Generative AI: Start small but scale fast

David Knox, the chief technology officer for industrials, energy and government at Oracle, urges federal agencies to take a measured approach to generative AI.

Tom Temin@tteminWFED

March 25, 2024 7:21 am

3 min read

Generative artificial intelligence is unlike any technology that’s come along in recent memory. One reason: You’d be hard pressed to find an application or process to which generative AI doesn’t apply. In some sense, it can do more than it cannot do.

That, plus the technology’s sudden emergence in media and at so many industry conferences and gatherings, has organizations worried. They want to avoid rapid obsolescence by failing to adopt generative AI right away.

David Knox, the chief technology officer for industrials, energy and government at Oracle, urges federal agencies to take a measured approach. He says you might be able to do anything with generative AI, but you can’t do everything. Knox recommends starting with what he called proving grounds — low-risk, low-complexity processes with which to try out AI.

Even in such entry use cases, it’s wise to test the work in an isolated “sandbox” environment while you evaluate the benefits and ensure the requisite security and privacy controls remain in place.

If the resulting AI-powered application does work as intended, agencies then need to be able to create “a path to production,” Knox says. That means knowing your compliance framework and requirements in advance, as with any new technology deployment.

Identifying the proving ground generative AI applications will enable users to operate in what Knox calls the “find/fail fast/fix” mode. For example, nearly every agency deals with human capital processes such as hiring, retention and performance management. Knox advises choosing a single process and using real data to test both the efficacy of generative AI and whether the resulting process retains those crucial compliance measures. Then, if it does, apply the process in a limited production scope before scaling to agency-wide use.

Given the wide potential of generative AI, nearly every federal process is a candidate for its application. Knox says a less intuitive, perhaps, but potentially large payoff candidate for generative AI would capture and preserve institutional knowledge held by the generation of federal employees eligible to retire, the boomers and late millennials.

Such people “have an incredible amount of institutional knowledge of just what are the programs, what are the processes, how to get things done,” Knox says. The information may or may not be written down, and if it is, often no one knows where. Recording knowledgeable people’s answers can create an unstructured database to which the application of generative AI can be ideal. He cautions that unstructured knowledge capture can be risky, especially when using generative AI to capture it. Knox advised a trust-first mindset, with human supervision, rather than using people’s answers to directly train AI.

Knox cited procurement as another internally facing function where such knowledge capture would have big payoff, asking senior practitioners about systems, policies, acronyms, norms of operating, compliance rails and other things to be aware of.

One externally facing area for applying generative AI is citizen serving applications. Here, Knox said, agencies can use AI for what he termed document understanding. He defined that as going beyond optical character recognition and parsing data in individual fields, and beyond converting documents to digital images by essentially turning them into an interactive knowledge base.

In all cases of AI deployment, Knox said, it’s important to keep in mind the human factor, because people often think AI will somehow replace them. Given the complexities of federal procurement, HR management, finances and accounting, grant-making and program management, Knox said, people will realize instead how AI will benefit them — not by replacing them, but rather by augmenting their decision-making and removing routine or repetitive tasks connected to their jobs.

“We didn’t get rid of people when we invented calculators,” he said. “We’re not going to do that with generative AI.”