• Goldman Sachs has restricted certain advanced AI coding tools to engineers, reversing access previously available to bankers and traders, according to people familiar with the matter.
  • The move follows an October policy update aimed at addressing productivity and risk concerns, not a blanket ban on AI.
  • The restriction highlights the growing governance challenge for banks deploying powerful AI tools while managing regulatory and operational risks.

Goldman Sachs Group Inc. has tightened access to advanced artificial intelligence coding tools, limiting their use to engineers and revoking permissions for bankers and other staff who had been using them for months, according to internal guidance updated in October and people familiar with the matter. The restriction, which covers tools such as Anthropic’s Claude Code and Microsoft (MSFT)’s GitHub, comes after a review of productivity and risk concerns, these people said, asking not to be identified discussing internal policy.

The change has sparked frustration among some employees who had incorporated the tools into their daily workflows. One person described the mood as “disappointed,” though others acknowledged the need for tighter controls. A Goldman spokesperson declined to comment.

Governance Over Access

The central issue is governance: who can safely use increasingly powerful AI tools inside a regulated bank. By limiting advanced coding tools to engineers, Goldman is drawing a sharper boundary around access, prioritizing oversight and testing over broad-based experimentation. The bank continues to develop and deploy its own AI assistant, GS AI Assistant, for tasks such as document summarization and data analysis, indicating the restriction is not a retreat from AI but a recalibration of risk.

The move follows comments from Goldman CIO Marco Argenti, who warned at an October 22 industry event about the risks of autonomous AI systems, emphasizing the need for containment, rules, and approval gates before agents can execute transactions. His remarks provide context for the policy shift, though they do not fully explain every restriction.

Goldman’s decision reflects a broader industry trend toward controlled enterprise AI. JPMorgan Chase & Co. has built its own LLM Suite, used by 200,000 employees, while prohibiting outside agents from operating within its environment. Banks in the U.S. and Canada have also restricted outside consumer AI agents’ access to customer accounts. The common thread: institutions want to control the environment in which AI acts.

The restriction comes amid strong financial performance. Goldman reported third-quarter net revenue of $18.68 billion, up 23% year over year, and diluted earnings per share of $17.01, up 39%. Equities revenue surged 72% to $6.44 billion, and investment-banking fees rose 21% to $3.22 billion. The bank employed 47,500 people as of September 30 and reported $4.10 trillion in assets under supervision.

Regulatory obligations remain a key driver. FINRA, the U.S. securities industry’s self-regulatory organization, has emphasized that existing securities laws apply to AI use, including supervision, recordkeeping, and fair dealing. Using a third-party model does not remove a firm’s responsibility. Its 2026 guidance stresses formal approval processes, testing, and continuous monitoring of prompts and outputs.

The long-term question is whether Goldman can preserve AI’s productivity benefits while preventing uncontrolled software development, data exposure, and operational errors. For now, the bank is signaling that access will be earned, not assumed. The restriction establishes a boundary; whether it proves temporary or permanent remains to be seen.

Correction: An earlier version of this article misstated the timing of the policy update. It was updated in October, not November.