2377) The second piece is about upskilling, and this is one of the reasons why we're so excited about Innovate US and all the work it's doing, both in these live trainings for everyone and also trainings inside government. But helping folks, creating those safe spaces for experimentation and upskilling people is really important. When it gets to redesigning roles, I think this kind of thing, again, gets tricky in government, but it's making sure to do those things thoughtfully and with the folks who are affected by those changes at the table is really important and preparing for those increased service needs. So figuring out, again, how can we focus AI's power on some of that back office work to make sure that it's possible to meet the needs. And through it all, modeling that steady leadership and knowing that showing, being transparent, listening, and showing adaptability and resilience are all critical. And I think it's in a lot of cases more critical than having the actual answers about what comes next, because we're just in a time of uncertainty and we're going to be for the foreseeable future.

[40:48](https://youtu.be/q0pxALf3AKU?t=2448) And so modeling that steady leadership, showing people that you're listening and that you're caring and that you're bringing them along is half of the battle here.

**Josiah Raiche**

[40:59](https://youtu.be/q0pxALf3AKU?t=2459) And then the last one of our five that we wanted to dive into was this idea of public trust. I think that we're in a time where, as we put on the slide there, right, public trust in government is quite low overall. And that varies pretty broadly from region to region and between organizations. But I think it's important to remember that our constituents don't necessarily draw very, very firm lines of distinction between various levels of government. And so as trust in institutions overall is low, even if you're not the one who supports the judiciary, if the judiciary in your area has lower trust, people don't see that distinction very clearly, as clearly as those of us inside do. So really thinking about trust, if we gain efficiency and lose trust, we're losing. And if we manage to save money, but we're not succeeding in delivering outcomes that people expect or need, I think we're in a tough spot. So AI can overall be used in ways that do build up people's trust and encourage people to participate more in their government and be more engaged.

[42:32](https://youtu.be/q0pxALf3AKU?t=2552) But we need to do that thoughtfully. So using trust, kind of the way I talk about this with my teams, I use, and maybe this is a silly example, but I use our local investigative journalism organization and tell people, hey, if they knew that you were using AI in this process, what kind of story would they write? And encouraging people to kind of think about that upfront. It helps people come up with a lot fewer really bad ideas when I frame things that way. But also it helps us to identify places where we would be comfortable using AI and having people poke at, why did you think this was a good approach? Why did you think solving this problem with a robot was going to make things better? So we've put a number of things on the slide here. I think staying human centered, staying transparent, those are crucial. Making sure that we're building accountability in at the very beginning and that all of our AI uses are clearly explainable and we can demonstrate that we have review processes in place and that people's experience with government is getting better overall, not worse, through our adoption of AI. And then being ready to pause, right? There are moments in our pilots that we've already run into, and I expect as we scale we'll run into more of these, where we need to say, okay, wait, we're not doing this expansion yet, or actually we're going to roll this back and go back to whatever our fallback was temporarily. Will we evaluate why this isn't working as we expected? Why people are not having the outcomes that we wanted

[44:35](https://youtu.be/q0pxALf3AKU?t=2675) them to have? Making things the same with AI as they were before isn't necessarily a win. We really want to be making things better and we want to make sure that our processes and our ability to deliver services remains agile and that we have the ability to adapt to changing conditions, new legislation, whatever the environment.

**Cassandra Madison**

[45:04](https://youtu.be/q0pxALf3AKU?t=2704) Okay, and I will, I guess I'll just say on that trust piece, so the reason why trust shows up as one of the five is I think this concept of like going from pilot to scale, I mean, there's been a couple little stories here and there of AI giving people wrong information in a government context, but I don't think like we haven't seen a failure at this scale, that it's like a front page news story across the country. And I, you know, we might see that over the next year, hopefully not. But as, as these things scale, and they become bigger news headlines, like there will be mistakes, inevitably, and so trust is going to be on people's minds. So you have this institutional trust conversation. And you also have, you know, the more we scale, and the more we accelerate, the greater the likelihood that there's going to be some bigger, more visible thing that then will compromise public trust around these systems. And so it's about making those good decisions and working incrementally, as Josiah said, and then also thinking about what is the response as those things happen if we end up boomeranging back a little bit with with public excitement and trust over this technology.

[46:19](https://youtu.be/q0pxALf3AKU?t=2779) So those were the big things that we wanted to cover. So just to recap, the five things that we're thinking about are this transition to autonomy, where, when, and how to use this and where the right boundaries are, to understanding that we're entering this period of acceleration that is not just driven by the technology, but also driven by national policy, and there are going to be implications of that, but also an opportunity for states and cities to really lead. Third is, and sorry, I'm noticing that there's a typo on the slide, but fourth is this workforce, acknowledging the shift, and investing in adaptation. And then fifth is trust, making sure you're earning it every day with every decision and that it's front of mind as we're delivering these pieces of technology.

**Josiah Raiche**

[47:08](https://youtu.be/q0pxALf3AKU?t=2828) And third, we're talking about scaling.

**Cassandra Madison**

[47:10](https://youtu.be/q0pxALf3AKU?t=2830) Thank you. Going from pilot to scale. So let me pull this down and we're going to give ourselves a little bit of time.

* * *

Original page: https://innovate-us.org/navigating-tech-governance-in-2025-policy-procurement-and-ai · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org
