Playbooks · Teams
Goldman Sachs limits Claude Code to engineers
Written November 9, 2026
Goldman Sachs cut Claude Code and Copilot for bankers and traders — engineers keep it. What the bank said, and what it means for your rollout.
Bloomberg's November 2 report on this change is paywalled. The details below come from outlets quoting that report, and Goldman Sachs has not published the internal guidance itself — so every fact here is secondhand, and the bank's reasoning beyond its quoted words is not public.
What happened
Internal guidance updated in October 2026 restricts Claude Code and GitHub Copilot to the bank's engineers, roughly 12,000 people, plus its "strats" group — quantitative specialists who also write production code. Bankers and traders who had used both tools for months lost access. The tools write, refactor, and run code, and their use had spread well beyond engineering teams, into parts of the business with no software development function — monitoring trading flows and generating content among them.
A Goldman spokesperson told Bloomberg: "After a period of experimentation where we assess a variety of factors including productivity and risk management, we put them in the hands of those who are best suited to leverage them for the long term." That sentence is the whole of the bank's public explanation. The bank has disclosed no security incident behind the change and no productivity data supporting it.
The reversal
The restriction reverses a position the bank promoted in public four months earlier. In July 2026, Goldman CIO Marco Argenti publicly praised non-technical staff for "vibe coding" — turning an idea into a fast prototype with an AI tool, then handing the result to an engineer. By October, access for that same group was withdrawn. There is also a precedent: in April, the bank removed Claude from its platform for Hong Kong bankers after reading its Anthropic contract strictly.
What it is not
This is not a ban on AI at Goldman, and it is not a ban on Claude Code generally. Engineers keep both tools. The bank's own assistant, GS AI Assistant, remains available to everyone at the firm and runs on models from several companies, including Anthropic's Claude. What lost its pass is narrower: third-party agents that write code and then run it on their own, in the hands of people whose work has no code review step.
The lesson for your rollout
A rollout without a risk story gets rolled back by role. Goldman ran months of open experimentation, then drew the line where code review already existed: engineers and strats, whose code flows through engineering review, kept the tools; everyone else lost them. If you are rolling coding agents out across a company, write down before the rollout who owns what the agent produces, and what review stands between the agent and production. Access decided by job title is the coarsest control there is — the review boundary underneath it is the control that actually holds.
Limits
- Bloomberg's November 2 report is paywalled and was not read directly. The restriction, the roughly 12,000 figure, and the spokesperson's wording are consistent across independent outlets quoting it, but they are secondhand.
- Goldman has not published the internal guidance, so exactly which GitHub Copilot features the restriction covers is unconfirmed beyond the reporting.
- No incident, regulator request, or productivity figure has been disclosed. Do not present one as the reason.
- The Hong Kong cut in April is reported by the Financial Times; it is included here as context, not as the same policy.
Related playbooks
- Claude Code pools budgets and reports costs via API
- Lock down Claude: IP allowlist and sign-out
- Claude Desktop Unified turns on Nov 10
- Claude for Manufacturing: 17 plant-office skills
Sources
- Goldman Sachs restricts Claude Code and GitHub Copilot to engineers — DEV Community
- Goldman Sachs Locks AI Coding Tools to Engineers — AI Tool Briefing
- Who gets the tool is a control. It should not be the only one. — Orca
Back to: Teams playbooks hub · Also: DHS DeepSeek order: inventory your stack · Decide whether Copilot should approve PRs · Copilot Automations learn from real outcomes