As part of the Government AI Campus, Apolitical is publishing a series of articles featuring leaders from the 2026 Government AI 100, a list recognising public servants around the world who are pioneering AI adoption, capacity building and regulation.

Apolitical spoke with Harrison MacRae, who specialises in the adoption and governance of emerging digital tools within state government. He has led first-in-the-nation efforts to pilot generative AI, including a statewide ChatGPT Enterprise experiment that informed use cases and shaped broader AI governance and training approaches.

There are a bunch of different challenges with this moment. First and foremost, and this has been the challenge of recent years, is how do we keep pace? So much is changing so quickly in the AI space. For governments, it’s really important to try to stay up to date, see around corners, experiment with new technology, but also not over-index on shifts that might last a month or two.

Balancing innovation with moving as fast as we can while also approaching this technology with appropriate stewardship has been an ongoing challenge.

Another challenge has been moving from piloting to scale. We’ve had some great success here in the Commonwealth with different types of pilots, really trying to explore as much as we can early. But moving from targeted groups and approaches to scaling to a workforce of 80,000 staff is really the next challenge, and one we spend a lot of time thinking about.

You led one of the first large-scale pilots of generative AI in state government. What problem was that pilot designed to address, and how did you structure the experiment to learn what AI could realistically deliver in a public sector context?

When this work started in late 2023, ChatGPT had been out for a while, and there were only a handful of other AI tools gaining traction. There wasn’t much understanding of when, where, and how folks could use these personally or professionally. In particular, there weren’t great examples of what it looked like for public sector employees to use these tools.

Governor Shapiro signed an executive order that established a governance board and guiding principles for how we would explore and implement AI in state government. Given where we were at the time, the board identified the need for a pilot to help answer some of these questions and inform what we did next.

Our goal was to understand what using generative AI actually looked like for our staff in their day-to-day, and to use those insights to inform our approach to the technology and contribute to the public sector ecosystem of shared information and research.

We tried not to go in with a premise of “everyone is going to find this helpful.” We were never going to capture the full breadth of an 80,000-person workforce, but 14 agencies participated, and we had a range of technical and non-technical folks, as well as tenured and fairly new folks to the Commonwealth. We were trying to build a comprehensive picture of what it looks like for our staff to use these tools: where they see opportunities and where they see challenges and risks.

During the pilot, how did public servants actually end up using generative AI in their day-to-day work? Where did the most meaningful value emerge, and what hesitations or barriers to sustained use did you notice?

We saw folks using these tools for a wide range of things. Over the course of over a year, it was really great to see people continuing to explore. The tools themselves continued to get better and open the aperture of capabilities as well.

People were using AI to help with text editing, writing, summarising information, navigating data, Excel and coding work. There were three types of personas that we saw amongst our staff, which we named innovation engines, bureaucracy hackers and strategic communicators.

Innovation engines used the tools to be more innovative, solve problems in new ways, brainstorm and come up with ideas about how they could change existing processes. Bureaucracy hackers were taking structured processes and figuring out shortcuts to help still keep themselves in the process, still provide their expertise, but move through steps more quickly. Strategic communicators used generative AI language-based tools to enhance and more quickly address communication needs for audiences they were working with.

To give a specific example, an HR e-learning team, which has to take policy or legal jargon and make it into content that is accessible for a varied workforce, found that AI tools helped them more quickly analyse, digest and turn around copy. They used their expertise to make final edits and ensure it met audience needs, but AI saved them a lot of time and allowed them to serve their agencies better.

“There are so many more ways that folks can apply AI tools in big or small ways than we would ever think, and we found that people are the subject matter experts of their own routine, their own day-to-day.”

It wasn’t perfect for everyone. There were adoption barriers around habit forming, adhering to data and privacy and security standards, and the learning curve of applying a generative AI chatbot to a work product. We didn’t find that there was any particular type of employee who struggled consistently. It was a mix, and different teams had different bandwidth or capacity to really dive in or not.

It was really energising to spend a lot of time talking to Commonwealth staff, who are very public service-oriented, and see them have that ‘aha’ moment of, “Wait, I actually figured out a way to use this that’s helping me not just do my job better, but fulfil that public service mission we have as a team.” It was a helpful reminder of why we need to keep pushing forward, learning more and finding ways to give folks the tools to meet the moment.

