GSA’s approach to artificial intelligence acquisition: ‘do it once’ and repeat, official says

Among the many aims of President Joe Biden’s artificial intelligence executive order is improving how quickly and efficiently federal agencies are able to acquire the technology.

Achieving that goal could mean establishing a governmentwide acquisition vehicle, such as a blanket purchase agreement and applying lessons learned from cloud acquisition, Sonny Hashmi, commissioner of the General Services Administration’s Federal Acquisition Service (FAS), told FedScoop in a recent interview.

“We want to leverage kind of our governmentwide buying power capability to do it once and repeat many times,” Hashmi said.

Purchasing AI is going to be different from software or hardware purchasing “because it’s not just a product that you’re buying,” he said. AI can include hardware, software, cloud, and is trained on data that comes from different companies or could be proprietary to an agency. 

“That becomes a very complicated acquisition environment,” Hashmi said. Making it more complicated: The tension between how fast AI capabilities are moving and government’s desire for stability with the things it acquires.

The following interview with Hashmi has been edited for length and clarity.

FedScoop: Tell me a little bit about the Federal Acquisition Service’s role in this executive order and what your piece of this is going to be going forward.

Sonny Hashmi: So as you know, the executive order is very comprehensive. It sets a very high level of expectation and vision for all of government to attack, and this is truly an all-of-government challenge. Both the challenge and the opportunity side requires coordination between all agencies — in the security space, the privacy space, and of course, the implementation side. The GSA, being a central agency, works very regularly on these kinds of things. We usually play a central role, whether you go back 10 to 15 years when the cloud was a new concept, GSA took the lead of developing a centralized capability, whether it’s through FedRAMP, or contracting vehicles to enable access to the right cloud solutions, right? 

I anticipate seeing a similar role for GSA as we move forward. In fact, some of our team members have been in collaboration with [the Office of Management and Budget] on this topic for some time already. So whether it’s creating a marketplace for capabilities, and how do you even define that? These are the kinds of questions we’re going to be asking ourselves. 

But ultimately, our role is going to be exactly what we do for any other category of product or service that the government needs. We create a marketplace, we identify all the providers who have capabilities in that marketplace, we set the standards for how we’re going to test and validate their claims, we want to make sure that it’s high quality in that process, and to the extent that requires us to develop a cross-agency coordination body, we also have the responsibility of doing that. And then ultimately, our job is to make it dead easy for agencies to focus on their missions, instead of having to figure out how to buy these complicated things on a one-off basis.

FS: What are some of the levers you can pull to help achieve faster and easier AI acquisition?

Hashmi: A few of the things that come to mind: We want to leverage kind of our governmentwide buying power capability to do it once and repeat many times. Purchasing AI capabilities is going to be very different than people know how to purchase software or hardware because it’s not just a product that you’re buying. You have to train that model on training sets. That training set comes from different companies. Maybe that’s internal to an agency. So it’s going to be a product, including hardware, potentially software, cloud, but services and training data all together, right? So that becomes a very complicated acquisition environment. And so what we want to do is to be able to build the right kind of repeatable acquisition environment, so it could look like a [blanket purchase agreement], it could look like some sort of an acquisition vehicle that then allows agencies to do very quick turnaround task orders against. 

So we’ve done that for cloud, for example, when we work with [the Department of Defense]. We’re able to create one environment where the DOD can transact and buy every time they need some cloud service. We’ve done similar kinds of things with, for example, in the telecom space where we spend a lot of time thinking through what the government’s back-office data and voice telecom needs are, but then once the agency is ready, they can very quickly do task order instead of having to do a full competition every time within the marketplace. So instead of spending a year doing an acquisition, we can turn it down to weeks, even, in some cases, right? So that’s exactly going to be the mode for AI, too. 

But before we do that, we’re going to get all agencies together to really understand what the common requirements are going to be. Because ultimately, if you don’t understand what the need is, then we can’t really build a solution that actually is gonna work for our customers. So we’re gonna be doing a lot of listening sessions both with industry as well as government partners. 

FS: What are the challenges you’re seeing? Have you seen anything common when agencies are trying to acquire AI right now?

Hashmi: I would say it’s early days, right? So we’re seeing a lot of early use cases; OMB has done a great job at identifying a long list of use cases that are potential opportunities for AI. We’re going to be looking through those use cases. Because it’s so early, especially in the generative AI space, we don’t see a lot of consistency yet. 

One could argue AI is already being used in the federal government, right? We are using AI within FAS, for example, to identify supply chain risk and be able to automatically kick off risky products and services off our contract vehicles. That’s been in place for some time — agencies using AI for imagery analysis, satellite management, like all sorts of different capabilities. But as we look at this, these are all very different solutions. So in the acquisition space, you need to kind of put like things next to like things so that you can do competition. So the closest analogy I can think of is cloud. Cloud is not one thing, right? [Infrastructure as a service] is very different than [platform as a service], which is very different than [software as a service]. It’s almost farcical to compare a product like, let’s say, Google Cloud services with DocuSign. They’re both cloud companies that serve very different use cases. And so when we think about the cloud marketplace, there’s different pools that you have to establish with different kinds of use cases, and I suspect in the AI space it’s going to be the same because AI is not just one thing.

FS: Are there any lessons learned from cloud that you could apply to the AI environment now?

Hashmi: Yeah, I think the biggest lesson is that the market is going to move faster than we anticipate. And I think the same thing happened with cloud. The market does not stagnate whenever there’s an impactful capability. You’re going to see many new companies start in a matter of weeks and months, you’re gonna see new capabilities scale, and if you’ve actually seen the progression from ChatGPT to GPT3, GPT4, even the foundational models are scaling very, very quickly and morphing very quickly. 

So we’re not going to have the ability to buy for one particular solution because by the time that capability comes online, the capability has already moved on. So we’re gonna have to develop a very agile, continuously evolving kind of an environment that can continuously ingest new capabilities. At the same time, government likes stability. We like to be able to test and validate that a product is actually going to solve a problem. We like to test and validate that a product meets security and compliance requirements, privacy requirements. And so this balancing act between stability so that we can have adequate insight and trust into a product and how fast the market is moving, that’s going to be the tension. And it’s going to be even more acute than in the cloud space. So that’s the challenge that we don’t quite know exactly how we’re going to navigate, but we are aware that this is something that we’re going to have to design in a new way.