“The bottleneck hasn't disappeared; it has simply moved. It’s no longer in the initial coding. It’s in the follow-up.”
– Darren O'Meara

‍

AI for mainframe modernization is advancing quickly. New tools, new claims, real progress, it’s all happening at once. At heart, every modernization effort still comes down to three core activities: understanding what you have, translating legacy code into something modern, and debugging and testing until it works properly.

The industry has made rapid progress applying AI to mainframe modernization, with each level of maturity building on the last. But is it — to mix metaphors — skating to the bottleneck? I present to you the three levels of AI maturity for mainframe modernization.

‍

Level One: Conversion-First

The foundational approach: translate the code, preserve the structure and logic, improve iteratively from there. AI has made this faster and more tractable than earlier generations of tooling, and that's no small thing. But it only moves one needle: time. And you’re leaving out lots of other important stuff: consider the kerfuffle over Claude’s COBOL capabilities—and our pov.

‍

Level Two: Insight + Conversion

Level Two is where AI is shining. Understanding a large legacy codebase traditionally took armies of analysts and years of effort, and AI compresses that dramatically. Same with translation, including the type that iteratively refactors JOBOL so it starts to resemble Java.

Speed helps you find problems faster, but it also exposes more of them, shifting the bottleneck to fixing what you’ve uncovered. So even if you’ve gotten to “idiomatic” Java faster, teams still get stuck in cycles of debugging and rework.

Put another way, the risk remains stubbornly high. You still don’t have the certainty that what you’ve rebuilt can help you change faster than before. “99.9% functional equivalence” and “production-ready” are not the same thing.

True, tools like AWS Transform are accelerating this level dramatically, making large-scale conversion and insight more accessible than ever. But we think that adds to the bottleneck (and the number of “humans in the loop” needed).

‍

Level Three: Behavior-First

This is where it gets interesting. Instead of starting with the code, you start with what the system does by capturing real execution data from the live system as your ground truth. Rewrite against that. Validate continuously. When one slice is done, it’s done - and ready to go into production, where you can start changing it easily.

It’s ready because the specifications the AI writes to are deterministic: derived from actual production data flows. You’re integrating that long tail of debugging and testing into the process of rewriting itself, because now you can be confident that AI can satisfy the scenario.
‍

——
‍
This is why the trajectory of AI maturity in mainframe modernization goes beyond doing the same things faster and reaches the actual business outcome most organizations want: the ability to change and adapt faster than anyone else.Constraints have shifted: from translation to comprehension and discovery, and now, to the most opaque and expensive work: debugging, integration testing, and cutover. The real question is no longer how fast you can rewrite the system, but how confidently you can change it once you do.
‍

News and Views

No one got hurt in the end. IBM’s stock rebounded, and the winner was that everyone listening came away slightly more informed about just how complex mainframe modernization is. Gartner’s First Take phrased it well (subscription required) and even mentioned Mechanical Orchard. More commentary linked above.

When 57% of UK adults use mobile apps to pay, it’s understandably unsettling when transactions you don’t recognise appear out of nowhere. That was the thought during the first part of March, when yet another IT glitch had folks asking a familiar question: have the banks’ legacy systems done it again?

Production is the hardest part when shipping code. And Charity Majors articulates that production is exactly where the rigor should go. Be still my heart: “AI changes things. It brings the cost of migration and instrumentation down dramatically at the same time as it turns the evolutionary rate of change into a bottleneck and existential risk.”
‍

From the Orchard

Using Claude to convert COBOL is a great starting point for speeding up mainframe modernization. But Ingo Wiegand, our VP of Product asks, “what will reduce risk?” Rob Mee, our CEO, also chimed in via a piece in CIO.com.

We joined forces with our partner Thoughtworks and noted industry analyst, Arun Batchu, for a webcast on what it takes to safely and confidently modernize mainframe systems. If you’re located in Chicago, Arun will be moderating an in-person discussion on April 7 as part of our Unfinished Business series. Reserve your spot.

Our COO, Edward Hieatt, will be speaking at the Enterprise AI Summit in San Jose on April 10 about ways to harness AI for software development, particularly in creating deterministic outcomes.

On April 15, we’re partnering again with Thoughtworks to talk about all things payments at the Future of Payments Forum in Amsterdam. Register here.

We’ll be at Google Next in April! Hear directly from a customer of ours, or attend this panel on AI’s role in transforming operations. And of course, we’d welcome the opportunity to meet.


Curious to learn more? Say hello@mechanical-orchard.com.‍

‍

—

View all newsletters here

‍