Goldman Sachs Just Pulled AI Coding Tools From Most of Its Staff. Here's Why That's Smart, Not Scared

What happened

Goldman Sachs spent the better part of a year encouraging traders and bankers to experiment with AI coding tools. In July, the bank's Chief Information Officer was publicly praising the practice, pointing to non-technical staff building quick prototypes with AI and handing them off to engineers.

That changed in October. Goldman updated its internal guidance and restricted access to Anthropic's Claude Code and Microsoft's GitHub Copilot to its roughly 12,000 engineers, pulling the tools from bankers and traders who had used them for months. The reason given was blunt: "productivity and risk management."

This wasn't a ban on AI. It was a narrowing of who gets to use which tools and for what. Traders and bankers had been using the tools to monitor client trading flows, simulate market moves and positions, and draft text and visuals. That's exactly the kind of work that sounds efficient until something goes wrong and nobody who built it can explain how.

Why this is the real story

The headline here isn't that a bank got nervous about AI. It's that one of the most sophisticated financial institutions in the world ran the experiment, got real data, and concluded that open access without a plan was the wrong move. Wall Street is shifting from open-ended AI trials to controlled deployment.

Goldman isn't alone in thinking hard about this. Bank of America's CEO has been vocal about the need for caution, specifically telling investment bankers to carefully check the accuracy of AI-generated pitchbooks and spreadsheets rather than hand off that judgment entirely. Inside Goldman itself, a partner leading one of the bank's flagship AI projects warned that heavy reliance on these tools risks dulling the critical thinking skills of the next generation of bankers.

That's the tension every large company is going to run into. AI tools make people faster. They don't automatically make people better at knowing when the output is wrong.

What this means for your company

If you're rolling out AI across departments, this is the cautionary tale worth reading closely. A few things stand out:

• Access without a framework creates risk, not just speed. Giving everyone the same tool with no guardrails means everyone is making their own judgment calls about what's safe to automate. That's fine until it isn't.
• Who uses a tool matters as much as what the tool does. Engineers at Goldman kept their access because they're equipped to review and own the output. Traders and bankers lost theirs because the review burden outgrew what made sense.
• "Experimentation phase" needs an end date. Goldman let this run for months before tightening it. Large companies rarely plan for that transition in advance, which means it happens reactively instead of on purpose.
• Governance isn't a brake on adoption. It's what makes adoption sustainable. The companies that figure out role based access, review processes, and clear ownership now will move faster later, not slower.

This is the exact gap an AI Audit is built to close. Before you hand a department a new tool, you need to know which tasks actually benefit from AI, who should have access, what needs a human checking the work, and how you'll know if it's actually saving time or just creating new risk. Custom agents and Copilot tools built around real workflows, not generic access for everyone, are what let a large company scale AI without ending up in a Goldman style reversal six months in.

The takeaway

Giving every department the same AI tool and hoping for the best isn't a strategy. Goldman Sachs just showed what it looks like when a major institution learns that the hard way and corrects course. Get the rules right before the rollout, not after.