Challenges and opportunities abound with AI-assisted updating of legacy software

When it comes to updating legacy software, AI can help—but IT pros need to be smart when they change things in an environment with potential knowledge gaps.

Antique software can still be mission-critical, and teams need to take care in how they manage transitions to newer platforms.

“One of the biggest challenges is, how do I step up and capture—as fast as possible—all the knowledge that has been already dumped into that code?,” Agustin Huerta, SVP of digital innovation at Globant, told IT Brew. “That code, in a way, is a representation of the history of how that software came to be.”

History lesson. Organizations are often trapped by their legacy infrastructure, in part through predatory licensing and knowledge capture, said Paul Holland, CTO at Astadia. That combines with brain drain as people leave the workforce.

“A knowledge gap is a big issue; people don’t tend to talk about that so much because it exposes a weakness and even compliance issues and governance issues,” Holland said. “Agentic AI can really help with this.”

History is important when you’re working on revamping legacy systems in order to contextualize the process for both human software architects and generative AI solutions. Huerta said that making sure parameters are understood prior to buildouts, or spec-driven development, is a useful way of approaching the problem, as is keeping perspective on the way we understand “legacy”—something a former Globant staffer told him.

“As soon as you ship code into production, that code is already legacy…it’s something you have to deal with,” Huerta said. “Legacy code is not necessarily years old, it may be a month old—as long as you have a lack of knowledge of what happened there, you have a challenge.”

Potential. AI models are good at reading and parsing out code, Holland said, and that capability is snowballing as machine learning improves. But even that has its limitations given that training models on different programming languages is time- and resource-intensive, often creating a snowball effect for the knowledge gap—but by no means an insurmountable one.

“I’m not saying it’s perfect yet, but it’s definitely getting better and better,” Holland said. “The idea is that you’ll have an agent or sub-agents that’ll be able to tell you what’s happening in any of these application code bases and document it and guide the program.”

“Of course, programmers aren’t going to be programming it anyway,” he added. “There’s another agent that’s going to do that.”