Instead of designing governance in the abstract, you shaped governance and training through live experimentation. What did the pilot reveal to you about how AI governance needs to work in practice?

There are a lot of things with generative AI that are built off our existing foundations of how we think about digital infrastructure, data and technology. A lot of our work is founded on existing privacy, data and security practices we have at the Commonwealth.

But there are a lot of new edge cases with generative AI that make us think a little differently about some parts of the governance process. This is a very impressive technological system, but it also has a lot of end user employee interactions that might be a little out of scope of classic technology policy or IT governance structures.

Being able to work with staff and collect their feedback helped inform the types of employee use policies we have: What are the actual questions folks have about usage? How can we address those? Do they reveal solutions that need to be technical or from a training and resources side?

We know this technology is changing very quickly, and it’s important for us to have a governance process that can iterate as well. It’s never going to keep up day-to-day, but we can keep refining it. That’s been a huge emphasis of the work.

What were the most significant risks or constraints you had to manage when you were piloting generative AI at scale? And how did you balance caution with the need to learn quickly?

With any new technology, hindsight will always let you know, did you move a little too fast or a little too slow.

We like to take an approach of recognising that there’s risk in not trying to understand and answer some of these questions. Even in 2023, we would hear from staff that people were starting to explore these tools and had a lot of questions about them. We either could be proactive or reactive.

“We didn’t want to take an approach of ‘AI everywhere, all at once’, but we took a very targeted, small-footprint pilot to try to be proactive, understand some of these questions for ourselves and use that to inform our work going forward. So far that has played out fairly well.”

Our general approach has been, “How can we be proactive in this moment as opposed to reactive, and where do we want to take some calculated exploration and learn from that to set ourselves up for success in the future?”

If another state or government were considering a similar generative AI pilot, what’s one lesson you would strongly encourage them to take seriously from the start?

I would say, on one level, it’s about ensuring you’re providing those feedback loops, engaging staff at all levels, giving folks the opportunity to surface ideas, and helping validate them rather than being too top-down and prescriptive.

As time has gone on, some common themes have emerged across organisations and teams regarding opportunities and successes. But I think that’s been less interrogated in government at large, and there are still a lot of opportunities to bring ideas from staff, big and small, about how they can use these tools.

I’d also offer that there are opportunities to take bespoke product solution technical approaches in terms of deploying AI. And there are also out-of-the-box or existing software that teams might use that have new AI features. Teams need to explore and make that decision of what’s the right way for staff to engage with this, to be those ‘humans in the loop’ and make sure that whatever tools we’re using, AI or not, we’re driving towards the outcomes our organisation wants to see.

Ensuring we have that feedback, scaling the pilot so you’re understanding how technical a tech stack you need, and bringing the employee voice in to inform the work are the key pieces.

Over the next two or three years, how do you expect generative AI to be used across government, and what should public servants start doing now to prepare for that shift?

It’s a tough question. If you asked the foremost AI experts three years ago what today would look like, you’d get a wide range of answers. So I try not to be in the prediction business too much.

Some things are clear. Work in recent years has affirmed the value folks can gain from adopting these tools and applying them to their work. We see these tools continuing to progress in terms of their capabilities, and it’s not uniform across all tasks or work areas at the same time.

I think understanding, as an organisation, what your mission statement is and what you’re trying to solve is essential. Among the range of capabilities that are developing and changing so fast, what are the pieces that are actually helping you get to that place? Having that problem–solution mindset at your core is how organisations will best set themselves up for success over the next three years.

Capabilities are improving. So we have to interrogate how we design the ways people are learning to use these tools and starting to apply them, and ensure it is helping solve the core challenge we have as public sector organisations. That needs to be the centre of gravity throughout any aspect of this world that changes from AI.

What excites you the most about governments using AI? What concerns you?

What excites me the most is that we have a lot of people who are really dedicated to their work and passionate about public service. I see the real opportunity as AI being another helping hand to help them do what they want to do, which is deliver for people in Pennsylvania. That’s ultimately what we’re here to do as public servants.

There’s also the fact that AI isn’t the right answer to every problem.

There are teams, workflows, data, existing tech stack and policy constraints that shape whether something meets the solution you need. AI is expanding so rapidly that we have to be measured and thoughtful in answering that problem-solving question and knowing if it’s the right tool.

The excitement is delivering better for people. The concern is making sure we’re using AI properly and not trying to make every problem fit an AI solution.