AI-Assisted Mainframe Development: Where We Are Two Years On

What's actually landed

• Code comprehension at scale.
• Understanding what a 30-year-old COBOL programme actually does, before touching it, is still the biggest bottleneck in most modernisation efforts — and this is where AI tooling has made the most tangible difference.
• Documentation generation.
• Programmes that had no meaningful documentation now often have a usable first draft, generated from the code itself.
• Test case generation.
• Faster coverage of legacy logic that nobody wants to touch without a safety net.

What hasn't materialised

Full autonomous rewriting of legacy COBOL into modern languages remains far less reliable than the tooling demos suggest. The core issue is unchanged: mainframe logic frequently encodes decades of undocumented business rules, and no tool — AI or otherwise — can infer intent that was never written down anywhere.

The practical takeaway

Treat AI-assisted tooling as an accelerant for the unglamorous work — comprehension, documentation, test coverage — rather than a shortcut past it. Organisations getting genuine value are the ones pairing the tooling with experienced mainframe engineers who can validate what the AI produces, not the ones trying to remove people from the process entirely. That combination — proven mainframe expertise plus the right tooling — is still the fastest, safest route through a modernisation programme.