Goldman's Claude Code rollback is about governance, not AI doubt
Goldman Sachs cutting traders and bankers off from coding agents isn't an AI retreat. It's a preview of how regulated firms will buy: controls first, hype never.
AI-generated image
Is it a bad sign for AI coding tools when a bank that was boasting about "vibe coding" in July bars most of its traders and bankers from them by October? We don't think so. It's a sign the tools work well enough to be dangerous in the wrong hands.
Goldman didn't ban AI, it fenced it in
According to Bloomberg's reporting as summarized by State of AI, Goldman Sachs updated its internal guidance in October. Anthropic's Claude Code and Microsoft's GitHub tools are now limited to the bank's engineers, about 12,000 people, after a review of productivity and risk concerns. Strats, who straddle trading, quant research and engineering, keep access.
Some traders are reportedly frustrated. They had been using the tools to monitor client flows, model positions and put together written and graphic material. The bank's in-house assistant, which draws on models from several companies, remains available firmwide. The restriction targets third-party agents that can write and run code on their own.
The tool isn't the risk. Unreviewed code is
The reasoning, as we read it, is simple. An engineer's code usually goes through review, testing and deployment controls. A script a trader builds to watch client flows may never see any of that, and it runs on the same machines used to handle client information. Add the information barriers banks must maintain between deal teams and trading desks, and you have a compliance problem no model upgrade fixes.
That's why we think the "AI is overhyped" reading misses the point. Nothing in the reporting suggests Goldman believes the agents don't work. The problem, in our view, is that they work for people who can't audit the output. A trader who can't tell how generated code handles edge cases is a loss waiting to happen.
Cost may also be a factor. Agent usage is billed by the token, and State of AI notes that other companies have watched spending climb. That's context rather than a confirmed reason here: The Decoder says the bank itself points to "productivity and risk management".
Vendors will win on audit trails, not benchmarks
Here's our thesis: in regulated industries, procurement will be decided by governance. Can the firm see what the agent read, what it ran and who approved it? Can access be scoped by role, desk and information barrier? If the answer is "trust us," the tool ends up confined to the one group trained to review it.
Goldman's own pattern points the same way. It built its internal AI platform around controls, and the back-office Claude-based agents for trade accounting and client onboarding, built with Anthropic engineers, are run by Goldman's engineering teams. The bank doesn't look anti-Anthropic to us. It looks like it wants AI inside a perimeter it controls. Anthropic's vendor lesson is uncomfortable: a better model doesn't buy you the trading floor, and an agent that typically has the user's full access is a feature on a laptop and a liability in a bank.
The timing is awkward, too. Goldman is a lead underwriter on Anthropic's IPO, per the S-1 cited in the report, so its bankers will soon pitch shares of a company whose coding agent most of them can't use. We'd call that irony, not scandal. Still, investors should probably consider that enterprise demand may be narrower than the headline adoption numbers suggest.
Right rule, clumsy rollout
We agree with State of AI on one point: Goldman is probably right to restrict, but handled it poorly. Its CIO promoted staff-built tools in July. Pulling access through a guidance update, with no replacement mentioned in the reporting, tells the people who found real uses that they were the problem.
The smarter move is to treat those frustrated traders as a free product backlog. Their workflows, such as flow monitoring and position modeling, are exactly what engineering could build properly, with review and barriers baked in. Shadow AI doesn't disappear when you ban it. It goes underground.
Treat it as a checklist, not a ban
If you run technology at a regulated company, don't read this as a reason to ban agents. Read it as a checklist: role-based access, logged actions, mandatory review before anything touches client data, and a sanctioned path for non-engineers to request tools. If you sell these tools, ship those controls before the next feature. The firms that gate by governance will keep buying AI. Vendors that can't show their work will find the door narrowing to the engineering floor.
No topics yet: start the first one.
More stories
The DRAM collusion suit survived, and AI can't be the alibi
A judge refused to dismiss the case against Samsung, SK Hynix and Micron. The AI boom explains some of what happened to memory, but it shouldn't shield the oligopoly from scrutiny.
Microsoft's outage shows resilience is a business decision
A global Outlook, Teams and Copilot failure is a reminder that bundling mail, meetings and AI under one vendor is a bet, and the board should own it.
United's AI agent ban is about the customer, not safety
United says it's protecting travelers, but its CEO's "another middleman" line gives away the real fight: who owns the relationship. Expect licensing, not a permanent wall.
CoreWeave's 16% yield is the market sending AI a bill
When a neocloud's debt yields more than 16%, the AI boom stops being a story about demand and becomes one about who can afford to keep borrowing.



