As OpenAI’s first government hire, Felipe Millon is on the front lines of artificial intelligence (AI) adoption across the public sector.
In a recent interview with MeriTalk, Millon, who leads government sales at OpenAI, said the government is confronting what he called “the biggest technological inflection point in history.” He urged agencies to continue to “unblock” AI access so civil servants can understand the technology’s trajectory and what it means for government missions.
“You don’t get fit by reading about working out; you have to go to the gym. And in the world of AI, going to the gym is using AI,” Millon said.
For OpenAI, he said, the importance of public sector adoption is tied directly to its mission: “Our mission is to ensure artificial general intelligence (AGI) benefits all of humanity. That does not happen without the U.S. government being involved.”
From access to adoption
OpenAI analyzed more than 1 million enterprise deployments to understand what drives real adoption – and one factor outweighs the rest: executive use, Millon said.
“The biggest predictor of whether a specific agency or customer will meaningfully adopt AI … is how much the executives use AI – by far,” he said. “Leadership sets the tone of the conversation.”
Millon said he is seeing similar patterns in federal adoption, citing AI familiarity and engagement among senior leadership at the Department of Health and Human Services (HHS) and the Department of the Treasury as key to unlocking broader use within those organizations.
Millon also highlighted the role of communities of practice – particularly those that form organically through self-volunteering – and the value of hands-on learning environments, such as hackathons, that encourage users to test boundaries and share best practices.
OpenAI’s vantage point on what works and what does not comes from its position at the frontier, Millon noted. Because the company is building the underlying models and moving quickly to get them in the hands of users, he said, his team “knows what’s coming” and can advise agencies on patterns they see across deployments.
From experimentation to operational use
Several workforce-scale deployments demonstrate where government organizations are successfully moving beyond pilots, Million noted.
A major example, he said, is the Commonwealth of Massachusetts rollout of ChatGPT Enterprise across its entire executive branch, which includes nearly 40,000 state employees. The state is coupling the rollout with role-based training and enablement for staff.
On the federal side, Millon cited accelerated adoption through OpenAI’s $1 General Services Administration (GSA) offer.
“We really leaned in with the federal government in our $1 GSA offer … We’ve had a lot of agencies that have leaned in with us. HHS is a big one … they’ve publicly shared that they’re saving tens of thousands of hours using ChatGPT Enterprise already. They have it available to their entire workforce.”
He also cited active deployments at the Treasury Department and the Office of Personnel Management (OPM), noting progress in day-to-day operations.
Beyond adoption of general-purpose tools, Millon pointed to more bespoke builds, including work with the Department of Energy’s national laboratories. He described an on-premises deployment at Los Alamos National Laboratory on its Venado supercomputer.
“Early last year, we actually deployed our model, for the first time ever, outside of our cloud infrastructure – on prem,” he said. “[The Los Alamos National Laboratory] is able to run our models against some of the nation’s most sensitive data to perform operations that they previously weren’t able to do at scale.”
The rise of AI agents
The discussion around agentic AI adoption grew through 2025, but reliability and tooling “weren’t quite there yet,” Million said. Now, he said, task-specific agents are beginning to click – particularly in areas such as deep research and software development.
Millon described a shift from “AI-assisted coding” to “true AI agentic coding,” where tools can execute larger blocks of work with less step-by-step prompting. At OpenAI, developers use agentic coding heavily, he noted. Developers act more like orchestrators who review and direct agents rather than coders who write every line by hand.
In government agencies and other large enterprises, Millon said the adoption of coding agents is still early, which he attributed in part to longer security and authorization timelines. But he predicted that as agent reliability improves – and agents can run for longer periods with fewer interruptions – adoption will increase.
That shift, he said, is often driven by moments of surprise when users see what the tools can do. “Sometimes we internally at OpenAI call them the ‘feel the AGI’ moments – that moment when you sit back, and you’re like, ‘I cannot believe a computer just did that,’” he said.
A practical blueprint for the next 12 months
To create more of those moments – and turn them into measurable impact – Millon urged agency leaders to take a two-part approach: a bottom-up workforce strategy and a top-down mission strategy.
Millon’s bottom-up advice was blunt: Make general-purpose AI tools broadly available; clarify what is permitted; and invest in training, enablement, and communities of practice.
Millon also argued that messaging around responsible AI, while important, has sometimes created uncertainty about what is permitted for routine work, such as summarization and research support. He urged agencies to provide clear guardrails so employees can use AI appropriately without fear of violating policy.
For the top-down strategy, Millon said leaders should start with their agency’s highest-priority mission outcomes. “What are the things your political leadership has said, ‘These are the three to five things we need to do’? … Think about those things first, then ask where AI can apply to them.”
That kind of strategic work typically requires small, empowered tiger teams, he said. “It has to be people who know the existing databases and where all the data sits,” he observed. “And then it’s about building for those specific use cases.”
Millon applauded the recent AI use case libraries but cautioned against trying to manage hundreds of AI projects at once. Instead, he encouraged leaders to prioritize a short list of high-impact efforts.
The simplest advice: begin
Millon’s advice to those trying to move quickly with AI was simple: start. “People need to learn by doing,” he said.
He referenced Ethan Mollick’s “jagged frontier” concept – the idea that models can be superhuman in some tasks and surprisingly weak in others. He said agencies have to test tools in context and set expectations for what “good” looks like.
“How do you know where that frontier is? By using the technology,” he said.
His bottom line: The pace won’t slow down for government – and agencies that wait for perfect clarity risk missing the window to shape adoption responsibly through real experience.
“Go get started,” he said. “Go do it. Like, right now.”