Sustainment is the crucial process of maintaining and supporting a system or project to ensure its continuous functionality and effectiveness over time. In the context of software and systems, sustainment involves regular updates, bug fixes, performance improvements, and adapting to new requirements. Without effective sustainment, even the most well-designed systems can become obsolete or inefficient.

Booz Allen was entrusted with sustaining a system that had been created by another vendor and existed only as a pilot for years. While the policy library project was designed to integrate with Salesforce, most of the code lived outside of Salesforce in a complex network of 20 Python repositories. As with many systems, considering its age and size (over 31,000 lines of code), tech debt was growing. Additionally, documentation and testing were limited or nonexistent, which made it challenging for new developers to quickly step in with confidence. As one software architect at Booz Allen noted, “Without tests, we didn’t have confidence we could refactor without breaking things.” The result was that some issues lingered longer than anyone wanted, and progress naturally remained slow as the team worked carefully to preserve system stability.

Then the agency introduced its AI-Assisted Software Development Pilot. Part of a larger initiative to find new ways of working with AI, the agency pilot put AI tools in the hands of software developers and implementation partners to see how we could invent new ways of working to simplify and accelerate our efforts. For our policy library sustainment project, this meant we suddenly had access to tools that would provide automated assistance and new insights into the codebase. 

Instead of painstakingly combing through code line by line, our team began using generative AI tools to better understand, test, and modernize the system with more confidence. For Kyle Knab, a Booz Allen software developer new to Python, this was both a challenge and an opportunity: “There’s that initial psychological barrier of taking your hand off the wheel and using AI. But, after the first few successful results, you just get more comfortable with it.”

A massive, thousand-line function that had already been flagged for refactoring—and was considered too risky to touch—deserved to be reconsidered using a test-driven development approach. With AI generating baseline unit tests and suggesting safer ways to restructure the code, Kyle was able to break the monolith into smaller, more manageable parts. As one architect noted, “Writing those tests could have taken a couple of days, given the size of the codebase compared to the size of the team. AI was able to generate tests in less than an hour. Before combining test-driven development and AI, we had spent multiple sprints adjusting logs to understand the issues in the code and got nowhere. Kyle made great progress in one sprint with AI writing tests and suggesting improvements.”

Beyond the big wins, AI also smoothed out the everyday grind. Copilot suggested comments, generated commit messages, flagged outdated patterns, and even helped build scripts that cut weeks of manual work down to seconds. Instead of spending time deciphering cryptic functions or building thousands of Salesforce metadata files by hand, the team could focus on solving problems that mattered. As one systems engineer noted, “In some ways, Kyle gets to do ensemble programming. He’s not having to do it alone. Copilot gives one point of view, and he brings another, giving reassurance that we’ll get the best possible refactoring.”

Perhaps the biggest change was cultural. At first, some developers hesitated to admit they wanted or needed AI. But results spoke louder than pride. Kyle’s perspective captured it best: “I’m not afraid of being replaced by AI. But anybody who’s a developer not using AI may be replaced by developers who do.” Over time, the team started seeing AI as a partner—one that gave them the confidence to stabilize sustainment and finally reduce the backlog enough to take on new enhancements. “The fact that we’ve taken on these new intakes speaks to the sustainment being under control. That wasn’t the case before,” Kyle noted. 

The agency saw the results: problems were resolved faster and a system that was once fragile now stands strong, with increased unit test coverage, better documentation, and higher quality. Across all policy library repositories, spanning three languages, 5,526 lines of code were added and 2,850 removed, bringing the total to near 33,745, a 10% increase over 8 months. In one instance, an intermittent issue that was unresolved for four months was researched and solved in a couple weeks. “For a newer Python developer like myself, it would have taken months. With AI, I’m almost done in two weeks,” Kyle said.  

For the Booz Allen team, the takeaway was undeniable: AI didn’t just simplify and speed up sustainment; it transformed the work. AI is no longer just a tool. It’s the foundation for building the future of federal systems.