IBM lost 13% of its market cap yesterday because Anthropic published a blog post about COBOL. Not a product launch. Not a working migration tool. A blog post, plus a playbook PDF. That is a remarkable amount of damage for what amounts to a positioning statement.

The claim: Claude Code can map dependencies across thousands of lines of COBOL, document workflows nobody remembers, and surface risks that would take human analysts months to find. All true, probably. LLMs are genuinely good at reading large codebases and producing structural summaries. Anthropic demonstrated similar capability with security analysis just last week. Reading code at scale is a solved problem, or close enough.

But reading COBOL is not the same as replacing it.

Anthropic's own blog is careful about this — more careful than the headlines suggest. The tool handles "exploration and analysis phases." Human engineers still define the target architecture, decide which business scenarios need manual validation, and manage the actual translation. Implementation happens one component at a time. The framing is explicitly incremental.

IBM's Rob Thomas pushed back with the line that "decades of hardware-software integration cannot be replicated by moving code." He's not wrong. COBOL systems running ATM networks and insurance claims processors aren't just code — they're code plus forty years of operational assumptions, regulatory compliance decisions, hardware-specific optimisations, and implicit business logic that exists in no documentation anywhere. The programme runs correctly because it has run correctly since 1987. Nobody alive fully understands why.

That's the actual problem. Not translating syntax from COBOL to Java. Any competent LLM can do mechanical translation. The problem is that COBOL systems encode institutional knowledge in their behaviour, and that knowledge evaporates the moment you rewrite the code in something else without first extracting every implicit contract the old system maintains with every other system it touches.

Claude can read your COBOL. It cannot read the forty years of institutional decisions baked into it.

The market reaction was absurd, which doesn't mean the underlying technology is useless. Reducing the cost of the analysis phase — the boring, expensive consultancy work of mapping what a system actually does — is a genuine contribution. That work currently keeps Accenture and Cognizant in business. If Claude Code can compress months of discovery into weeks, that changes the economics of modernisation projects that were previously too expensive to even start.

However. Cheaper analysis doesn't mean cheaper migration. The analysis was never the hard part. The hard part is testing, validation, regulatory sign-off, and the paralysing fear that somewhere in two million lines of batch processing logic there's a conditional branch that handles a scenario that occurs once every eighteen months and will bring down the payment network if it's missing from the new system.

No LLM solves that. Not yet.

Sources: