What jumps out to me is not that Goldman tightened access. It’s that they apparently let traders and bankers build useful little internal tools with Claude Code and then, after a review, pulled the rug back out. That sequence says a lot about how enterprises still don’t really know what to do with agentic coding once it escapes the engineering department.
I don’t think the “risk and productivity” explanation is bogus. It actually makes sense. A trader script that watches flows or models positions is not the same thing as a toy app sitting in a sandbox. If the tool can read, write, and run code with the user’s privileges, then the blast radius is whatever that employee can touch. On a bank desk, that’s not a small concern. Information barriers alone would make me nervous. The article is right about that.
What I’m less convinced by is the idea that the clean fix is simply “engineers only.” That feels tidy in a compliance memo, but it also sounds like a way of pretending the real issue is user class rather than process ownership. If a banker built something genuinely useful, the bank should probably have a way to absorb it into engineering, not just shut it down and call it a day. Otherwise you get the worst of both worlds: local experimentation gets encouraged, then punished once it becomes visible.
The timing is what makes this awkward. Goldman’s CIO was reportedly praising trader-built AI tools in July, and now the same institution is tightening the screws while it is also a lead underwriter on Anthropic’s IPO. That doesn’t mean hypocrisy in some dramatic sense, but it does make the messaging look sloppy. If you want people to experiment, you need to be clear about what happens when the experiment works. Otherwise everyone learns the same lesson: don’t build anything too useful.
I also think this is a good reminder that “AI coding tool” is not one thing. There’s a big difference between an assistant that helps an engineer draft code and an agent that can autonomously write and execute programs on a trader’s workstation. People keep collapsing those into the same bucket because the demos look similar. In production, they are not similar at all.
The bit I’d watch next is whether other firms start drawing the line by role, or by workflow ownership. Role-based restrictions are easy to explain, but they’re a blunt instrument. Workflow-based controls are messier, yet probably more honest. If a non-engineer wants to use Claude Code to build something, someone in engineering should own the artifact before it touches anything real. That seems like the actual lesson here.
Reference: Goldman Sachs Is Pulling Claude Code From Its Bankers and Traders | State of AI
