AI can rewrite COBOL and IBM investors are panicking

Even though Big Blue suggested using AI to do just that – years ago

A blog post by Anthropic pointing out that Claude Code can refactor legacy COBOL applications appears to have sent IBM’s share price down by double digits.

Although COBOL is everywhere, especially in sectors like finance and government, the number of people who understand it is shrinking every year.

Developers retire, documentation is out of date, and only a handful of universities are teaching COBOL to new students.

Rewriting COBOL applications isn’t easy. As Anthropic says, “You aren’t just updating familiar code to use better patterns, you’re reverse engineering business logic from systems built when Nixon was president.”

AI tools like Claude Code and OpenAI’s Codex, however, can massively speed things up, modernising “in quarters” what could have taken years. They can map dependencies, document forgotten workflows and identify hidden risks.

So, why does this matter to IBM, and why did its share price take a 13% hit soon after Anthropic published its blog?

It comes down to IBM’s mainframe business: the core of Big Blue’s original success.

Those COBOL applications aren’t just legacy software, they require legacy hardware to run – that is, mainframes. Moving off of COBOL means, finally, moving off of mainframes as well.

Investors appear to have been spooked by Anthropic’s post, going on an IBM selling spree. The thing is, Anthropic isn’t the first to make this point about AI: IBM itself highlighted the capability three years ago.

Back in 2023 IBM developed a tool called ‘watsonx Code Assistant for Z’ designed to refactor COBOL apps to Java (which has its own problems, but at least it’s a widely understood language).

SaaS businesses, too, have been affected by the spread of AI. Share prices are falling at the likes of ServiceNow, Salesforce, Atlassian, Adobe and HubSpot.

That’s because customers are seeking to build their own tools using AI and cutting the number of suppliers, or at least the seats they pay for. "If an AI agent can handle the workload of five people, a business only needs to buy one license instead of five," Aurelian Research wrote this month.