Legacy Modernization with AI in 2026: Real Tools, Real Risks

For years, “modernizing with AI” was more promise than product. In 2026 that changed: the major vendors now offer concrete tools to analyze COBOL, convert SQL Server databases, and upgrade .NET applications. At the same time, Gartner issued one of its toughest warnings on the subject.

Both things are true at once. The tools are real, and so is the risk of overestimating them. This article sums up what’s available, what analysts are saying, and how to use AI with a method in a modernization project.

Gartner’s warning

As reported by CIO Dive, Gartner published a blunt prediction on June 18, 2026: “More than 70% of mainframe exit projects started this year will fail due to overestimation of generative AI capabilities,” in the words of Alessandro Galimberti, VP Analyst at the firm.

The accompanying recommendation is just as relevant. Gartner suggests using generative AI to modernize the mainframe in place, and deciding on full exits case by case rather than as a default strategy. The firm also estimates that 75% of mainframe-exit vendors will adjust or end their business models by 2030.

That doesn’t mean AI is useless. In late July, CIO Dive cited Morgan Stanley, which is modernizing 17 million lines of code with AI assistance. The difference lies in the approach: AI as an accelerator within a method, not as a shortcut to rewrite everything at once.

What vendors already offer

Between June and August 2026, AWS, IBM, and Microsoft announced concrete capabilities:

Tool	What it does	Keep in mind
AWS Transform: SQL Server to Aurora PostgreSQL (August)	Offline conversion from uploaded DDL: tables, schemas, stored procedures, and functions. Validates functional equivalence and adapts dependent .NET applications	SQL Server conversion is only available in US East (N. Virginia)
AWS Transform continuous modernization (August)	Connects to GitHub, GitLab, or Bitbucket, scores repositories on tech debt, security, and “agentic readiness,” and opens pull requests with fixes	Generally available in all AWS Transform regions
AWS Transform for mainframe (June 16)	New traceable “reimagine” workflow for z/OS COBOL and PL/I	Focused on z/OS
IBM Bob Premium Package for Z (June 26)	COBOL, PL/I, assembler, and JCL; “Z Understand” static analysis including Control-M/TWS schedulers; metadata repository for estates of 10,000+ programs	Sales-led add-on to IBM Bob
GitHub Copilot .NET modernization agent (July 9)	Assess → plan → execute workflow, in Visual Studio, VS Code, Copilot CLI, and the Copilot app	In public comments on the announcement, users report it handles class libraries well but struggles with ASP.NET Core projects

A clue in how the tools are designed

There’s a telling pattern in how these tools are built. AWS talks about validating functional equivalence and a traceable workflow. IBM instructs its model to answer from analysis metadata first. Microsoft structures its agent into assessment, planning, and execution.

In other words, the vendors themselves assume a generative model alone isn’t enough: it needs static analysis, traceability, and verification. Anyone framing a project as “we hand the code to the AI and it gives us back the new system” is ignoring what the tools themselves take for granted.

Why AI alone isn’t enough

• Business rules are buried. Code shows what a system does, but not always why. A literal conversion can reproduce historical bugs or drop exceptions the business relies on.
• Dependencies go beyond the code. Scheduled jobs, file-based integrations, reports, and shared databases are part of the system too. That’s why IBM includes analysis of schedulers such as Control-M and TWS.
• Equivalence has to be proven. Without test cases and representative data, there’s no way to claim the new system behaves like the old one.
• Results vary by project type. The comments on the .NET agent are a good example of why a pilot should come before committing to a timeline.

SQL Server 2016: the clock has run out

For many Latin American companies, the most urgent case isn’t the mainframe but the database. SQL Server 2016 reached the end of extended support on July 14, 2026. Microsoft offers paid Extended Security Updates (ESU) for three years, but that buys time; it doesn’t solve the problem.

It’s a good moment to decide with a method: upgrade to a supported version, migrate to another platform — for example, Aurora PostgreSQL with help from AWS Transform — or pay for ESU while a plan is executed. Each option has different implications for the applications that depend on that database.

A three-phase method

1. Assess

• Inventory applications, databases, scheduled jobs, and integrations.
• Static code analysis and a dependency map; this is where AI helps a lot.
• Classify by business criticality and frequency of change.
• Check the support status of each component, such as SQL Server 2016.

For a first pass at this phase, you can use our legacy system self-assessment.

2. Trace

• Document business rules and validate them with the people who know the operation.
• Link every new component to the legacy component it replaces.
• Build equivalence tests with representative data before converting.

3. Modernize incrementally

• Start with a well-bounded module with visible value.
• Apply patterns like Strangler Fig or expose the current system through APIs.
• Review every AI-generated deliverable with automated tests and human review.
• Decide on a full platform exit case by case, as Gartner suggests.

What about Delphi, VB6, or PowerBuilder systems?

Not all legacy runs on a mainframe. Latin America is full of critical applications built in Delphi, Visual Basic 6, or PowerBuilder, often connected to SQL Server. The announcements reviewed here aren’t aimed at those technologies, but the method is the same: assess, trace, and modernize piece by piece, with AI to read, document, and propose, and people to decide and validate.

Frequently asked questions

Should we exit the mainframe relying on generative AI?

Gartner recommends using generative AI to modernize in place and evaluating full exits case by case. A full exit built on high expectations of AI is precisely the risk scenario its prediction describes.

Do the AWS, IBM, or Microsoft tools replace a modernization team?

No. They speed up analysis, conversion, and fixes, but business rules, equivalence testing, and architecture decisions still require human judgment.

What should we do if we still run SQL Server 2016?

Its extended support ended on July 14, 2026. You can pay for ESU for up to three years while you carry out a planned upgrade or migration.

Conclusion

AI has already changed the economics of modernization: there are tools that analyze, convert, and propose fixes at scale. But Gartner’s warning is clear: overestimating generative AI is a fast way to make a project fail. The value lies in combining these tools with rigorous assessment, traceability, and incremental delivery.

If you’re weighing how to modernize a critical system, see how we support legacy system modernization together with specialized strategic partners, our approach to AI for business, or get in touch to review your case.