Federal agencies need clearer artificial intelligence (AI) governance and a plan for agentic AI if they want to move from experimentation to production, Dell Technologies CTO and Chief AI Officer John Roese told MeriTalk.
In a recent interview, Roese said the government AI market is entering a more consequential phase, with agencies moving beyond pilots and chatbots toward production systems that can improve mission delivery, speed, scale, and efficiency.
But he warned that agencies lacking governance risk getting stuck in what he called proof-of-concept (POC) purgatory.
“I predicted that the word of the year this year would be governance,” Roese said. “The reason I said governance is that we see a direct correlation between companies and organizations that have strong governance … and the most advanced AI adoption and impact.”
“The previous year, I said it would be agentic. I kind of got that one right,” he added.
Governance separates pilots from production
Roese explained that Dell sees a clear pattern across the public and private sectors: Organizations with strong AI governance are the ones most likely to move into production.
“Governance is you have a clear operating model of how you’re going to use this technology and what you’re going to use it for,” he said.
That discipline, he said, separates organizations that scale AI from those that remain trapped in pilots.
“Organizations that lack that governance, that cannot make a decision about what to do and what not to do, get stuck in the POC prison,” Roese said. “They, quite frankly, cannot get out of it.”
Roese said the technology itself is not the biggest obstacle for agencies. Instead, he said, leaders must decide which mission problems matter most and direct resources toward getting those use cases into production.
“At Dell, we chose four areas: sales, services, supply chain, and engineering,” he said. “We focused on them maniacally at the expense of anything else … got them all into production and transformed our company.”
Roese said that experience is now central to Dell’s work with government customers, particularly as agencies look for practical ways to move from AI activity to measurable outcomes.
AI enters its second phase
Looking ahead over the next 12 to 18 months, Roese said federal leaders should prepare for a shift from the first generation of generative AI toward a second phase defined by agentic AI.
“Right now, at this moment in time, the number one thing everybody should be aware of in the AI world is we are shifting from the first phase to the second phase in real time,” he said. “It won’t wait for you. You don’t get to decide to do it three years from now.”
The first phase of generative AI, Roese said, was largely about unlocking data through tools such as chatbots. The second phase, he said, is about AI systems that can perform work more autonomously.
“It’s opening up this giant surface area to digitize the work of the world to change where autonomous work is done,” Roese said.
That shift has major implications for infrastructure, security, data architecture, and workforce training, he said. Agents require more tokens, must operate closer to where work happens, and introduce new risks because they are more powerful and autonomous than traditional IT tools.
“We have never in human history had this kind of autonomy in technology,” Roese said. “There are no examples of IT tools that exist today that are fully autonomous like what agents can do.”
Data, not models, may be underestimated
While much of the AI market continues to focus on model releases, Roese said models are becoming less of a differentiator for success.
“The piece that’s overhyped is models,” Roese said. “There are now many sources, at least four major sources of frontier models, that at any given time, one of them might be slightly better than the other. But they’re all really good, and they’re all really accessible, and they’re all really powerful.”
What is underestimated, he said, is data – particularly whether agencies can make specialized institutional knowledge accessible to agentic systems.
“Bad data, bad AI, that is a truism,” Roese said. But in the agentic era, he added, data quality alone will not be enough.
“The more important part is data accessibility,” he said. “An agent without access to things like knowledge graphs cannot be a specialist. And if you think about government, government agencies are the home of specialization.”
For agencies with deep subject-matter expertise, Roese said the challenge will be organizing data in ways that allow AI agents to use it effectively. Without that step, agencies may struggle to deploy agents that can truly support specialized mission work.
Dell leans on scale and ‘customer zero’ experience
Roese said Dell’s differentiator in a crowded AI market is twofold: its scale as a technology provider and its experience applying AI internally through what the company calls its “customer zero” approach.
“Many of our competitors, we simply outscale them, and that’s a good thing,” Roese said of Dell’s position in the market. “The more important one … is that we have leaned into this technology, and largely people recognize us as one of the most aggressive early adopters of the technology.”
Roese said many of his meetings with government leaders are not traditional sales conversations, but discussions about Dell’s own AI implementation lessons. Those include what failed, what worked, how the company developed governance, and how it moved projects into production.
“We are simply bigger and more scalable than our competitors, and I believe we are a much more aggressive and early adopter of the technology,” he said.
For government CIOs and mission leaders trying to move quickly but responsibly with AI, Roese’s advice returned to governance and focus.
“Experimentation and prototyping are not success; that’s just activity, not outcome,” he said.
The goal, he said, should be to identify the highest-impact use cases, get them into production, and build from there.
“It doesn’t matter about 50 things in your pipeline. It matters about the two things that get over the finish line,” Roese said.