RAG Could Unlock AI’s Human Potential in Government

Retrieval-augmented generation (RAG) is a technique that is still fairly untapped in artificial intelligence use within the government, according to OpenAI Partnership Manager Alex Bonnell.

“RAG is the ability to say you know what, I actually know the information I want to have a relationship with,” Bonnell said at the AI in Action Workshop in Washington, D.C. Thursday. “So much of human-machine teaming is going to be about how we show up, what is my ambition?”

RAG integrates large language models with external knowledge bases, enabling AI to deliver more accurate, context-aware answers and perform tasks beyond its training data. The technique could help federal employees tap into their own expertise and pool insights to turn scattered data into actionable intelligence.

Bonnell said in her experience with RAG, she is able to “draw on the totality of my thinking versus just having to redo that every time.”

She added that RAG enables users to create their own personal relationships with AI, which — when pooled together with colleagues’ relative experiences — will open new opportunities and break down silos.

“You’re going to see individuals start to curate their own little knowledge universes. But that also means that we don’t work in vacuums,” she said. “You’re going to start to see teams have collections of, ‘hey, here’s all the information on this program or something that we’ve curated together.’”

AI Cuts IT Tool Sprawl

Bonnell, who previously served as CIO and Digital Capabilities Directorate director of the Air Force Research Laboratory (AFRL), noted she helped the lab consolidate 183 different tools, systems and platforms to 47 when she introduced RAG and empowered users to “solve their own problems.”

Bonnell said that though AI’s potential is exciting, the relationship between the human and the machine will be key to using it effectively.

“I am ultimately using this tool to curate the kind of place and impact that I want to have,” she said. “I think there’s such opportunities we think about, not just AI literacy, but when we think about human potential, how is it that we’re intentional about how we show up?”

She added that AI is the first “truly intimate technology” because it allows users to interact and engage with it on a personal level relative to their experience.

“What you’re good at, what you know, your incredible expertise and experience — you are going to query or have a different relationship with knowledge than I am, and you’re going to get to show up in your terms,” she said.

The Four Stages of AI Adoption

Bonnell outlined four stages of AI adoption beginning with people exploring the technology, learn how to use it and see what is possible. She emphasized that when people interact with AI on their own terms, “they’ll often take it up much faster than if you introduce it professionally because it kind of lowers the risk feeling.”

The second stage comes as users confront and move past their fears or misconceptions, gaining the confidence needed to make the most of AI.

The next stage is when users recognize how AI can save time, energy and effort, which can sometimes trigger guilt over feeling like they are taking shortcuts. She stressed that AI training should show users “it’s not cheating, you’re optimizing. We actually want you to bring the most efficient you to work, but actually telling someone you’re going to feel this way and it’s okay, it’s actually a sign that you’re advancing, that you are using it in a way that’s useful.”

The last stage marks the point where AI becomes a routine part of the job, with users more likely to embrace it as a tool rather than resist it out of fear or uncertainty.

Targeted Use for Maximum Impact

Bonnell noted how AI won’t be effective in all aspects of life or work, but there are places where it can be influential specific to the user’s experiences.

“[AI] shouldn’t be used for everything, but for where it can, where you have clarity on who you want to be and the potential that you want to be able to fulfill, I think it has the ability to support that or contribute to that in ways that we didn’t even see six months ago,” she said.

She warned that the speed of AI development could change dramatically, but that the effectiveness of AI adoption primarily comes down to a user’s willingness to augment their work, rather than use it purely as a shortcut to critical thinking.

“If you decide to show up as someone who looks for shortcuts, who looks to cheat, who doesn’t value the opportunity to learn and to grow and the investment that we’re making in you, you’re going to do that, whether it’s with AI or anything else,” she said.