Goldman Sachs restricted some AI coding tools, including Claude Code-style agents and similar GitHub environments, for non-engineering roles such as traders and bankers, Bloomberg Law reported. The bank cited productivity and risk concerns when staff without engineering controls use agents that write and run code. The policy does not ban all generative AI, but narrows who may operate powerful coding agents. It fits a wider bank trend of allowing models by role rather than with a single firm-wide switch.

What it means: Coding agents change how work gets done, yet in a financial firm a bad script can touch client data, risk models or trading systems. Limiting them to engineers recognises that code review, tests and environment separation are not the same on a trading desk. Greek banks and fintechs should publish approved-tool lists by role and forbid unsupervised code execution against production. Over-strict rules without sanctioned alternatives can push staff toward shadow use of personal accounts. Balance means approved sandboxes, prompt logging where lawful, and training on what must never be pasted into external models. Risk committees should treat coding agents as a new class of third-party software with vendor assessment. Goldman’s move will pressure peer firms to write down policies before a leak or faulty automation forces the issue. Clear escalation paths for business users who still need automation — through platform teams rather than desktop agents — will matter as much as the ban itself.

Read the original on Bloomberg Law

Related service: Custom software

Related stories