Goldman Sachs just told its bankers and traders to put the coding tools down. Per Bloomberg’s November 2 report, internal guidance updated in October restricts Anthropic’s Claude Code and Microsoft’s GitHub Copilot to the bank’s roughly 12,000 engineers, pulling access from the traders and bankers who’d been using both tools for months.

Four months earlier, Goldman’s own CIO was holding the opposite tools up as a feature, not a liability.

Quick Summary: What Happened

Detail	Info
What changed	Claude Code and GitHub Copilot access restricted to engineers only, per internal guidance updated in October 2026
Who lost access	Bankers and traders who had used both tools for months for tasks with no software development need
Who kept access	Goldman’s roughly 12,000 engineers, plus its “strats” group (quant specialists who also write production code)
Why	Goldman cited “a variety of factors including productivity and risk management” after what it called a completed “period of experimentation”
The reversal	CIO Marco Argenti publicly praised non-engineer “vibe coding” as recently as July 2026
Official source	Bloomberg, Nov 2, 2026

Bottom line: Goldman ran the experiment, let AI coding tools spread past engineering into trading and banking, and then pulled them back. It’s the clearest enterprise walk-back yet on letting non-engineers build with agentic coding tools.

What Actually Happened

Goldman had given bankers and traders broad access to Claude Code and GitHub Copilot (tools built for writing and shipping software) for jobs that had nothing to do with writing software. Per Bloomberg, people in those roles used the tools to monitor trading flows and generate written content, the kind of general-purpose assistant work you’d expect from a chatbot, not a coding agent. That’s the detail that makes this story different from a routine licensing cleanup: the tools didn’t stay in their lane, and nobody stopped them for months.

The reversal wasn’t quiet. Goldman spokesperson Nick Carcaterra 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.” Translate that out of corporate-speak and it reads as: we let this run wide, we looked at what came back, and the answer wasn’t “keep going.”

The access that remains is narrow and specific. Goldman’s roughly 12,000 engineers keep Claude Code and Copilot. So does the bank’s “strats” group, quantitative specialists who sit between trading desks and engineering and write production code as part of the job. Everyone else at a firm with tens of thousands of employees is back to doing banking and trading without an agent that can touch a codebase.

Why Did Goldman Reverse Its AI Coding Tool Policy?

1. The tools spread past their intended use. Claude Code and Copilot are built for software development. Bankers and traders were using them as general-purpose assistants, which is a different risk profile than the one the bank signed off on.
2. The “experimentation” window closed. Goldman’s own language frames this as the end of a trial period, not an emergency shutdown. It reads as a scheduled recalibration rather than a scramble.
3. Risk management caught up with adoption. An AI agent capable of writing and executing code sitting on a trading desk’s machine is a materially different control problem than the same agent on an engineer’s machine, where code review and deployment gates already exist.
4. Productivity didn’t clearly justify the exposure. Carcaterra’s “variety of factors including productivity” phrasing suggests the output didn’t obviously outweigh the risk once someone actually measured it.

The Argenti Reversal

Here’s the part worth sitting with. In July 2026, Goldman CIO Marco Argenti was publicly describing a workflow he liked: a trader or someone in asset management gets an idea, builds a quick prototype with AI (“vibe coding,” in the term everyone was using by mid-2026) and hands it to an engineer from the bank’s pool to scale for production. That’s not a grudging tolerance. That’s a CIO citing non-engineer AI coding as a feature of how Goldman worked.

Three months later, the policy that enabled exactly that workflow got pulled. Argenti didn’t frame vibe coding as a mistake in July. He framed it as the point. Whatever changed between July and October, it wasn’t his read on the concept. It was what actually happened once “a trader has an idea” turned into thousands of non-engineers across trading floors and banking desks running agentic coding tools for months with no software-development need at all.

