How much do you really trust agentic AI to make a mission-critical decision without you?
On Oct. 8, Leadership Connect and Quantifind co-hosted a rare, actionable webinar featuring senior government leaders from NIH, Commerce’s Bureau of Industry and Security (BIS), the Department of the Air Force, and Quantifind’s senior leadership.
The result was a robust conversation covering where agentic AI belongs in the federal government, how much autonomy you should give to AI, and what it really takes for public sector leaders to trust AI with mission-critical decisions.
Attendees included CIOs, CTOs, and chief AI and chief data officer teams; acquisition, procurement and vendor-risk leaders; cybersecurity, risk and AI governance leads; research and health agencies modernizing their IT; mission and program leaders; and AI test, training and workforce leads from across the federal government.
Missed Our Webinar? Read This!
When Leadership Connect’s James McKenna polled the audience about where their organizations stand with agentic AI, the most common answer was that they are still exploring or learning about AI use cases.
The challenge for many agencies who have decided to adopt agentic AI is getting their people to actually use it. To address this, George Forbes shared a scientific approach to solving that problem.
George Forbes, Chief Information Officer, Office of the Chief Information Officer, Bureau of Industry and Security (BIS) for the U.S. Department of Commerce
Six months into his role at BIS, Forbes was blunt. His bureau’s AI goals aren’t ambitious enough yet because AI fluency is still low. His fix comes from marketing, not IT.
“We found our 14%, which were our early pioneers, and then 41% were our early adopters. That was critical mass. So, what is 14% of our adopted target community? That’s 200 people? Find your first 200 people,” Forbes said.
Most agencies have rolled out AI tools, but far fewer have people using them well. Forbes’ point is that adoption does not happen on its own. At the Air Force, his team used the same S-curve concept that marketers consider when launching new products.
An S-curve describes how a new tool spreads from a few enthusiasts, speeds up as their colleagues see it working, then levels off once most people are on board. Over time, this pattern typically looks like a stretched-out S.
In his previous role at the Air Force, Forbes’ team found that the first 14% of users gave them hands-on help and let them pull in the next wave until adoption hit critical mass at 41%.
For CIOs and workforce leads whose AI tools are deployed but barely used, this is a plan you can copy. Turn the percentages into names, find your first 200 users, and help them achieve AI fluency one on one. Then use those early users to bring in everyone else.
“If you do leave it to chance, then you leave it to chance with that steepness of that curve. If you want it to be steep, put effort into it,” Forbes added. “If you are okay with it being very shallow, that is what most likely will happen. That is how we have used it; that is how we intend to use it. We have seen great success!”
Dr. Jennie Khun, Chief Information Officer and Director, Office of Information Technology for the National Institutes of Health (NIH), U.S. Dept. of Health and Human Services (HHS)
“I love George’s comment about AI literacy,” Khun said. “That’s such an important part of framing the conversation. When do you know people are familiar and are able to explain what is happening? I think for us, and probably a lot of the agencies and organizations in general, [the question] is how do we help people help themselves?”
Khun picked up on Forbes’ point about AI literacy and took it straight to security. Her answer isn’t to clamp down. She explained that “helping people help themselves” means accepting what is already happening.
The goal is to change how agencies think about what an agent is. Agents aren’t users. They aren’t just tools. They’re non-human identities that can make decisions and act at machine speed.
“A few years ago, automation was scripts, service accounts, and API keys. Now, these agents really need to be treated like a full identity,” Khun said. “They are unique objects. We have to authenticate and authorize them. They have to be monitored. It has to be revocable.”
For any federal leader who needs buy-in, Khun’s lesson is simple. “Nobody’s super excited about security unless you work in security,” she said. Her suggestion was to frame it around what people care about. “This prevents your work from being stolen. This makes sure that your data sets are accurate.”
