On November 2, Bloomberg reported Goldman Sachs had pulled Claude Code and GitHub's coding tools away from most of its bankers and traders. The headlines wrote themselves. "Goldman bans AI." "Wall Street turns on vibe coding." Some traders were frustrated, and said so.
I read the story twice. I don't see a ban. I see a leadership team doing the one thing most companies never get around to... ending an experiment with a decision.

What Goldman did
Here are the facts, as reported by State of AI and AI Chat Daily, both citing Bloomberg:
- Internal guidance updated in October limits Claude Code and GitHub's coding tools to engineers. Goldman has about 12,000 of them, out of roughly 47,000 staff.
- Strats, the people who sit across trading, quant research and engineering, keep access.
- Traders had used the tools to monitor client trading flows, model market moves and positions, and build written and graphic material.
- The restriction covers third-party agents able to write and run code by themselves. The in-house GS AI Assistant is still rolling out to about 46,000 employees.
- In July, CIO Marco Argenti praised "vibe coding" by non-technical staff, who built tools and passed them to engineers to improve.
A Goldman spokesperson, Nick Carcaterra, said the bank makes AI tools available "to determine the best use cases across different parts of the firm." ANOTHER News reported him saying the bank put the tools in the hands of those best suited to use them after a period of experimentation.
So Goldman ran a trial. It looked at the results. It drew a line by role. Everybody still gets AI. Only engineers get the agent with its hands on the keyboard.
You might hate where they drew the line. But they drew one. Most companies I talk to haven't.
Your pilot is still running
I've sat in a lot of leadership meetings over the years. Here's a pattern I see again and again. Someone launches a pilot. People get access. Usage grows. Nobody writes down what success looks like, so nobody ever calls it. Six months later the "pilot" is the production environment, and nobody decided anything.
AI coding tools make this worse, because the output doesn't stay in a chat window. It runs. It touches data. It ends up in a shared drive, a scheduled job, a dashboard the sales team checks every morning.
I wrote earlier this year about how banning AI won't stop your team, because people route around a flat "no." I still believe it. Goldman's move is different. They didn't say no. They said "this tool, these people, for this reason." The traders still have an AI assistant. They lost the part where a non-engineer ships running code next to client data.
The real question is ownership
Strip away the AI headlines and this is an old fight. Who owns the software once it exists?
Argenti's July model sounds great on a slide. The trader closest to the problem builds version one. An engineer picks it up and makes it solid. Fast learning, no waiting in a queue for engineering time.
Now think about what happens at scale. Thousands of traders build tools. Engineers don't have the hours to pick up all of them. So most of those tools never get handed over. They keep running on a trader's machine, written by someone with no training in testing, access control or audit trails, on the same machine they use for client information. The trader gets promoted, moves desks, or leaves. The tool keeps running. Nobody owns it.

Banks have lived through this already. Not with AI. With Excel.
When JPMorgan lost over $6 billion in the 2012 "London Whale" trades, the investigations found a pile of problems. One of them, as summarized on Wikipedia's account of the loss, was "woefully deficient accounting controls in the CIO, including spreadsheet miscalculations" which caused large valuation errors. The spreadsheet didn't cause the bad trades. It did help hide how bad they were.
Every bank I know of has a team whose job is to hunt down spreadsheets running critical processes nobody signed up to support. Vibe coding is end-user computing with a faster engine. If you thought spreadsheet sprawl was bad, wait until the spreadsheet writes its own macros and calls an API.
So when Goldman says "engineers only," I hear "if code runs here, someone trained to own it has to own it." It's an ownership rule. It isn't an anti-AI rule.
Where Goldman got it wrong
I don't think Goldman nailed this. Two things bother me.
First, the whiplash. In July the CIO cheered non-technical staff building tools. By October the guidance took the tools away. The traders who leaned in hardest got burned the worst. If you're a leader, remember this one. The people who trust your enthusiasm first are the ones you hurt most when you reverse without explanation. Next time you ask them to try something new, they'll wait to see if you mean it.
Second, the frustration leaked to Bloomberg. When your people take their complaints to the press instead of to you, your communication failed before your policy did. I've spent years researching bad bosses. 99.5% of the people we surveyed said they've had one or more types of bad boss. A lot of what makes a boss "bad" isn't cruelty. It's decisions dropped on people with no reasoning attached. "Here's the new rule" lands differently from "here's what we learned, here's the risk, here's what you get instead."
I'm not sure about this part: none of the reports I found describe how Goldman explained the change to traders. The frustration might come from the decision itself, not how it landed. I'm reading between the lines.
What to steal for your company
You don't run a bank. Your regulators might be lighter. The ownership problem is the same. Here's what I'd take from Goldman, and fix along the way.
1. Put an end date on every AI experiment
Write down the end date and the question you're answering before you hand out access. "By March 31, we decide who keeps Claude Code and why." No end date means no decision, and no decision means the pilot becomes policy by accident.
2. Draw the line by what the tool does, not by what it is
Goldman didn't pull AI. It pulled agents able to write and run code on their own. A chat assistant drafting an email carries a different risk than an agent with shell access. Sort your tools by blast radius, then decide who gets which.
3. Give every running thing an owner
Any script, agent or automation touching production data gets a named owner who knows how it works and answers the phone when it breaks. If nobody will sign up, it doesn't run. This one rule kills most shadow tooling without a ban.

4. Build the handoff before you need it
Argenti's idea of traders prototyping and engineers hardening was a good idea with no capacity behind it. If you want non-engineers building, budget engineering time to receive their work. Otherwise you're asking for prototypes you have no plan to support.
5. Explain the reversal like an adult
If you pull a tool back, say why. Say what you learned. Say what people get instead and when the decision gets reviewed. Your early adopters deserve the reasoning. They're the ones who'll test your next idea.
Governance isn't the opposite of speed
The loudest voices in AI right now treat every restriction as fear. Some restrictions are fear. Goldman's looks more like a team deciding where speed belongs and where ownership has to come first.
The real failure mode isn't a bank taking a coding agent away from traders. It's the company down the road with 400 vibe-coded scripts, no owners, no inventory, and a CEO who thinks the pilot is still a pilot.
So here's my question for you. How many tools are running in your company right now with nobody's name on them? If you don't know the number, start counting this week.