Goldman Sachs has taken advanced AI coding tools away from most of its bankers and traders. Under internal guidance updated in October and reported by Bloomberg on November 2, Anthropic’s Claude Code and Microsoft’s GitHub Copilot are now limited to the bank’s roughly 12,000 engineers, out of a workforce of about 47,000.
Don’t mistake this for a firm turning its back on AI. Goldman still runs its own AI assistant for every employee, and it still uses Claude heavily behind the scenes. What changed is who gets to hand a coding agent a task and let it run.
A Goldman spokesperson, Nick Carcaterra, framed the shift as the natural end of a trial period. He said the bank makes certain AI tools available experimentally, then puts them “in the hands of those best suited to use them over the long term,” citing productivity and risk management as factors.
The experiment is over.
From vibe coding to engineers only
Just months ago, Goldman was encouraging the opposite. In July 2026, chief information officer Marco Argenti publicly promoted vibe coding for non-technical employees: the workflow where a staffer turns an idea into a fast prototype with an AI tool and then hands the result to an engineer for finishing.
Bankers and traders took up the offer. According to Bloomberg’s reporting, they used Claude Code and GitHub tools to monitor trading flows, build models of how market moves would hit their positions, and draft written and graphic materials.
October’s guidance reversed that. Most bankers and traders lost access, and reports say the change drew complaints inside the firm. One group kept its access: the strats, hybrid staff who mix trading, quantitative research, and software engineering. The line Goldman drew runs through job function, not seniority.
What stays in place
Nothing about the restriction touches Goldman’s firmwide AI setup. The GS AI Assistant, launched to all staff in June 2025, stays available to everyone, running behind the bank’s firewall on models from OpenAI, Alphabet, and Anthropic. Only the third-party agents that write and run code lost their pass.
Goldman also keeps building with Claude in the back office. Anthropic engineers spent six months at the bank helping build Claude-based agents for trade accounting and client onboarding, per Bloomberg. Those agents are run by Goldman’s engineering teams, which is exactly the arrangement the new guidance preserves.
Nor is this Goldman’s first narrowing of AI access. In April 2026, the bank removed Claude for bankers in Hong Kong after reading its Anthropic contract strictly, the Financial Times reported. Anthropic replied that Claude had never been officially supported there. That restriction was about location; October’s is about job role, and it is far broader.
Why banks are nervous about agents
Goldman has not said what its evaluation found, and that silence matters. Reporting does not establish whether non-technical users were less productive, created identifiable risk, or simply ran up a token bill nobody wanted to defend. But the reason banks keep coming back to is simple: an AI coding agent is less like a calculator and more like a junior employee with a login.
An agent works with the same system access as the person who prompted it, and its output can run before an engineer reviews it. On a trading desk, that difference is the whole story. Under US financial rules, the bank answers for supervising and recording what happens on its systems, not the software vendor. Letting thousands of non-engineers loose with agents that can touch production data is a compliance headache.
Goldman is not alone in that worry. Mitsubishi UFJ Financial Group, Japan’s largest bank, concluded that Claude Code’s capabilities could bypass its information-access restrictions and security constraints, Risk.net reported in July, and built its own version of the tool instead. Stu Taylor, who oversees AI for MUFG’s global markets business outside Japan, said all of the bank’s AI tools use a human-in-the-loop approach, with a person reviewing and approving actions. General-purpose agentic tools, he said, make those controls hard to maintain.
Why this matters
Goldman is the most informed buyer in this story, and it just bought the narrower version. Its engineers worked alongside Anthropic’s for months, and Goldman, with Morgan Stanley, is a lead underwriter on Anthropic’s IPO, which filed its S-1 on October 26. Soon, Goldman’s bankers will market shares in a company whose coding agent most of them can no longer use.
That irony aside, the decision sets a template other big firms will study: try everything, measure what happens, then fence the tool to the people trained to catch its mistakes. For workers, the message is blunt. AI coding skill is becoming a job qualification, not a perk handed to everyone with a laptop. For AI companies, the message is blunter still. Enterprise-wide deployment is not the same as enterprise-wide access, and the gap between the two is where their next sales pitch will be won or lost.
Frequently asked questions
Why did Goldman Sachs restrict AI coding tools to engineers?
Goldman says it makes certain AI tools available experimentally, then puts them in the hands of those best suited to use them over the long term, citing productivity and risk management. The bank has not disclosed what its evaluation actually found.
Which AI coding tools did Goldman Sachs restrict?
Anthropic’s Claude Code and Microsoft’s GitHub Copilot. Both are now limited to Goldman’s roughly 12,000 engineers, while bankers and traders have lost access.
Do any non-engineers at Goldman still have access?
Yes. So-called strats, staff who blend trading, quantitative research, and software engineering, are exempt from the restriction.
Has Goldman banned AI tools for most employees?
No. Its firmwide GS AI Assistant, which runs on models from OpenAI, Alphabet, and Anthropic, remains available to all staff. Only the advanced coding agents were restricted.
Sources: Bloomberg News, The Decoder, Financial Advisor Magazine, Financial Times, Risk.net