Goldman Sachs bans staff from advanced AI coding tools over risk

What happened

Goldman Sachs has banned traders and other employees from using advanced AI tools for coding, citing productivity and risk concerns, according to Bloomberg. The restriction was reported by Finextra and The Decoder, both of which drew on the Bloomberg account.

The move places a clear limit on how one of Wall Street's largest banks allows its staff to use generative AI in day-to-day technical work. The reporting frames the decision as a response to two linked pressures: whether the tools genuinely raise output, and whether they introduce risks that the firm is not prepared to accept.

Why it matters

AI coding assistants have spread quickly across software teams, and many organisations have treated them as low-risk productivity boosts. A decision of this kind suggests that in regulated, high-stakes environments such as trading, the calculation is more complicated. Code written or modified outside approved controls can affect trading systems, data handling and compliance, so the question for leaders becomes where the tools can be used safely, not simply whether they are allowed.

For experience and operating-model leaders, the story is also about the internal employee experience. When staff lose access to tools they have come to rely on, the gap between perceived and measured productivity becomes visible. Organisations that deploy AI without clear guardrails, or that withdraw it abruptly, risk eroding trust with the people who must work within the new rules.

The Renascence take

Most coverage will read this as a story about AI being too risky. We see it as a story about governance design: the decision shows that access to AI is itself a service that needs rules, owners and feedback loops. Behavioural economics would predict that employees will look for workarounds wherever a tool is valuable and its boundaries are unclear, so bans without alternatives tend to shift behaviour underground rather than remove it.

Blanket prohibition is a blunt instrument, and it rarely removes the demand behind it. The more useful question is what the tool is for, which tasks it can safely support, and what checks must sit around it. A customer-obsessed operator should map where staff are already using AI, design sanctioned routes that are faster than the workaround, and measure both productivity and risk before deciding what to permit.