Goldman Sachs Restricts AI Coding Tools from Traders

Goldman Sachs has restricted the use of artificial intelligence coding tools by traders and banking operations staff. The tools in question include Claude Code and GitHub Copilot, and the company is transitioning to allow primarily engineers to use them going forward. The company cites 'productivity and risk management' as the reasons.

This policy shift is noteworthy because of the stark contrast with statements made just a few months earlier. Marco Argentti, the company's Chief Information Officer, was still publicly praising 'zero-shot coding' by non-technical staff as of July 2025, a technique that allows users without programming knowledge to generate code automatically by giving natural language instructions to artificial intelligence. Within a short period after that, the policy reversed direction entirely.

Zero-shot coding refers to a method of use where people without programming knowledge can have artificial intelligence automatically generate code simply by giving instructions in Korean or English. As generative artificial intelligence has become more widespread, this usage pattern has expanded beyond engineers to other professions. However, experts have also pointed out the challenge that it is difficult for specialists to verify the quality and safety of generated code. Especially in highly regulated environments like financial institutions, this emerges as a risk issue.

The 'risk management' rationale presented by Goldman Sachs plainly demonstrates this concern. If traders or bank employees write code using generative artificial intelligence, and that code is incorporated into actual trading or operational systems, there could be risks of unintended behavior or information leakage. By limiting usage to engineers, the company appears to be aiming to ensure thorough code review and quality control.

This move can be positioned as an example symbolizing the transition from the 'expansion phase' to the 'management phase' of artificial intelligence adoption. The pattern of companies first deploying generative artificial intelligence tools broadly and then reviewing usage scope and permissions is seen not only in the financial industry but also in other sectors. It can be argued that such reviews are particularly likely to occur in applications like code generation, where errors carry direct risks to business operations.

What will be watched in the future is whether the scope of restrictions expands further or moves toward reopening under specific conditions. As the phrase 'balancing productivity and risk management' suggests, Goldman Sachs itself is not rejecting artificial intelligence adoption. The approach appears to be one of continuing to use the technology within engineer-only limits while accumulating knowledge about safe operations. Since large financial institutions' responses often influence standard-setting for the entire industry, it will be necessary to watch their moves alongside those of other banks and regulators.