Adam Mulliken, Chief Product Officer for Quantifind
Before the conversation could move on, Quantifind’s Chief Product Officer Adam Mulliken jumped in:
“I think we as leaders acknowledge that the governance framework is actually the durable part of what we’re all building. The models will change. The tools will change. I think Brad [Worley], focusing on test evaluation, probably has the best job security out of any of us because that’s really what’s required and is going to persist.”
Later during the panel discussion, Mulliken explained what makes AI work, which is understanding the job before you automate it. Pointing a general-purpose tool at a problem rarely succeeds, he said. The customers who succeed come ready to explain how they finish their work.
“I bring up the example of quality assurance. Maybe … you do QA on 10% of the work that’s maybe the most critical, or that’s how much budget you have for QA. That status quo of 90% gets no QA is maybe a bigger risk than taking AI and running it fully autonomously across 100% of the things that you do.”
For mission and program leaders, that flips the usual question. The risk isn’t only what AI might get wrong. It’s everything already going unreviewed.
For CIOs, CTOs and chief AI and data officers, Mulliken’s advice was to look past chatbots. Chatbots are a good start because they spark curiosity and help teams find use cases. But putting agents to work means rebuilding processes from the ground up.
“The objective there is just mapping out that process. Like we’ve been talking about, being able to really understand and get into some of the tradecraft and latent information that maybe isn’t so obvious. How is this work actually getting done today? And being able to map that out and work with a partner that can help you design the right solution for that kind of purpose.”
Maj. Brad Worley, Chief, Cyber, Digital and AI Programs, Test and Evaluation, U.S. Department of the Air Force
Mulliken’s mention of “tradecraft” struck a chord with Maj. Brad Worley, who quickly agreed. Worley calls it tacit knowledge: the “tricks of the trade” that never make it into the policy document. It also proved Mulliken’s earlier joke: if governance is the part that lasts, the person who tests whether AI can be trusted may have the most secure job on the panel.
Specialized knowledge is also why Worley leans so heavily on industry. When McKenna asked the panel how they rely on their industry partners, Worley didn’t hesitate:
“We couldn’t do it without our partners. And industry is really leading the way in a lot of different things.”
Building on an earlier point Forbes made, Worley said the Air Force now tells partners the outcome it needs instead of handing them a list of technical requirements. What matters is whether a solution is resilient and delivers the mission result, not exactly how it gets there. He also pointed to how openly the AI community shares what works:
“Everybody is so open to talk about this stuff and say, hey, how do we make this better? How do we get this right?”
For acquisition and procurement leaders, that means writing solicitations around outcomes and letting partners propose how to get there. For CIOs and chief AI officers, it means you don’t have to figure this out alone. And it brings Mulliken’s point full circle: partners can only build the right solution if you share how your work is completed.
Biggest Agentic AI Takeaway
Partners can build the technology, but Worley was clear that trust is still the agency’s job. Earlier in the hour, as the panel questioned whether AI should be treated as just another tool, he offered the line that sums up the whole conversation:
“I’ve never had a hammer tell me which nail to hit.”
A hammer waits for you to decide. An agent can decide for you. So, how much do you really trust agentic AI to make a mission-critical decision without you? As these four leaders made clear, trust isn’t something you assume. It’s something you build.
Every panelist kept coming back to the same things: AI literacy that leads to real adoption, agents managed as identities, governance built to outlast the tools, and partners who understand how the work really gets done. The main takeaway for the attendees was to build those four, and you’ll know exactly which nails the hammer can pick on its own.
Don’t Miss the Next Conversation
What you just read is a handful of quotes from a full hour. The people who joined live heard every answer in full, shared their own questions with the panel, had the opportunity to earn a CPE credit and saw in real time where their peers stand. They also heard what did not fit here:
- How Quantifind spent a year with the Defense Innovation Unit building AI-enabled vendor due diligence
- Why Forbes is moving BIS to outcomes-based, “as-a-service” buying
- How NIH uses AI centers of excellence to build literacy
- Why Worley believes human attention is the one resource in AI that doesn’t scale
- What each panelist expects in the next two years
Watch the full webinar on demand to catch the rest.
