Goldman Sachs Restricts Traders’ Use of Claude Code: An Internal Retreat by an Anthropic Underwriter

In October 2026, Goldman Sachs quietly updated an internal guideline: Anthropic’s Claude Code and Microsoft’s GitHub would be available only to engineers. The directive drew a clear permission boundary—about 12,000 engineers retained access, while roughly 35,000 traders, bankers, and analysts were cut off. Bloomberg reported that the change has already caused internal dissatisfaction, with some traders frustrated by it.

After Bloomberg disclosed the news in early November, it quickly drew attention, partly because of the delicate timing: in the same month the restriction took effect, Anthropic filed for an IPO on October 26—and Goldman Sachs was the lead underwriter for the listing.

From “Everyone Codes” to “Engineers Only” in Just Three Months

In July, Goldman Sachs CIO Marco Argenti was still publicly praising the firm’s internal “vibe coding” experiment. He described to Bloomberg a smoothly operating collaboration model: a trader or asset manager has an idea, quickly builds a prototype with AI tools, and hands it to engineers for production-grade refinement. Argenti saw it as an effective way to let the people closest to the problem explore first and shorten iteration cycles.

This logic is persuasive in a lab context. But it concealed a structural risk: thousands of employees without engineering backgrounds were running AI-generated and AI-executed code on their own work machines—the same machines that held client trading information and position data.

Three months later, Goldman chose to end the experiment. Goldman spokesperson Nick Carcaterra said in a statement that after evaluating factors including “productivity and risk management,” the firm decided to “give the tools to the people best suited to use them over the long term.” The other side of that sentence: those who had been using the tools were deemed not to be the “best suited” candidates.

What Was Cut Off Was Not AI, but the Authority to Execute Code Autonomously

The key to understanding the restriction is distinguishing two types of AI use cases. Goldman’s internal AI assistant—a self-built product integrating models from OpenAI, Anthropic, and Alphabet—remains open to the entire firm. Goldman’s partnership with Anthropic has not been interrupted either: Anthropic engineers previously worked on-site at Goldman for six months, helping build Claude agents for trade settlement and client onboarding processes; those systems run in the background and are managed by engineering teams.

What was taken back was third-party tools such as Claude Code and GitHub that can autonomously write and execute code—in other words, the direct execution authority of agentic AI in the hands of nontechnical staff.

Goldman’s exemption of a special group called “Strats” from the restriction offers supporting evidence. Strats is a role type unique to Goldman, spanning trading, quantitative research, and software engineering; they are essentially finance professionals with full technical backgrounds. They kept access to the tools, while traders and bankers with purely business backgrounds lost it. The boundary was not drawn by whether someone touches AI, but by whether someone has the ability to judge the safety of AI-generated code.

The Hong Kong Precedent and the Logic of a Global Pullback

This is not the first time Goldman has drawn a boundary around Claude. In April 2026, Goldman removed Claude from the internal AI platform for its Hong Kong bankers. At the time, the Financial Times reported that Goldman had interpreted its contract terms with Anthropic strictly, and Anthropic said Claude had never been officially supported in Hong Kong. ChatGPT and Gemini remained available locally. Two months later, JPMorgan also removed Claude from its approved list in Hong Kong.

The Hong Kong restriction was based on contractual terms and regional compliance; the new October restriction is based on job type. Both moves point to the same core proposition: inside financial institutions, access to AI tools is being systematically tightened, and each tightening comes with an explicit rationale—compliance, risk, or the fit between tools and roles.

This trend is not unique to Goldman. Dealroom reported that JPMorgan has also pushed employees to match lower-cost models to routine tasks and imposed a blanket block on third-party AI agents accessing customer accounts. JPMorgan CEO Jamie Dimon explicitly mentioned the risk of “rogue agents” on the third-quarter earnings call and noted that when customers authorize external AI tools, liability shifts. Citigroup, Bank of America, and Wells Fargo have also taken similar measures. Wall Street is moving from open experimentation to controlled deployment faster than outsiders expected.

The Awkward Narrative Created by IPO Timing

The Goldman restriction on Claude Code came exactly as Anthropic was racing toward an IPO, a coincidence that provides the most tension-filled angle for outside interpretation. As Anthropic’s lead underwriter, Goldman should in theory be one of the strongest endorsers of the AI company’s value. In reality, it is also the decision-maker removing Claude Code from the workstations of 35,000 non-engineering employees.

This is not a contradiction. Goldman’s retreat targets a specific use case—nontechnical staff directly operating AI agents that can execute code—not a rejection of Anthropic’s overall capabilities. Goldman still uses Claude deeply: back-office agents, engineer tools, and a multiyear technology partnership with Anthropic. What retreated was the permission boundary, not technical trust itself.

But this distinction is too subtle for the market narrative. What outsiders see is this: a staunch Anthropic partner and IPO underwriter removed Claude Code from its own traders’ computers on the eve of the listing. The symbolic meaning of this signal far exceeds its practical meaning.

The Real Speed of Enterprise AI Deployment

Goldman’s case reveals a cycle systematically underestimated in AI industry promotion: enterprise adoption of AI tools is not linear expansion but a wave-like rhythm of “open experimentation → risk assessment → permission contraction → controlled expansion.”

It took Goldman only a few months to go from allowing all employees to use the tools to restricting them to engineers. The trigger was not tool failure but tools running at the wrong boundary—nontechnical staff executing unreviewed AI-generated code in sensitive data environments. This was a foreseeable risk, but in the optimistic “vibe coding” narrative it was not formally priced in until after the experiment ended.

This pattern is especially typical in finance. A bank’s core constraint is not compute or model capability but the integrity of data governance and code review pipelines. A piece of code generated by AI, modified by nontechnical staff, and executed directly on a production machine—no matter how small—is a potential compliance hole. Goldman’s decision is an institutional expression of that judgment.

Over a longer cycle, this tightening may not be the end point. Goldman’s “engineers first” model is logically compatible with the “prototype → engineering” process it described in July: let people with technical judgment fully use the tools first and build a repeatable safety framework, then open them to a broader set of business staff in a controlled way. The wave-like AI rollout Argenti described—developers first, others following—may still be evolving slowly after the October restriction, just more cautiously than expected.

But one thing is already clear: inside financial institutions, enterprise deployment of AI coding tools is not a procurement decision but a code governance issue. Who has the right to run unreviewed code, on which machines, and with what level of data—the answers to these questions will determine the real pace of agentic AI expansion in finance, not AI vendors’ IPO valuations.

Sources: - [Goldman Sachs Restricts AI Coding Tools to Engineers](https://finance.yahoo.com/technology/ai/articles/goldman-sachs-restricts-ai-coding-150508601.html) - [Goldman Restricts Traders and Bankers From Using Some AI Tools - Bloomberg](https://www.bloomberg.com/news/articles/2026-11-02/goldman-restricts-traders-and-bankers-from-using-some-ai-tools) - [Goldman Sachs limits AI coding tools for traders and bankers](https://qz.com/goldman-sachs-restricts-ai-coding-tools-traders-bankers-110226) - [Goldman Sachs pulls AI coding tools from traders and bankers - The Decoder](https://the-decoder.com/goldman-sachs-pulls-ai-coding-tools-from-traders-and-bankers/) - [Anthropic IPO plans - CNBC](https://www.cnbc.com/2026/10/26/anthropic-ipo-live-updates-ai-s-1-filing.html) - [JPMorgan blocks outside AI agents from customer accounts](https://qz.com/jpmorgan-blocking-outside-ai-agents-customer-accounts-101326)