Your agency isn’t ready for AI
![](https://cdn.nextgov.com/media/img/cd/2021/12/11/humanmachine/860x394.png?1639281703)
To truly take advantage, government must retool both its data and its infrastructure.
![human machine interface](/media/ps360/GIG/EDIT_SHARED/AI_Automation/humanmachine.png)
Artificial intelligence has been around for more than half a century, yet in many ways it's poised to be 2019's hot new thing for federal IT. The commercially available toolkit is growing rapidly, a new executive order emphasizes AI's transformative potential, and several provisions of the President's Management Agenda explicitly depend on intelligent automation. Perhaps more importantly, projects across government -- from chatbots that deliver citizen services to purchasing-data analytics -- are proving that AI can be put into production.
So far, though, success stories are mostly small and scattered. FCW recently gathered a group of specialists from across government to discuss what's needed to bring AI into the federal IT mainstream. The discussion was on the record but not for individual attribution, and the quotes have been edited for length and clarity. Here's what the group had to say.
First, get your data house in order
The ability to make sense of unstructured data is a key selling point for AI solutions, but the group stressed the limitations that come with sloppy datasets.
"It's all about people interacting and providing value to the data," one participant said. "Then the machine can be intelligent about it."
Part of the challenge is settling on standards, but multiple participants said the real key will be effective ontologies that connect and "translate" between multiple standards. "If we can really make our data smart, then the AI and the machine learning are just going to be phenomenal," one participant said.
Even though users of specific datasets have very specific needs, another executive said, "there's an upper ontology that we all should agree to -- [such as] what is a person? -- those kinds of simple things. If we agree to that, then everything underneath is based on that domain. We agree that this means that, and here are all the things that are associated, and yes, that's the parent of this. You have to do that hard work, and once that hard work is done, then wow. You can…use it for everything."
The government is already in catch-up mode on those efforts, another participant said, adding: "We can't get ahead of AI because it's already out. People are running it, and they're actually doing things."
"If you don't do those simple foundational things, I think we'll be creating chaos," a third participant warned.
![FCW Perspectives](/media/ps360/GIG/FCWNow/Logos/fcwperspectives.png)
Participants
Gil Alterovitz
Presidential Innovation Fellow
Department of Veterans Affairs
Jim Chen
Professor of Cybersecurity, College of Information and Cyberspace
National Defense University
Shelby Hritz
Federal Sales Leader
Google Cloud
Robert Monroe Jr.
Computer Scientist
Defense Department
Jim Rahai
IT Specialist
Environmental Protection Agency
Donna Roy
Executive Director, Information Sharing and Services Office
Department of Homeland Security
Dave Shepherd
Program Manager
Department of Homeland Security
James St. Pierre
Deputy Director, Information Technology Laboratory
National Institute of Standards and Technology
Shannon Sullivan
Head of Federal
Google Cloud
Pamela Wise-Martinez
Chief Enterprise Architect
Energy Information Administration
Note: FCW Editor-in-Chief Troy K. Schneider and 1105 Public Sector Media Group Chief Content Officer Anne A. Armstrong led the roundtable discussion. The Dec. 11 gathering was underwritten by Google Cloud, but the substance of the discussion and the recap on these pages are strictly editorial products. Neither Google Cloud nor any of the roundtable participants had input beyond their Dec. 11 comments.
Second, fix the infrastructure
According to multiple participants, another key ingredient is getting agencies' IT infrastructures in order. Cloud-based services power many advanced analytics tools, so preparing data and workloads for the cloud is critical.
"There are powerful algorithms out there," one participant said, "but do you have the backing infrastructure to support it?"
Another participant divided agencies into three key segments. First, "you have people who are terrified of the cloud. They want to say they're doing something in the cloud so they take their existing processes and put them in. But there's no point to it. It's probably more expensive than when they were running it on-premises."
The second segment has accepted cloud technology, and "you've just got to get your arms around your data," the executive said. "You need to start labeling data, and you need to have data that you can use." The third segment is "more of the cutting edge, and that's where we've really been successful. Health sciences, Department of Energy, NASA, some of the leading Defense Department organizations -- they're embracing [AI], and they're the ones that are really ready."
When one participant cited U.S. Citizenship and Immigration Services as a prime example of AI readiness, another said, "The only reason they can do that now is because for five years they were in the crux of their digital strategy and [then-CIO] Mark Schwartz put them on the right path. They can actually do something now that they're in the cloud. It's tied to your modernization strategy -- the way you're building applications and the way you're treating the data."
Other agencies are starting to follow suit. "Somewhere around 27 percent of our IT systems are now either in the planning process or in the migration process," one executive said. "Our goal for this year is 35 percent. The more we focus on how to get to the cloud and the security and the training, the more we can understand that getting the data to the cloud is probably the point."
Finally, learn to trust the algorithms
The group also said the "black box" of how AI solutions reach their conclusions must be better understood before broad buy-in is possible.
"There are biases in your data, and there will be biases in your results," one executive said. "I'm not sure we're ready for maximum use of machine learning and AI until we persuade folks it's safe."
Agencies are looking for help in that area, another participant said, adding: "I think there are a lot of pockets of good work happening, but I don't know that at the federal CIO level, the strategy and the President's Management Agenda around data is going to meet the need of where we want to go with machine learning and AI."
A third participant said: "I really think the next landing spot is the privacy impact testing of the algorithms themselves -- getting through that process of explaining what the algorithm is doing in a way that everyone, including the public, understands."