Four months ago, Goldman Sachs' chief information officer was publicly celebrating the fact that traders and bankers were using AI coding tools to build their own software. In July, CIO Marco Argenti praised "vibe coding" among front-office staff — a trader sketches an idea with AI, hands the prototype to one of the bank's 12,000 engineers, and something useful comes out the other end.
That experiment is now over. According to TechStrong, citing Bloomberg, internal guidance updated in October limits Anthropic's Claude Code and Microsoft's GitHub Copilot to designated software engineers and quantitative "strats" — the hybrid trading-development specialists. The traders and bankers who had been using the tools for months lost access, and several are reportedly frustrated: the tools had worked their way into daily workflows for monitoring client trading flows, modeling market moves, and drafting content.
The bank's official framing is "productivity and risk management." "We make certain AI tools available to our people to determine the best use cases across different parts of the firm," said Goldman spokesperson Nick Carcaterra. "After a period of experimentation where we assess a variety of factors including productivity and risk management, we put them in the hands of those who are best suited to leverage them for the long term."
What makes this more than one bank's IT policy is what it says about the "every employee is a developer" pitch that has powered enterprise AI sales for two years. On a trading floor, an AI coding agent operates with the user's permissions — which means it can touch material non-public information, cross internal data barriers, and produce code nobody has audited. That is a governance problem engineers are trained to contain and traders are not. "Goldman has said out loud what a lot of banks believe: AI-built software is only safe in the hands of engineers," said James Harrison, CEO and co-founder of Genesis Global. "A trader can build something that looks like an app in an afternoon, but code that isn't in production hasn't created any value yet."
The timing adds a layer of irony. As The Inference notes, the clampdown lands roughly two weeks before Anthropic's IPO pricing — and Goldman is the lead underwriter. One of Anthropic's most prestigious enterprise buyers just told it that seat expansion has stopped.
This is also Goldman's second retreat of the year: it earlier restricted Hong Kong staff from using Claude over data-security concerns, and peers like JPMorgan have tightened oversight to keep staff from running expensive frontier models for basic automation. The industry direction is unmistakable — AI coding tools are being pulled back inside the engineering org, with formal channels deciding what the front office gets.
Watch for two things. First, whether other banks follow with their own pullbacks — Goldman's move gives them cover. Second, what replaces the tools for the traders who lost them: TechStrong reports the withdrawal has already created workflow bottlenecks with no internal substitute. If the bank can't route front-office demand through an approved path, the shadow IT it just banned will simply reappear under a different name.
— via "TechStrong"