Staff Reports
Goldman Sachs quietly pulled the plug on a lot of its staff using advanced AI tools. In October the firm updated internal guidance to limit access to certain coding and agentic products — think Anthropic’s Claude Code and Microsoft’s GitHub coding tools — and now mostly engineers get the keys. Traders and bankers who said the tools made them faster are not happy. This is the big development everyone inside the bank is talking about right now.
Why Goldman says it did this
The bank says it ran a review and decided the productivity gains weren’t worth the risk without tighter controls. A Goldman spokesperson, Nick Carcaterra, explained the company experiments with tools and then narrows who can use them after weighing productivity and risk. Goldman Sachs Chief Information Officer Marco Argenti has praised non-technical staff for prototyping with models — “vibe coding,” he called it — but he also says such tools need proper supervision and engineering hardening before they touch production systems.
Who loses access — and why traders are grumpy
The move mostly affects traders and investment bankers who had been using these AI coding agents to monitor client flows, model positions, produce charts and write content. These people say the tools saved them time and made tasks easier. The bank carved out exceptions for “strats” — the quant teams that mix coding and trading — and for the engineering ranks. That exception raises questions: if some non-engineers can keep access, why close the door for others who proved the tools can help?
Security, governance — or elites protecting the toolbox?
There are real risks here: agentic tools can run code, touch sensitive data, and behave in ways no one expected. Big firms and regulators are right to worry about leaks, mistakes, and weak audit trails. But there’s also a danger of smothering useful innovation under a pile of bureaucracy. Goldman’s approach centralizes development in engineering teams, which should reduce errors — but it also hands engineers tight control over who gets to build and iterate. Call it risk management or gatekeeping, depending on your appetite for red tape.
What to watch next
The key questions now are simple. Will Goldman publish clear rules about which tools and use cases are banned? Will other banks follow with similar role-based limits? And will regulators use this moment to tighten rules across the board? If the goal is safer systems and protected client data, fine. But if the result is slowing down real improvements to how markets work, firms and policymakers will need to rethink where they draw the line.
Goldman’s October guidance is a reminder that AI in firms is no longer a toy. It is also a test of how much trust we place in front-line workers to innovate versus how much control corporate IT wants to keep. The right answer lies somewhere in the middle — and until Goldman shows the public the balance it struck, people who like both safety and speed will keep rolling their eyes.
