The outfit has launched its own homebrew DevGen.AI in January, based on OpenAI’s GPT, and it’s already chewed through nine million lines of code. Morgan Stanley global head of technology and operations Mike Pizzi said that the code has saved developers 280,000 hours. 

Rather than spitting out shiny new Python scripts like your average AI coder, DevGen.AI translates old languages such as COBOL or even firm-specific gobbledegook into plain English specs. Once translated, human developers can re-engineer the stuff in modern code, minus the head-banging.

Pizzi said commercial tools were decent at writing new code but pants at understanding older languages. “We found that building it ourselves gave us certain capabilities that we’re not really seeing in some of the commercial products,” he said.

Apparently, the bank didn’t fancy waiting for off-the-shelf tools to catch up. Instead, it trained DevGen.AI on its own codebase, even including languages that were never popular outside a Morgan Stanley server room.

Now, 15,000 developers across the bank use the tool not just for translation, but to dig out specific code bits for regulators or internal queries. It can even fully translate small code chunks into modern languages, though Pizzi admits the results might not always be efficient or pretty.

That’s why humans are still needed to turn the tool’s output into properly structured, performant code. The real power lies in generating plain English blueprints that describe what the code actually does, he said. That’s a skill going the way of the dodo, as fewer devs understand the old languages.

Morgan Stanley has already got hundreds of AI use-cases in production, including tools to automate workflows and churn out new apps. But none of it works without modern, standardised code architecture, Pizzi said.

“You’re always modernising in tech. Today, with AI this becomes even more important,” Pizzi said.