Most big companies have a system like this. It pays the staff, or settles trades, or books freight, and people steer around it in meetings. It works fine. The engineers who built it have retired, the documentation peters out somewhere around 2009, and plenty of the business rules sit inside nested IF statements that one senior developer kind of remembers.
Ask anyone who has tried to replace one, and they’ll tell you the new code was never the hard part. Understanding the old code well enough to touch it without breaking payroll is what eats the budget. And that’s precisely where generative AI has started to help. It reads unfamiliar code, explains what a module does, follows dependencies, and drafts the documentation nobody bothered to write. Vendors have noticed. Hanoi-based Newwave Solutions, for one, now sells AI-assisted conversion as part of its modernization work, including a tool for moving Visual Basic 6 apps onto VB.NET.
Morgan Stanley offers the clearest example so far. The bank built DevGen.AI in-house on OpenAI’s models and switched it on in January 2025. Five months later, it had gone through roughly nine million lines of legacy code and saved developers an estimated 280,000 hours. Most of that came from turning COBOL and other aging languages into plain-English specs that engineers could rewrite from. The bank’s tech chief was candid that fully automatic translation isn’t there yet. Reading the old stuff was the win.
Why nobody wants to touch it
Legacy software is basically a pile of quick fixes, most of them undocumented. A screen that looks simple might rely on a hidden rule, an old integration held together with tape, or a manual step only two people in finance know about.
Switching it off isn’t an option. These systems carry something the business needs daily, so every change happens while the thing is still running, which makes everyone slow and jumpy.
Knowledge is the other gap. The original builders have usually left, and younger developers may never have written a line of COBOL, Perl or VB6. Someone has to work out what the system does, and what might snap, before any real modernizing starts.
Where AI genuinely helps
Engineering judgment still matters. What AI changes is the starting point, which ends up far better than a blank wiki page and a nervous team.
Reading code nobody recognizes
Business rules get spread across thousands of files, and the official docs drifted from reality years ago. AI can sweep a large codebase, explain modules no one recognizes, sketch how everything connects and pull out rules buried deep in the logic. A team that once needed months to get its bearings can now get there in days.
Finally writing the manual
From that same analysis, AI drafts module summaries, explains APIs, walks through workflows and helps rebuild technical specs. New hires get something to read. The company stops depending on two people’s memories.
It does need a human check. Old systems are full of exceptions that come from how the business really works, and a model can’t guess a rule that only lives in someone’s head.
Translating and tidying
When a language or framework has to go, AI can translate sections of legacy code into something newer, flag duplicated logic, improve readability, and suggest cleaner structure. Treat the result as a draft. It still needs review, testing, and an architecture check before production, for the same reasons covered in the real risks of AI coding.
A new front door on an old house
Ripping everything out at once rarely makes sense. AI can help draft API definitions, service wrappers, and integration code around the existing system, so newer apps can call old functions while the core stays put. Good software integration for AI agents and modern apps buys time: open up access now, replace pieces later when the numbers justify it.
Tests before anything moves
The real fear is changing behavior nobody knew existed, and many legacy systems have barely any tests. AI can generate unit tests, regression scenarios and test data from the existing code and workflows, then help compare old and new outputs so surprises surface early. Anything business-critical still gets a person’s sign-off. The coverage just shows up much sooner.
Picking a starting point
This decision is about the business more than the code. Which applications need attention now? Which can be left alone? Which need deep surgery?
AI can help weigh complexity, dependencies, usage, and technical limits, and lay out whether to refactor, rebuild, or leave things as they are. The final call stays with the business. A crusty old app that barely matters may not deserve a budget yet, however ugly the code is.
The economics are moving, too. In McKinsey’s 2026 State of AI survey, 32% of organizations said they’d skipped buying at least one software product or feature because they could build it with agentic coding tools, and about one in five are already scaling coding agents. Old build-or-buy maths that said “leave it alone” is getting redone.
Doing it in stages
Pick something that hurts. A system causing real operational pain, or one with obvious business value, makes a better first project than whichever is oldest.
Write down what exists before changing anything: the stack, databases, frameworks, dependencies and integrations, and especially the corners nobody documented.
Refactor in small pieces. AI suggestions are a first draft, and every change gets reviewed and tested before it’s merged.
If the interface matters, let AI rough out early designs and UI code, then put the prototypes in front of real users before production changes.
Start with low-risk components and leave the core for later. Agree on each phase’s scope, tests, and rollback plan before anyone writes code.
Build the delivery pipeline along the way. Automated tests and CI/CD make releases repeatable and cut out manual steps.
Then watch what happens after release. Performance, errors, and user behavior will tell you whether reliability and usability actually improved, and what to tackle next.
Short on people?
Launch day isn’t the finish line. Someone has to deploy, monitor, and support every change, and AI can help there with deployment scripts, configuration files, release notes, and log analysis, or by summarizing errors when something starts misbehaving.
Plenty of teams simply don’t have the headcount to do all this alongside their day jobs. Custom offshore software development services are one way to fill that gap. Newwave, for example, runs dedicated teams, staff augmentation, QA, DevOps, and post-launch maintenance out of Vietnam, with ISO 27001-certified security practices. Whoever takes it on, the working rule doesn’t change. AI drafts. Experienced engineers check.
So what’s different now?
Generative AI won’t make legacy modernization a one-click job, and the most credible case studies say as much. It does shrink the slowest, most expensive part of the work, which is figuring out what the old system does in the first place. Systems that once looked too risky to touch are getting easier to size up, document, test, and improve.
For the first time, teams can see clearly what they’re about to change before they change it. Which system at your company does everyone still tiptoe around, and what would it take to finally understand it?
