Goldman Sachs pulls Claude Code, the AI that writes code, from its traders and bankers

In July, Marco Argenti, Goldman Sachs's chief information officer, was still praising "vibe coding." The idea: traders and managers with no technical training build a prototype themselves with AI, then hand it to one of the bank's 12,000 engineers to make it robust. Four months later, the bank is backing off.

According to Bloomberg, which reviewed internal guidelines updated in October, Goldman Sachs now reserves Claude Code, Anthropic's coding assistant, and the coding tools from GitHub, a Microsoft subsidiary, for its engineers alone. Most other employees, including traders and bankers, are losing access. The news was published on November 2.

Why the bank is closing the door

Goldman Sachs is sparing with explanations. The bank says the decision follows a period of experimentation, during which it assessed several criteria, including productivity and risk management. It says it now reserves these tools for those best placed to use them over the long term. "We are making certain AI tools available to our teams to determine the best uses across the different parts of the firm," said spokesperson Nick Carcaterra, quoted by the site PYMNTS.

The decision has not gone unnoticed internally. According to Bloomberg, some traders are frustrated: for months, they had been using these tools to monitor their clients' order flows or to produce documents and charts. According to a write-up of the article, the "strats," the team that combines programming, quantitative research and trading, are exempt from the rule.

Caution shared by major banks

Goldman Sachs is not an isolated case. Its competitor JPMorgan has already banned outside AI assistants from acting on its clients' accounts, and its chief executive, Jamie Dimon, has warned against "uncontrolled" AI agents. Bank of America limits agents' access to repositories, and Citigroup lets them neither log in nor execute transactions. The difference here is that Goldman is not targeting outside tools: it is restricting tools it had deployed itself.

What a company can take from this

Letting non-developers produce code that touches company data raises concrete questions. Who reviews that code, who maintains it, and what data passes through the assistant? Before rolling these tools out widely, a few precautions are in order:

• define which job functions have access, and for which uses;
• forbid putting a script into production without review by a developer;
• check what data the assistants can read and send outside the company;
• measure the real gain rather than the number of licenses opened.

The bank does not say whether it has seen an incident, or how many employees are affected. The question remains open: is this a pause, or the end, in finance, of the dream where everyone programs?