One of the world's biggest banks has decided that AI coding tools are for engineers. Bloomberg reported on 2 November that Goldman Sachs has limited advanced coding tools, including Anthropic's Claude Code and a Microsoft GitHub coding tool, to its engineering staff, and that most other employees, including traders and bankers, have lost access. The move is a useful signal for anyone outside an engineering role who has started building with AI at work, because it shows how the question of who gets these tools is changing as experiments turn into budgets.
What was reported
The details come from Bloomberg's report, which is behind a paywall, as summarised by The Decoder and Techstrong.ai. We could not read the Bloomberg article in full, so the points below rest on those summaries and on the quotes from Goldman's spokesperson they include.
- Goldman has restricted advanced AI coding tools, naming Claude Code and a coding tool from Microsoft's GitHub.
- Engineers keep access, and Techstrong.ai adds quantitative "strats", specialists who work between trading and software. Most other staff, including traders and bankers, have lost it.
- The change is based on internal guidelines dated October.
- Traders had used the tools for months to monitor trading flows and create content.
The reports do not say whether Goldman has changed its wider relationship with Anthropic or Microsoft, and the internal guidelines have not been published.
What Goldman said
Goldman spokesperson Nick Carcaterra said the bank makes certain AI tools available "to determine the best use cases across different parts of the firm", according to Techstrong.ai. The Decoder reports that he cited "a variety of factors including productivity and risk management" and said the tools will go to people "who are best suited to leverage them for the long term." The Decoder sums this up as the experimentation phase being over. None of the reports name a specific incident behind the decision.
The change is a turn from Goldman's earlier tone. In July, chief information officer Marco Argenti praised vibe coding, the practice of building software by describing it to an AI, by non-technical staff. The Decoder says the plan he described was for a trader to build a quick prototype with AI and hand it to an engineer to scale, and it reports that the rollback has caused frustration inside the firm.
Why a bank would draw this line
Coding agents like Claude Code do more than suggest text. They can write and run code, connect to data and build small tools that other people start to rely on. In a regulated bank, a tool built by a trader can touch market data, client information or risk calculations, and someone has to be accountable for checking it. Limiting the tools to engineers puts that work with people who already follow code review and testing rules. The cost is that useful tools built by people closest to the problem become harder to make.
There is also a budget question, since productivity is one of the two reasons Goldman gave. Once a trial ends and a company pays for every seat and every task, it is natural to ask which users get the most value per dollar, and engineers are the easiest group to justify.
This is also the second time this year that Goldman has narrowed access to Claude. In April, Reuters reported that the bank removed Claude for its bankers in Hong Kong, citing a Financial Times report that the move followed a strict reading of its contract with Anthropic, while Gemini and ChatGPT stayed available on its internal platform. Anthropic told the FT that Claude had never been officially supported in Hong Kong.
Goldman is not the only large organisation tightening its AI tool list. Last week Sweden's Government Offices blocked ChatGPT and Claude on staff computers and steered staff to Microsoft Copilot, according to Decrypt and The Next Web, which cite Swedish public radio. Some staff there reportedly keep access through an internal platform, as we covered in our piece on the Swedish decision. The reasons differ in each case, and all three show employers moving from open trials to formal approval.
What it means for non-engineers who build with AI
We think this is where workplace AI is heading in general. As companies move from trials to paid rollouts, access will follow evidence of value and a clear answer on risk. People outside engineering who have built useful tools with AI are in a stronger position if they can show what the tool does, what it saved and who checks it. Without that, a policy written for safety can remove the tool without anyone weighing what is lost.
What to do this week
If you use an AI coding tool outside an engineering role, prepare short answers to five questions before your company asks them:
- Which task do you use the tool for, in one sentence?
- How many hours a week does it save you, roughly?
- What company or client data does it touch?
- Who reviews what it builds before anyone relies on it?
- What breaks, or slows down, if it is taken away tomorrow?
If you manage a team, ask the same questions of the tools your team already uses. To compare the assistants your company might approve, see our AI tool comparisons.