That gap (enthusiastic public endorsement, then a quiet internal pullback three months later) is the most useful data point in this story for anyone running a smaller version of the same experiment. We’ve covered Goldman’s caution with AI access before: the bank has been willing to bet real money on Claude in regulated functions, but it’s also shown it will cut access fast when something about the exposure doesn’t sit right. Goldman staff in Hong Kong lost Claude access entirely back in April over data-boundary concerns specific to that office. This firm moves fast in both directions.

Why This Matters

If you’re evaluating Claude Code or GitHub Copilot for your own non-engineering staff, Goldman just ran the pilot program for you, at a scale almost nobody else can replicate, and published the result: it works well enough that people keep using it for months, and it creates enough risk that a bank with Goldman’s compliance muscle decided the tradeoff wasn’t worth it long-term.

The specific risks aren’t abstract for a bank. Code-capable agents touching material non-public information, data barriers between desks, and systems that interact with live trading infrastructure are a different category of exposure than a marketing associate using a chatbot to draft copy. An agent that can execute code is an agent that can, by design, take actions — not just generate text. Giving that capability to people outside the review and deployment processes engineering teams already have built around it is a governance decision, not just a licensing one.

That’s a narrower point than “AI coding tools are risky.” AI code assistants compared head to head are generally judged on how well they write and refactor code for people who already know how to evaluate what comes back. Goldman’s problem wasn’t code quality. It was putting a code-executing agent into the hands of people whose job isn’t to evaluate code at all.

What Are Your Options Now

If you’re a Goldman employee outside engineering who lost access, the bank’s own statement is the actual answer: these tools now go through engineers and the strats group. Route coding needs through that channel rather than working around the restriction.

If you’re evaluating a similar rollout at your own company, a few things are worth taking from Goldman’s experience directly:

• Scope access to the job, not the org chart. Vibe coding as a prototype-then-handoff workflow (the thing Argenti described in July) is different from letting non-engineers run agentic tools unsupervised for months. The failure mode here wasn’t the prototype step. It was the tools becoming general-purpose assistants with nobody tracking what they were actually being used for.
• Set a review date before you launch, not after something goes wrong. Goldman’s language (“a period of experimentation”) implies this was reviewable from the start. Build that checkpoint into your own rollout instead of discovering you need one.
• Separate “can write code” from “can be trusted with code execution.” The tools that make vibe coding appealing (fast prototyping, natural-language-to-working-app) are a different risk surface than a tool executing against production systems. Treat them differently in your access policy.

The Bigger Picture

This lands in the middle of a broader skittishness across banking about agentic AI and security. Techstrong reported that JPMorgan Chase, Bank of America, Citigroup, and Charles Schwab have moved to block third-party AI agents, tools like Meta’s Muse, from accessing customer accounts and executing transactions, citing security risks and the threat of automated agents sweeping idle cash out of low-yield deposit accounts. That’s a different specific risk than Goldman’s internal tool restriction (external agents touching customer money versus internal coding tools touching trading desks), but it’s the same underlying posture: banks deciding agentic AI needs a harder boundary than the industry assumed a year ago.

It also follows a pattern Goldman itself has set this year. The bank has been one of Anthropic’s most visible financial-sector bets, deploying Claude in genuinely regulated workflows — and simultaneously willing to yank access somewhere else entirely when the exposure looked wrong, as it did in Hong Kong in April. Goldman isn’t anti-AI. It’s a bank that treats AI access the way it treats any other system with regulatory exposure: expand it where the controls exist, cut it where they don’t.

For the vendors, the stakes are different but real. Anthropic and Microsoft both lost a meaningful chunk of paid seats at one of the most closely watched enterprise accounts in finance, not because the product failed, but because the customer decided the access model was wrong. That’s a harder problem to fix with a product update than a bug report is, and it’s the kind of enterprise signal other banks’ compliance teams will have seen by now too.

