Goldman Sachs has restricted Anthropic's Claude Code and Microsoft's GitHub to its engineers, cutting off most traders and bankers who had been using the AI coding tools at work, according to a Bloomberg News report that resurfaced this weekend as one of the most-discussed stories in the technology community.
Internal guidance updated in October limits the tools to the bank's engineers, Bloomberg reported, following a review of productivity and risk concerns. The change means employees on trading desks and in the investment bank — roles that had enthusiastically adopted AI coding assistants — can no longer use them for daily work. For more context on this story, see our ongoing AI news.
Who still has access
Not everyone at the bank is locked out. According to a second person familiar with the matter cited in reporting on the change, "strats" — Goldman employees who work across trading, quantitative research and software engineering — still have access to the tools.
The restriction has frustrated some traders, a person familiar with the matter told Bloomberg. They had been using the coding agents to monitor client trading flows, model market moves and positions, and put together written and graphic material for clients and internal use.
A Goldman spokesperson, Nick Carcaterra, told the news service that the bank makes certain AI tools available to staff "to determine the best use cases" across the firm — a framing that suggests the restriction reflects an ongoing evaluation rather than a permanent ban.
A fast reversal for a bank that championed 'vibe coding'
The pullback is a striking reversal for a bank whose own technology leadership was, as recently as July, celebrating exactly the kind of use the new guidance curtails. Chief information officer Marco Argenti told Bloomberg in July that traders and bankers had been building their own tools with AI — a practice often called "vibe coding" — and praised the ways non-technical employees were putting coding agents to work.
That enthusiasm is now constrained by the practical realities of deploying cutting-edge AI tools inside a regulated bank. Coding agents can process large volumes of sensitive material, and financial institutions face strict obligations around client data, record-keeping and supervision. The October review that preceded the restriction reportedly weighed productivity gains against those risk concerns.
Anthropic's IPO complicates the picture
The timing is awkward for another reason: Goldman Sachs is a lead underwriter on Anthropic's initial public offering, the hotly anticipated listing that the ChatGPT rival is pursuing at a reported valuation target of roughly $1.5 trillion. Anthropic's IPO has drawn intense attention amid November's AI stock sell-off, with Bloomberg reporting that some prospective buyers value the company closer to $1.1 trillion.
The bank has also been here before with Anthropic's regional restrictions. In April, Goldman staff in Hong Kong lost access to Claude after Anthropic cut off availability in the territory, a change the AI company attributed to its licensing terms amid tightening US export rules on advanced AI models in China. That episode showed how quickly access to frontier AI tools can vanish for operational reasons; the new internal guidance shows the bank itself now drawing the boundaries.
Part of a broader enterprise recalibration
Goldman is not alone in tightening the taps. Computerworld reported this month that Microsoft has extended undisclosed usage restrictions on Copilot Chat for enterprise users, adding to a growing list of caps and quota changes that companies have quietly imposed on AI tool usage this year. As enterprise AI bills balloon, corporations are moving from experimentation to hard-nosed decisions about who gets access to expensive frontier models and for what.
The pattern marks a new phase in enterprise AI adoption. In 2024 and 2025, banks and technology firms competed to hand powerful AI tools to as many employees as possible, and productivity studies fueled the frenzy. In 2026, the questions have shifted to governance: which roles genuinely benefit, what data the tools touch, and whether the productivity gains justify the cost and compliance exposure.
Why banks treat coding agents differently
The distinction between engineers and everyone else is deliberate. A coding agent wired into an engineer's workflow sits close to systems the bank already treats as governed: code repositories, review processes, audit trails. A trader spinning up an ad-hoc script that ingests client flow data lives outside most of those controls — exactly the kind of shadow IT that compliance departments spend their lives hunting down. The fact that strats keep their access suggests Goldman is not abandoning AI, but folding it into roles where supervision already exists.
There is also the matter of where the data goes. Claude Code and similar agents send code and context to third-party models in the cloud, and banks have long required that client-identifying information never leaves their own managed environments. A tool that reads client positions or trading flows and ships them to an external service for inference is a category of risk that general productivity chatbots simply do not touch.
What it means for the AI industry
For Anthropic and OpenAI, stories like Goldman's are a warning about the durability of enterprise revenue. Coding agents have been one of the fastest-growing product categories in AI, but if large financial institutions conclude that the tools belong only in engineering teams, the addressable market for premium-priced AI subscriptions shrinks.
For now, Goldman's engineers keep their agents, and its traders and bankers will have to go back to doing things the old way — at least until the bank's review of "best use cases" concludes. The episode is a reminder that inside the world's most regulated industries, the path from AI experimentation to institutional adoption runs through risk committees, not just demos.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
