CBP says AI coding tools accelerate legacy software migration by about 40%

U.S. Customs and Border Protection says AI coding tools are helping it move old software components into modern systems about 40% faster. The reported gain concerns work on its Automated Commercial Environment, a long-running modernization effort where developers still check the tools’ output rather than accept generated code without review.

Thomas Mills, executive director of CBP’s Cargo Systems Program Directorate, described the results during a panel at the recent ATARC AI Summit. Federal News Network published his remarks on November 5, including an excerpt featured on its Ask the CIO program.

Working through decades of code

The Automated Commercial Environment, or ACE, contains infrastructure developed in the early 2000s, Mills said. CBP also recently retired a mainframe dating to the 1980s. His account places the AI-assisted work inside an existing modernization program, rather than describing a new system built entirely with AI.

The migration work reaches several kinds of software: stored procedures in databases, database components themselves, and code that parses incoming messages. Mills identified C and C++ among the older languages involved. He said some of that code uses languages he knows but others do not, highlighting the expertise needed to understand what is being replaced.

Mills said the original estimate for modernizing the older infrastructure was five to seven years. For components expected to take a long time, he described a 40% compression in the timeline. That was a claim about particular modernization components—not an announcement that the entire ACE program had finished or received a new completion date.

Understanding the old system still takes people

The tools’ usefulness extends beyond producing replacement code. Mills said they help staff understand what existing database procedures do and give them a clearer picture of the software. In his description, code generation supplies insight into older components as well as material that people must inspect before trusting it.

CBP is also using AI for automated data extraction and embedding it in data-governance work to clean up information, Mills said. He warned that large language models—the systems behind many generative AI tools—still produce bad results when given bad data. The agency’s cleanup work is intended to give subsequent AI use a more accurate dataset.

Two-week deliveries help expose costs

For many software projects, CBP sets two-week development sprints to deliver capabilities incrementally. Mills said those small releases help the agency estimate the total cost of delivering a capability. He recalled an initial surprise over cloud spending and said breaking development into smaller pieces helps predict what that spending will be.

Those iterations also feed the case for further funding. Mills described showing what a use case delivered and what the team learned, then using those findings to seek money for expansion. If funding does not follow, he said, CBP cannot scale the work as much as it wants. His approach is to start with small pieces, fail quickly when necessary, and identify ways to scale.