Anthropic recently made waves with their messaging around using AI to modernize COBOL systems. The headlines were exciting. The LinkedIn thought pieces were breathless. And our customers started calling.

Can we use AI to rewrite our application in C#?
What tools should we be using?
How do we get started?

We understand why people are asking. When a company like Anthropic says the future is here, you naturally wonder what that means for your 30-year-old business-critical system. But after spending the better part of a year deep in the trenches of AI-assisted development on legacy codebases, I owe you a frank conversation about where I think things actually stand.

The tools are real. The gap is human.

Here’s what catches people off guard: The AI tools genuinely work. They can read legacy code. They can write in languages that haven’t seen a new tutorial in decades. They can trace business logic through thousands of lines of spaghetti and tell you what’s really happening. I recently used AI to crack a bug in one of the brownest brownfield codebases you can imagine. The customer’s team was dead in the water with no clear path forward. As a developer with only a surface-level understanding of the codebase, I directed the AI’s work, starting from a loosely defined bug report, and we banged it out together in two hours. Humans verified the fix didn’t break anything. Then, after the dust settled, AI scanned the entire codebase and confirmed that the same problem (a tiny detail in a field expansion that got overlooked and came back to bite them years later) didn’t exist anywhere else. That original bug had been lurking for 15 years because humans aren’t perfect either.

So, the tools aren’t the bottleneck. The people holding the tools are.

These are not Fisher-Price toys

Here’s how I think about it: These tools are flaming rocket-powered chainsaw scalpels. In the hands of someone who has spent serious time building intuition for what AI can and can’t do, who understands software architecture, who can think like a business analyst and a software archaeologist at the same time, they are extraordinary. Genuinely transformative.

But outside of UI work (where they really are toddler-approved, Fisher-Price safe), they are not safe for beginners. There is no guard rail that will save you from confidently wrong AI output that you don’t have the experience to recognize. And when you’re working on systems that run actual businesses, these tools are simultaneously ultra blunt, extraordinarily jagged, insanely sharp, and slightly radioactive depending on how you hold them.

The uncomfortable truth: If you have to ask how to use AI to rewrite your legacy application, you’re probably not ready to do it. That’s not an insult. We don’t hand jet-boosted power tools to first-year medical students and point them toward the operating room for the same reason.

The one thing that actually is easy (relatively speaking)

UI modernization. If you have an API layer in front of your data and business logic, building a modern front end is the one area where AI tooling has genuinely become accessible to a much wider range of developers. The front-end JavaScript ecosystem is exactly what these AI tools were trained on, and they’re good at it. Really good.

So if you’re sitting on green screens and you have API access to your data (through something like our Harmony Core web services framework or a similar layer), you have a real, practical, near-term opportunity to put a modern face on your application without touching the engine underneath. Get the best business analyst or ultra-senior developer you can find, someone who has been in the trenches with real users for a few decades, sit them down with Claude or Codex, and let them go. The most powerful UI/UX cyborg available to you right now is one of those people paired with AI tooling. The result? Your users get an interface they’re willing to show new hires and a user experience that doesn’t require sticky notes all over the monitors to help them manage janky workarounds.

Everything else on the list—code understanding, testing, refactoring, full translation? You need to know what you’re doing.

The unicorn problem

Someone who can pull off a full AI-assisted legacy modernization needs to be a rare combination: business analyst, software architect, software archaeologist, and advanced AI practitioner. All at once, all in one person.

Why one person? Because coordinating that work across a team multiplies the difficulty enormously. The whole reason the solo-practitioner model works is that the context lives in one head. Split it across three people, and you’re juggling flaming chainsaws in a relay race.

People like this exist, but there aren’t many of them. And the ones who are good at it have been doing nothing but this for months or years. They didn’t read a blog post and start converting mainframe code over the weekend.

Beware the pitch

This brings us to the part nobody wants to talk about. As AI hype around legacy modernization grows, so will the number of consultants and vendors telling you they can do this for you. Many of those pitches will sound compelling. Some will even use the right vocabulary.

The problem: To someone who hasn’t been deep in this work, a credible pitch and a bogus pitch look identical. The person who can actually deliver and the person who will burn through your budget and leave you with a half-converted mess will say a lot of the same things in the sales meeting.

We’re not saying everyone offering AI modernization services is running a con. The ratio of real capability to marketing is just heavily skewed right now, and most organizations don’t yet have the internal expertise to tell the difference.

What you should do right now

Stop waiting for permission to start. Stop waiting for the perfect tool or the perfect prompt template. Here’s what’s practical today:

If you’re not already, start using AI on your daily work. Bug hunting, code review, understanding unfamiliar parts of your codebase, writing tests. Use AI as a thinking partner, not an autonomous agent. You’re building intuition for what it gets right and what it gets wrong, and that intuition is the foundation for everything else. And do not trust the AI. Its lies look identical to its truths. The good news is that it will 100% tattle on itself if you ask the right follow-up questions.

Fix your UI. If you have API access to your data, a modern front end is within reach. If you don’t, ask us about Harmony Core. This is the lowest-risk, highest-visibility win available to you right now.

Do the refactoring you’ve been putting off. AI can dramatically increase your confidence in refactoring work by helping you understand impact, trace dependencies, and verify changes. Be kind to your future self. Stop amassing a garbage dump and start cleaning house, with AI as your safety net.

Build the pyramid from the bottom. A model we like says getting good at AI is about 80% just doing the work, 10% learning from others, and 10% refining your approach. You can’t start at the top; you have to put in the hours. There’s no checklist that substitutes for that.

Your language and your applications aren’t left out

One thing I want to be clear about: AI can write DBL. I have many lines of DBL code with my name on them in git commits over the last nine months, and I didn’t write them. If you’re a DBL developer, you’re not cut off from using AI just because you’re working in a language that doesn’t trend on Hacker News.

And once you get comfortable using AI as a development tool, the next question you’ll ask is “What can AI do for my users and my application?” The answer: A lot. But the prerequisite is the same as it is for UI modernization. Your data and your business logic both need to be visible. When a change happens, you need to be able to trace it back to what triggered it and understand why. Tyler Akidau, CTO of Redpanda and keynote speaker at our 2025 DevPartner Conference, laid this out really well in his session on event-driven architectures and intelligent applications. Once your data is prepared and streaming within your DBL application through our Streaming Integration Platform, you can plug in any third-party AI tool you want. Get those layers in place, and the entire AI ecosystem opens up to you.

The honest outlook

Will AI change everything about how teams build and maintain legacy applications? Yes. Over two, five, ten years, nothing will look the same. But nobody knows what the shape of “different” looks like yet, and anyone who tells you they do is selling something.

There’s a waterline between “accessible” and “expert-only,” and right now, most of the interesting legacy modernization work sits below it. Every time that line moves, even a little, the world changes. We’re watching for those shifts; if we see one, we’ll raise an alert. In the meantime: Use the tools. Build the intuition. Fix your UI. Hunt your bugs. And be deeply skeptical of anyone promising instant modernization, because even the best AI practitioners in the world would tell you that “instant” doesn’t belong in this conversation.