Goldman Sachs pulled Claude Code and GitHub's AI tools from traders and bankers. Give business builders a path to production instead

Goldman Sachs gave non-engineers AI coding tools, then took them back. The lesson is about the path to production, not the tools.

What happened

Bloomberg reported on November 2 that Goldman Sachs has limited advanced AI coding tools, including Anthropic's Claude Code and Microsoft's GitHub offerings, to its engineers, based on internal guidance from October. Traders and bankers had spent months using them to monitor client trading flows and create written and graphic materials. According to The Decoder, the rollback is causing frustration inside the firm.

The reversal is sharp. In July, CIO Marco Argenti praised vibe coding by non-technical staff: someone in trading or asset management would build a quick prototype with AI, then hand it to one of the firm's 12,000 engineers to scale for production. A Goldman spokesperson cited a variety of factors including productivity and risk management, and said the experimentation phase is over. PYMNTS notes this comes as banks spend millions on AI tools while weighing cost, productivity and cybersecurity risk.

My take

I see the same pattern in smaller companies, just without a press release. A sales manager builds a Zapier or n8n flow with an AI assistant, it works, and six months later it is quietly running a core process with nobody owning it.

The fix is not banning business users. They know the problem better than anyone. The fix is a clear handoff:

1. Business users can prototype freely, but only against test data or a sandbox CRM.
2. Anything that touches client records, money or outbound messages goes through a review before it runs live.
3. Every production automation has a named owner, a run log and an off switch.
4. Credentials for live systems sit with the builder, not inside a personal AI tool.

That keeps the speed of letting the person closest to the work sketch the solution, while the risk stays where someone can see it. If your team already has a few homegrown automations, start by listing them. You will usually find at least one doing something important with no owner.