Migration and modernization at scale

COBOL migration: know what you have, switch off the rest

By Xavier Franco, CEO de Elternativa · Grupo Zemsania
· 7 min read

Written with the help of AI and reviewed by Xavier Franco

A vast archive of shelves in darkness where only one aisle is lit, with a person at the far end.
In this article

Almost every large company I know has a system people talk about in low voices. It works, it invoices, it balances the books every night… and two people understand it. Often it is written in COBOL, runs on a mainframe that costs more every year, and nobody dares to touch it. Here is how I see it, and why I think the time has come to face it head on.

It is not only your problem

If it is any comfort, the United States government has the same problem, on a different scale. In July 2025, the GAO, its audit office, reviewed 69 federal legacy systems and singled out the 11 most in need of modernization. Eight of them use outdated languages; the two at the Department of the Treasury run on COBOL and assembly language. They are between 23 and 60 years old. Seven carry known vulnerabilities. And only three of the eleven have a complete modernization plan.

I am not telling you this to scare you. I am telling you because the pattern is the same everywhere: the most critical system is the one that is least understood, and that is why it never gets touched.

In practice: write down the names of the people who understand your COBOL system. If the list fits on one hand, the risk is no longer technical: it is about continuity.

The real fear: not knowing what it does

COBOL migration was never scary because of the language. It was scary because nobody fully knew what the system did. Decades of changes, business rules hidden in a batch program, integrations that only one person understands and documentation that stopped being updated years ago.

That is why, for me, the first step is not rewriting: it is understanding. And understanding has to be something you can verify. If AI «interprets» your code, it can get it wrong. Code analysis has to be deterministic: the same result every time, with every rule linked to the line of code it comes from. AI helps to explain and document it, but it does not invent what the program does.

This gives you something many organizations have never had: functional and technical documentation extracted from the code itself, which says what the system does today, not what someone thought it did.

In practice: before asking for a quote to migrate, ask for the system to be documented from the code. If your team reads it and says «we didn’t know it did that», it has already been worth it.

What isn’t used gets switched off

Here is the saving nobody talks about. When you take inventory of a system that is decades old, a lot of code turns up that no longer runs: programs nobody calls, reports nobody requests, processes that stayed behind just in case.

That is not migrated. It is switched off. Migrating what adds nothing means paying twice for something useless.

What stays is what matters: the business, the rules and the data. What goes is the code that runs them (batch and JCL, CICS transactions and 3270 screens) and, in the end, the host that supports them.

In practice: ask how many programs in your system have run in the last twelve months. The difference from the total is what doesn’t need to be migrated.

No big bang and no host downtime

The other big objection is risk. Nobody wants a cutover weekend with the whole business stopped. Neither do we.

That is why the migration is done process by process:

  1. Read-only inventory. The code is scanned without touching anything.
  2. Documentation from the code and classification: what stays, what goes and what is switched off.
  3. Rebuild with orchestrated AI, in the language you choose: Java (Spring Batch for batch and JCL), .NET or, if it is calculation, your data platform. The AI instances rewrite, test and document; your team validates.
  4. Cutover with parallel running. The new and the old process run at the same time and are compared case by case, with your real data. When the last process matches, the host is switched off.

You choose the target language. And what doesn’t run isn’t migrated.

In practice: ask to be shown how an old process will be compared with its new version before cutting over. If there is no comparison on real data, there is no guarantee.

Why it is viable now

For decades, COBOL migration was unviable because of the effort: too much code, too many people, too many years. What has changed is capacity. With orchestrated AI, work that used to require an army is shared among several instances working in parallel, with method, and a team that validates. On the website we publish what we measured on our customers’ code: up to 4 times more industrializable migrations, COBOL included.

And migrating is not just saving the host. It is no longer depending on two or three key people and recovering a system that your team can understand and evolve.

And the people who know it?

There is a question that is almost never asked out loud, but it is on the table: what happens to the two or three people who know the system? My answer is clear: they are indispensable, and in a project like this they become the most valuable people on the team.

They are the ones who know why a rule is there, which process cannot be touched at month-end and which odd case shows up once a year. In a well-run migration, those people stop being the only ones who read the code and move on to validating: they review the documentation extracted from the code, confirm the rules and sign off each process before cutover.

What changes is where their knowledge lives. Today it is in their heads; afterwards it will be written down, linked to the code and within reach of whoever comes next. Your key people bring value, not risk.

By the way, this is not a new problem. In 2019, the same GAO had already identified ten critical federal systems between 8 and 51 years old, one of them in COBOL, and noted that nearly 80% of federal technology spending went on operating and maintaining what already existed. Six years later, in 2025, it was again calling for modernization plans. The cost of not deciding also adds up.

In practice: sit your COBOL experts down with the documentation that comes out of the code. If they validate it, half the project is done.

Where to start

  1. Choose a bounded subsystem, with business value and an owner who knows it.
  2. Scan it read-only and extract the documentation from the code.
  3. Mark what isn’t used and switch it off.
  4. Rebuild a first process and compare it in parallel with the original.
  5. Decide with that data whether to continue with the next one.

No step forces you to commit the whole system. And each one leaves you with something useful, even if you stop.

How we help you

In Code and application migration we migrate COBOL and mainframe, Visual Basic 6, Access, Oracle Forms and PL/SQL to the language you choose, keeping the business rules and the data, with parity measured case by case on real data and no big bang.

We do it with the Datagentis migration module: it connects read-only, takes inventory of what you have, detects what is not used, generates the functional and technical documentation and the work plan, and runs in your environment.

How many people in your company really understand your COBOL?

In a demo, on a sample legacy application, we show you the inventory, the documentation extracted from the code, the rebuild and the comparison.

Frequently asked questions

Can you migrate COBOL to Java without stopping the mainframe?

Yes. You migrate process by process, with the new and the old system running in parallel, and the host is switched off when the last process matches.

What language do you migrate COBOL to?

Usually Java (with Spring Batch for batch) or .NET. If the program is a calculation over data, it can go to your data platform.

What happens to the business rules hidden in the code?

They are documented from the code itself before migrating, and every new process is compared with the old one using real data.

Do you have to migrate the whole system?

No. What doesn’t run isn’t migrated: it is switched off.

Keep reading