Our Take

I don’t read this as Goldman souring on AI coding tools. I read it as a bank admitting its first access policy was too loose, which is a completely different and much less interesting story than “Goldman bans AI,” but it’s the accurate one. Argenti’s July comments weren’t wrong on the concept. Handing a prototype to an engineer for production hardening is a sane workflow. The policy failure was letting the tools drift into general-purpose use by people who were never supposed to be the audience for a coding agent in the first place, for months, without anyone apparently tracking it closely enough to catch that drift sooner.

What I’d push back on is the idea that this settles anything about whether non-engineers should have agentic coding tools at all. Goldman’s specific failure mode (a code-executing agent becoming an ambient assistant with no usage boundary) is a rollout design problem, not evidence that the underlying workflow is unsound. Companies without Goldman’s regulatory surface area and without material non-public information sitting on every desk have a lot more room to let the Argenti version of vibe coding run than Goldman did. The lesson isn’t “don’t let non-engineers near AI coding tools.” It’s “know exactly what they’re using it for, and check back before three months turns into a policy nobody planned.”

Frequently Asked Questions

What did Goldman Sachs actually change?

Per Bloomberg’s November 2, 2026 report, internal guidance updated in October restricts Claude Code and GitHub Copilot access to the bank’s roughly 12,000 engineers and its quantitative “strats” group. Bankers and traders who had used both tools for months lost access.

Which AI coding tools were restricted?

Claude Code, built by Anthropic, and GitHub Copilot, built by Microsoft’s GitHub. Both are agentic coding assistants designed to write, edit, and execute code, not general-purpose chat assistants.

Why did Goldman reverse its policy on non-engineers using these tools?

A Goldman spokesperson told Bloomberg the bank was assessing “a variety of factors including productivity and risk management” after a completed “period of experimentation.” Bankers and traders had been using the coding tools for tasks unrelated to software development, including monitoring trading flows and generating content.

Did Goldman’s CIO support “vibe coding” before this reversal?

Yes. In July 2026, CIO Marco Argenti publicly described a workflow where a trader or asset-management employee would build a quick AI-generated prototype and hand it to an engineer for production. That endorsement predates the October restriction by about three months.

Who still has access to Claude Code and Copilot at Goldman?

Goldman’s roughly 12,000 engineers and its “strats” group, quantitative specialists who write production code as part of their role.

Is this part of a wider trend in banking?

It appears to be, though not in identical form. Techstrong reported that JPMorgan Chase, Bank of America, Citigroup, and Charles Schwab have been blocking third-party AI agents from accessing customer accounts and executing transactions over security and deposit-sweeping concerns, a different specific risk than Goldman’s internal engineer-only restriction, but part of the same 2026 pattern of banks tightening the boundaries around agentic AI.

Does this mean AI coding tools are too risky for non-engineers generally?

Not necessarily. Goldman’s specific concerns (material non-public information, data barriers between trading desks, and code-executing agents operating without engineering-style review) are heightened by the bank’s regulatory environment. Organizations without that exposure have more room to extend agentic coding tools to non-engineers, provided usage is scoped and monitored from the start.

Where can I learn more about Claude Code or GitHub Copilot for enterprise use?

See our guides on Claude Code Routines for enterprises, GitHub Copilot’s token billing model, and a head-to-head comparison of AI code assistants.

---

Last updated: November 3, 2026. Sources: Bloomberg — Goldman Restricts Traders and Bankers From Using Some AI Tools · Bloomberg — Goldman Staff in Hong Kong Lose Access to Anthropic’s Claude · Claude Code · GitHub Copilot · Techstrong — U.S. Banking Giants Block AI Agents Over Security, Rate-Sweeping Threats.

Related reading: Anthropic Bets $1.5B on Claude in Banks and Hospitals · Claude Code Routines: What Enterprises Should Know · GitHub Copilot Token Billing: What It Costs Now · AI Code Assistants 2026: Cursor vs Copilot vs Codeium vs Tabnine · Vibe Coding in 2026: The Best Platforms for Building Apps by Describing What You Want