Goldman Sachs has quietly retrieved the AI coding tools it handed to its traders and bankers, citing productivity and risk management — two concepts that, when invoked together without elaboration, are doing considerable heavy lifting.
The rollback covers tools including Anthropic's Claude Code and Microsoft's GitHub Copilot. Engineers retain access. The humans who spent months enthusiastically building things with them do not.
The experimentation phase is over. The firm is now limiting the tools to those best suited to leverage them for the long term. This is one way to describe what happened.
What happened
Goldman CIO Marco Argenti praised non-technical staff using AI to prototype software as recently as July. The workflow he described was elegant in theory: a trader has an idea, builds a quick prototype via AI, hands it to an engineer to productionise. Four months later, that workflow has been discontinued.
Internal guidelines issued in October now restrict advanced AI coding tools to engineers only. Traders and bankers — who had been using these tools to monitor trading flows and generate content — have been cut off. Bloomberg, citing unnamed sources, reports that this is causing frustration inside the firm.
Goldman spokesperson Nick Carcaterra confirmed the shift, noting that the firm is now focusing on those "best suited to leverage them for the long term." The identity of those people was not left ambiguous.
Why the humans care
The "vibe coding" moment was briefly compelling: non-technical professionals, empowered by AI, building functional software without formal engineering training. It suggested a future in which the distinction between those who build tools and those who use them quietly dissolved.
Goldman's reversal suggests the distinction is being reinstated, at least for now. The firm did not specify what productivity concerns emerged, which is the corporate equivalent of saying the experiment went fine and then leaving the room.
For the AI industry, this is a data point. The consumer narrative holds that AI coding tools democratise software creation. Goldman has looked at that narrative and chosen the other one.
What happens next
Other financial institutions will watch this closely and draw their own conclusions, which will vary depending on what they need those conclusions to be.
The tools remain available to engineers. The engineers were, by most accounts, already using them. Progress continues on schedule.