DoD created Tradewind as its capability to develop and fund solutions to challenges in fielding AI and machine learning, as well as digital and data analytics tools.
Bonnie Evangelista, the Tradewind execution lead within the CDAO, said Wednesday at ATARC’s federal IT modernization summit that AcqBot generates text much like ChatGPT, and is part of Tradewind’s effort to accelerate DoD’s contracting process.
“We have a working prototype. Whether it’s that, or it’s changing people’s mindsets about how you can use acquisition authorities to go faster or to get a better end-state for your user, we’re trying it all, because we want to go after the high-value tasks,” Evangelista during a panel moderated by Federal News Network.
Evangelista said the contracting process within DoD is rife with manual processes and “antiquated process mediums to do business,” such as filling out PDFs and paper forms. The CDAO is standing up its own contracting shop, which Evangelista said she’ll lead once it’s up and running.
Evangelista said the AI tool doesn’t make any contracting decisions, and that there’s a human in the loop reviewing and validating the AI-generated text at every point in the process.
“We’re trying to start training the model to help us build a problem statement from nothing. Can somebody put a descriptive title in the tool, and help it develop a pretty articulate problem statement?” she said.
The CDAO is looking for feedback and user testing to train AcqBot and improve its functionality.
“You don’t expect glory overnight. No AI has an out-of-the-box solution … I don’t think it takes a ton of data just to prove the concept, and if we can continue to prove the concept and then create a demand signal for this type of tool, that’s where we can make it optimize the technology,” Evangelista said. “When AI comes out of the box, it’s like children, and even children have to learn and grow, just as much as AI also has to learn and grow. I’m hoping that, because the technology has enabled us to basically solve really hard pain points in my functional lane, we’ve given it a chance to be enduring.”
Evangelista said AcqBot is part of Tradewind’s “go big” approach to AI and demonstrating what’s possible with the technology.
“We’re not going to settle for the low-value stuff. We want to, if we can, break the mold and break the glass ceiling, from an acquisition perspective,” Evangelista said. “I’m not interested in automating a workflow that is not helping emerging technology get into the hands of soldiers faster. I’m more interested in, can I speed up the entire process so that we can point, click and buy? Is that even possible? Can vendors write proposals on their phones? Those are big bold statements, and those are the things that, from a practitioner’s perspective in contracting, those are the things we’re considering.”
That go-big approach to automation is also happening across the civilian federal government.
U.S. Citizenship and Immigration Services is taking the next step in implementing robotic process automation by rollout out “semi-unattended bots,” according to Meikle Paschal, USCIS’ program manager for RPA.
Paschal said these semi-attended bots solve a problem with attended bots, which prevented employees from completing tasks while the bot was running on their computers.
“I have a job to do, I can’t sit here idly while this bot is running on my computer,” Paschal said.
Paschal said semi-attended bots allow employees to have a bot running in a second window of their computer, “like picture-in-picture,” and allows employees to continue working on their own tasks while the bot is running.
“It allows people to use their computers again, so now we’re doubling your capacity right there,” Paschal said.
USCIS this year is trying to understand how best to credential fully unattended bots, and is looking to other agencies to solve this governmentwide challenge.
“What we’re trying to do is figure out how the credential will work, so that we can have these perpetual bots running more than eight hours a day. They’re coming, we’re just figuring out some of these challenges,” Paschal said. “The benefit is that at DHS, a lot of our systems, a lot of our security requirements are similar enough that if one of us figures it out, then we should all be sharing that information with each other.”
Paschal said USCIS is shifting away from focusing only on “low-hanging fruit” opportunities for automation, and instead raising its sights on high-impact use cases.
“If you really want to make the biggest dent with your resources, go and automate a high-impact, high-value workflow or process and get your resources back. You want your people to be able to focus on the hard things that only a person can do,” Paschal said.
USCIS also built an automation that uses optical character recognition to bring more than 30,000 paper applications for asylum requests and feed them into the agency’s content management system.
“Now that person has an opportunity to come to the States and seek asylum. Those are the opportunities where it’s changing the way that we do business,” Paschal said.
Paschal said the semi-automated process, which includes an employee manually validating what’s been added to the system, takes less than 10 minutes. Before the automation, it took USCIS staff about 30 minutes per application to manually process.
“Typically, when you have a problem, you just hire more people and throw more resources, more people transcribing. We don’t have to do that, and not only is this process being done at multiple service centers, but we just standardized the way to do it,” he said.
Greg Singleton, the chief AI officer at the Department of Health and Human Services, said the National Institutes of Health is automating some of the workflows of grants applications — specifically who in the agency should review tens of thousands of incoming grant applications.
“There’s this big bottleneck in terms of saying, ‘Well, who are we going to give this grant to? Are going to give it to the expert on Zika? Are going to give to the expert on AIDS? Are we going to give to the expert on proteins?” Singleton said.
Automation tools using natural language processing help NIH identify keywords, and then figure out which agency experts should be reviewing the grant applications.
“Rather than sitting in a queue for 30, 60, 90 days, it just goes over to the expert that gets the money into the hands of researchers faster, and makes a significant difference in how research is conducted,” Singleton said.
Singleton said HHS is looking to automate workflows for high-volume administrative tasks, such as responding to Freedom of Information Act requests or responding to congressional inquiries.
“Things have increased in volume over the past 10 to 20 years, but our workforce hasn’t increased to keep up. We’ve all gotten more efficient, we can instantly send messages over to each other, and we can generate more material, more information, but it’s basically the same number of people. So how do we adapt? Do we either slow down, or do we do a worse job? Or do we come up with tools and solutions to allow us to excel?” he said.
Singleton said agencies have demonstrated that they can “solve the easy stuff with RPA.” The Centers for Medicare and Medicaid Services, for example, use automation to help process approximately 12 million requests for reimbursement a day from care providers across the nation.
“If we can take these systems and get them into a mode where we automate the easy stuff, and then we’re left with the, let’s say 15% of the workflow that’s weird, or we don’t know what to do here, or that’s a question. And then the workers can apply their expertise,” he said.