Goldman takes AI coding tools back from traders and bankers

What happened

Bloomberg reported on 2 November that Goldman Sachs has cut most of its employees off from advanced AI coding tools, including Anthropic’s Claude Code and GitHub Copilot. The report cited internal guidelines from October. Engineers keep access. Traders and bankers, who had been using the tools to monitor trading flows and create content, don’t, and the change reportedly frustrated them.

It’s a sharp turn. In July, chief information officer Marco Argenti publicly praised vibe coding by non-technical staff. He described a trader or asset manager building a prototype with AI and handing it to an engineer to scale. A Goldman spokesperson, Nick Carcaterra, said the decision reflected a “variety of factors including productivity and risk management.” He said the experimentation phase is over, and that the tools will go to the people “who are best suited to leverage them for the long term.”

My view

This isn’t a verdict on AI coding. It’s a verdict on software nobody owns, running with access nobody designed.

A script that watches trading flows stops being a prototype the first morning someone relies on it. From then on it has users, inputs that change and a failure mode nobody has tested. The plan in July depended on a handoff: the trader builds, the engineer scales. In practice the handoff is the step that doesn’t happen. The prototype works well enough, the engineer is busy, and the code keeps running on the desk.

The real exposure is access

The reporting cites productivity and risk management, without detail. Here is the risk I’d expect any bank to find in this setup. A tool a trader builds with an AI coding assistant runs with the trader’s own entitlements. Whatever the trader can open, the tool can read, join and send on.

That matters more than it sounds. A person with broad access still works one system and one screen at a time, and the firm’s controls assume that. A generated script doesn’t. It can pull positions, client activity and research into one dataset in seconds, and leave the result in a spreadsheet, a chat or a shared folder with none of the original labels. Data that sat in separate places, some of it behind information barriers, is now commingled in an artifact nobody classified. The coding assistant also sees whatever the trader pastes into it.

None of this needs a bad actor. It only needs a helpful tool, broad entitlements and no check between the request and the data.

Why it matters

For the industry. Every firm that opened AI coding tools to non-engineers has the same question to answer: what can these tools reach, and where does the output go? Withdrawing the tools answers it bluntly. Most firms will need a better answer, because the demand from the business won’t go away, and spreadsheets have shown for decades that desk-built tools find a way. Expect more firms to narrow access first and rebuild it with controls, rather than leave it open.

How it connects to our work. This is the problem Context Before Query is built around: authorization is checked at the data for each request, before anything is retrieved, rather than inherited from whoever started the tool. Agent Onboarding makes the second point. An agent or generated tool gets its own bounded identity and an owner, not a copy of the user’s access. In practice that means four things:

• Sort by consequence, not job title. Exploring data and drafting analysis can stay open in a contained environment with logging. Anything that touches trading data, client content or a decision someone else relies on needs a named owner and an entry in the register before it’s used.
• Give tools their own access. A tool should reach the specific datasets its purpose needs, through governed paths that apply the same information barriers as everything else, and nothing else the user happens to hold.
• Carry labels to the output. If the inputs were restricted, the result is too. Exports and shared files keep the label of the data behind them.
• Make the handoff a real path. A desk-built tool worth keeping moves to engineering with a date, a reviewer and acceptance tests. If nobody will take it on, it’s retired.

“Who is best suited to use these tools” is the right question. I’d answer it by task, by what the tool can reach and by what the result can affect. Once those controls exist, I’d expect the answer to include more than engineers.

Sources

• Goldman Sachs pulls AI coding tools from traders and bankers, The Decoder, 2 November 2026, reporting Bloomberg’s story of the same day.

AI adoptionBankingData accessAI governance

Drafted with AI research assistance and reviewed by Richard White before publication. Linked sources are the record; check them for the full reporting.