Marina Fox, who leads the DotGov Domain Services in Office of Governmentwide Policy at the General Services Administration, said the agency is turning to artificial intelligence make sure agencies comply with Section 508 through its Solicitation Review Tool (SRT).
“On FedBizOpps.gov at least 500 new solicitations are posted every day. It’s a lot and that’s just one of the sites. There are others where you can find solicitations. We realized it’s not sustainable and it’s not scalable. We could hire 1,000 people, but it wasn’t going to work,” Fox said at the recent ATARC big data summit. “We need AI and machine learning to be able to scan through entire population posted to FedBizOpps and separate the ones that are IT. Then scan those using a predictive model and predict their compliance with Section 508.”
GSA has to rely on several algorithms to ensure agencies are meeting all the different parts that make up Section 508 standards.
Fox said GSA built the algorithms by identifying the key words based on data from training courses for contracting officers who were taught to manually review solicitations.
“We predict the compliance of all the solicitations. We post them on a web portal and then subject matter experts and agency users can log in and see what’s on their docket. In version one, you still need human intervention in order to provide useful feedback to make the predictive engine much smarter,” she said. “We are currently migrating and getting ready for production. We will go into production using the cloud.gov platform. The entire application is built using open source technology so we are scalable, agile and can always add more use cases.”
This means, Fox said, GSA could eventually use this same approach with AI for contract clauses required for sustainability or cloud security or so many other mandated provisions.
Previously, Fox said GSA personnel would review one percent of all solicitations based on a random sample in a span of 6-to-12 months. Those that didn’t comply with Section 508, GSA would send letters to those agencies asking them to fix those already awarded contracts.
“There was no tracking mechanism to see if they corrected it,” she said. “And there have been lawsuits where GSA and the government have been sued for non-compliance.”
Fox said by using AI, GSA got out from underneath an expensive contract that would only “scratch the surface” resulting in low compliance with Section 508 requirements.
“[Compliance] is really an issue and it starts with the solicitation quality. If the government doesn’t release a solicitation that contains federal regulatory requirements, then technically companies don’t have to abide by that,” she said. “The education component is important, but also because the procurement system across the government is not consistent, statements of work get posted and they all have a variety creative, non-structured data so our job is to catch it all. When we realized that this was a real need, we had a data scientist as part of our staff and it all started by just developing some scripts in python and trying to see how we could scale it.”
New IT accessibility rules catch up to 17 years of changes
In the future, Fox said the solicitation review tool will have a prediction and recommendation engine, and customer agencies can fix the solicitation in the web portal. The end goal is for the tool to alert agencies before an RFP is even posted on FedBizOpps.gov and what has to be fixed.
Fox said her office started this effort to use AI to look at solicitations about a year ago by building the predictive model using 1,000 RFPs, which GSA previously reviewed manually.
Then, GSA tested the AI models for about nine months by scanning FedBizOpps.gov every 24 hours.
“We had an extensive user acceptance testing, using Section 508 coordinators and other accessibility experts who are reviewing the solicitations that this tool predicted as compliant and non-compliant so that we can catch false positives and false negatives,” Fox said. “The prediction rate, the accuracy, is 95 percent, which is very high for non-structured data. We have great confidence in the tool. We started with a hard case because there is so much written about it, the law is so extensive and there are so many exceptions that have to go into place. We believe now that we’ve tackled the hard one, other ones will be a little bit easier.”