AI Briefing: November 3, 2026 — Goldman Sachs reportedly limits AI coding tools to engineers.

WHAT WAS REPORTED

According to coverage of the Bloomberg report, Goldman Sachs allowed bankers and traders to use AI coding assistants for months, then changed course in internal guidance published last month. The tools, named in coverage as Anthropic’s Claude Code and Microsoft’s GitHub Copilot, are now meant for the bank’s engineers. Traders had reportedly used them to track how clients were trading, simulate shifts in markets and positions, and draft text and visuals. Staff outside engineering who relied on them are reportedly told to request alternatives or use the bank’s in-house GS AI Assistant.

WHY THE BANK MAY HAVE CHANGED ITS MIND

The stated driver is a review of productivity and risk, though I found no detail on what the review measured. The context is suggestive. In July, Goldman’s chief information officer Marco Argenti spoke favourably of “vibe coding” by staff outside technical roles. In August, a partner reportedly warned that heavy reliance on AI could erode bankers’ reasoning. In April, the bank had already restricted Claude access for bankers based in Hong Kong. Code written by a trader and never reviewed by an engineer is, for a regulated firm, an unmanaged model: it can touch client data, feed decisions, and leave no audit trail.

WHAT IS STILL UNCONFIRMED

Outlets disagree on how to describe it. Some call it a ban; others say access was narrowed. Dates also vary between “updated in October” and “announced in October 2026”. The number of affected staff has not been disclosed, and I could not open Bloomberg or the internal guidance. Goldman’s only reported comment is that it makes certain AI tools available to its people to determine the best use cases across the firm. Nothing here suggests the bank is retreating from AI generally.

WHY THIS MATTERS FOR TEAMS BUILDING ON AI

This is the same tension our vibe-coding hangover piece described, now showing up in a large regulated company. The pattern to learn from is the line Goldman drew: AI-generated code is welcome where an engineering process surrounds it, and restricted where it is not. Three habits are cheap. Decide which roles may ship code at all, and route non-engineers’ scripts through review. Log which tool produced which artefact, so an audit can reconstruct it. And offer a sanctioned alternative, because a ban without one pushes people toward personal accounts, which is the riskier outcome.