Agencies now have access to a new artificial intelligence tool that has been test driven by the General Services Administration for the last eight months.

Through the USAi platform, agencies can take advantage of capabilities that include chat-based AI, code generation and document summarization.

“We started with 10 users, and then we moved to 100 users, and then 1,000 users, and eventually rolled it out to all employees across GSA,” said David Shive, GSA’s chief information officer, in an interview with Federal News Network. “The business portfolio of GSA is very broad, and they started to apply this thing across those broad business domains. They started to do things like writing code that would satisfy some code development needs that they had completed in hours, not days, weeks or months. They started doing data analytics across multiple data sets, that would normally take days and weeks, and they were doing it in hours.”

Shive said the employees really started to show the impact of the USAi tool during Friday demonstrations last spring led by Stephen Ehikian, then the acting administrator and now deputy administrator.

Shive said 60% or 70% of those demonstrations were employees using USAi to “automate or augment the work that they were doing by reducing the drudgery of their day-to-day work, or making it so they didn’t need to swivel between five systems to do one thing or they’re writing code snippets to automate the drudgery of their work day.”

GSA is offering USAi as a no-cost shared service.

Agencies just have to sign a memorandum of understanding with GSA and determine what databases they want to integrate with the large language models.

Shive said agency customers will have access to USAi through a GSA-hosted single tenant cloud environment.

“We don’t need to be in the business of running other agencies’ technology, not because we would be bad at it, but because the business mission of each agency is very different. The risk profile of each agency is very different. It doesn’t make sense to have a one-size-fits-all for everybody,” he said. “We’ve developed a multi-tenant architecture where each tenant, the underlying infrastructure, is all the same, but each tenant is independent. Each agency has full control over the entire stack, including things like the provisioning of their users, assessing that telemetry and those behavior patterns across people and technology, and making risk management framework based decisions on what’s acceptable to them and what’s not.”

Access to six commercial AI models

Shive said several agencies are already interested in using the platform, but he wasn’t ready to say who or how many.

Through USAi, GSA is offering three different options and access to six different commercial AI large language models, including Amazon, Microsoft, Google, Meta, Open AI and Anthropic.

One of the services is a basic chatbot.

The second is the API layer to make it easier to integrate their datasets into the chatbot. “The connections and the mechanics and stuff are the same for every tenant, but how they configure those is up to each agency to determine,” Shive said.

The third service is the console, which is a management dashboard that lets each agency tailor the tooling and the create user requirements based on things like risk tolerance.

Shive said USAi received its authority to operate at the FISMA moderate level.

“The idea there is that we built something in what we call a model garden. The idea there is that not all models are the same. They all have different strengths and weaknesses, and we want agencies to be able to get access to whatever model they need that they think is suitable for the work that they’re trying to do,” he said. “Also knowing that those strengths and weaknesses are leapfrogging themselves on a very rapid pace, it is a wildly dynamic industry environment out there, this gives agencies the ability to try before they buy at scale. In that rapidly changing, highly dynamic space, they are able to point whatever work they’re trying to do at the highest value model. Value can be defined any number of ways. It’s not just cost. It’s usability, resistance to bias, resistance to cyber attack and other things. We give them unfettered, full visibility into that telemetry, and then they can make the decisions based on whatever their internal criteria is because that is the domain of chief information officers to make those decisions at the agency level. We just give them the tools to be able to do that.”

GSA started building USAi about 20 months ago, but its AI journey started more than a decade ago.

The agency took several important steps to get to the point where USAi could emerge.

Oversight process for use cases

Zach Whitman, GSA’s chief data scientist and chief AI officer, said at the recent 930gov conference, sponsored by the Digital Government Institute, that once generative AI became the next iteration of the capabilities, the agency established a generalized fair use policy for commercial offerings. He said that gave employees the ability to start to think about how they can start to use GenAI tools as part of their day-to-day workflows.

“One of the major things we understood was folks were afraid. They were afraid of how it would be perceived to use the tool as part of their work. They were afraid of data exfiltration problems. Am I putting the wrong thing in the tool? They were culturally afraid of being perceived as cheating or not necessarily doing the work and having an advantage. And there were a number of different hurdles that we were facing, some technical, some from the use case perspective, should I be doing this?” he said. “It became a clear mandate for us not only technically to figure this out, but also culturally. In the policies that we’ve been receiving from the White House, we’ve had a huge advantage.”

Whitman said GSA set up several gates for the AI tools to jump through before being offered widely across the agency.

GSA created its own evaluation factors based on the mission area as well as its own safety team made up of technical and subject matter experts.

Whitman said an executive team took all this feedback and made risk-based decisions about which tool could be used.

Through this process, Whitman said GSA saw a need for a common AI platform, which became USAi.

“What we did was we ended up building a platform using a lot of open source tooling. Again, nothing very original. It was all off-the-shelf concepts where you would take a chat bot interface and then you would end up building an API layer that interfaced across multiple hyperscalers, which allowed you access to a variety of different models,” he said. “You had a way to measure the behavior of the users on that system. Now the interesting piece of pulling a bunch of models together is it gives you a really good opportunity to compare models based on the use case. We understood that certain models were better than others, certain newer models were better than others, certain larger models were better than others. But that’s not always the case, and so what we ended up doing with the telemetry was less about monitoring the users and more about how we discern and make buying and business decisions based on the different models.”

Whitman added that GSA’s pilot and data showed the chatbot solves a majority of the use cases employees were interested in using AI for.

He said the chatbots met the basic standards of user experience and they didn’t require customization.

“What we are doing is making sure that we can have interoperability between the different models and maintaining the newest models as quickly as possible, so that the agency has access to these latest models, and then we run our evaluation set against them to understand which model is best suited for which purpose. And we’re doing it in a way where we try to convey to a subject matter expert, not an AI specialist, but a subject matter expert, to make a decision as to which model they need for their specific use case,” Whitman said.

Copyright
© 2025 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.