Goldman Sachs has taken back traders' and bankers' access to AI coding tools. On November 2, 2026, Bloomberg reported, citing internal guidance Goldman updated in October, that Anthropic's Claude Code and Microsoft-owned GitHub Copilot will now be available only to engineers, and traders and bankers will no longer be able to use them. A Goldman spokesperson said the decision took productivity and risk management into account and that the experimentation phase is over. This is not a temporary outage caused by a technical fault. It is a shift in a major bank's stance on AI coding tools: from encouraging everyone to try them, to restricting them by role.
What traders lost, and what engineers kept
Before access was withdrawn, traders and bankers had already used the two tools for plenty of work beyond writing code, including monitoring client trading activity, building market models, and producing written and graphic material. Bloomberg reported that some traders were unhappy about losing access. Those who keep it are roughly 12,000 engineers, plus the strats teams that span trading, quantitative research, and software development. Goldman's argument is that the tools should go to the people best suited to using them long term, rather than everyone continuing to experiment. That changes what the tools are: no longer a productivity trial for the whole firm, but production tools for engineers.
Encouraged in July, withdrawn in October: why the turn?
The change of direction happened within a few short months. In July 2026, Goldman chief information officer Marco Argenti was still publicly encouraging non-technical staff to try vibe coding: using AI tools to build a quick prototype, then handing it to an engineer to finish. By October, that same group of non-technical staff had lost access. Goldman spokesperson Nick Carcaterra disclosed no specific productivity figures or risk incidents, saying only that the firm had weighed a variety of factors, that the experimentation phase was over, and that the tools would stay with those best placed to use them over the long term. Bloomberg's reporting also disclosed no specific incident or internal audit behind the withdrawal, so outsiders cannot tell whether access was pulled because the tools worked poorly in trading, or because code and data were flowing somewhere they should not.
Goldman has not banned AI outright; it has narrowed who may use these two advanced coding tools. The full text of the internal guidance has not been made public, and it remains to be seen whether internal applications already built will be affected, and whether other banks will follow with role-based limits of their own. For financial firms, the signal is clear: getting AI tools into the office is the easy part. Drawing the line over who may use them, and at which step, is the real management problem.
