Goldman Sachs Pulled AI Coding Tools From Its Bankers and Traders While BNY Lets Everyone Build

In July, Goldman Sachs' chief information officer was cheering the bankers and traders who "vibe coded" their own tools. By October, the bank had taken those tools away from them.

Goldman has barred most traders and bankers from AI coding tools including Anthropic's Claude Code and Microsoft's GitHub, Bloomberg reported. Internal guidance updated in October limits them to engineers. Bankers had been using them for months before a review of productivity and risk led to the change.

Traders lose their tools

Some traders are not happy. They had used the tools to follow how clients were trading, simulate shifts in markets and in their own positions, and draft text and visuals. The directive does not cover "strats," the employees who mix software engineering, quant research and trading.

Spokesperson Nick Carcaterra described it as the natural end of a trial. "We make certain AI tools available to our people to determine the best use cases across different parts of the firm," he said. After experimenting and weighing "productivity and risk management," the bank put the tools "in the hands of those who are best suited to leverage them for the long term."

Goldman has not disclosed a security incident or put numbers on the productivity findings behind the decision.

A fast reversal

The turn is sharp. In July, CIO Marco Argenti spoke warmly about traders and bankers building prototypes that the bank's pool of about 12,000 engineers then polished. A month later, Goldman partner Chris Churchman, co-chair of its Global Banking and Markets AI working group, warned that AI risks degrading the analytical skills of the next generation of bankers.

Goldman is hardly anti-AI. It built its own assistant on models from OpenAI, Anthropic and Alphabet. But it has drawn lines before, restricting staff in Hong Kong from using Anthropic's Claude earlier this year.

The worry is easy to picture. Financial software can run perfectly and still spit out plausible numbers built on a wrong pricing assumption. A coding agent with internal access can pull confidential data, link restricted databases or push changes around approval processes.

Who gets the keys

Not every bank landed in the same place. BNY lets non-engineers build on its Eliza platform inside a governed environment, and lists more than 400 AI-enabled solutions in production, with AI writing roughly 40% of its code. MUFG went another way and built its own alternative to Claude Code after finding the tool could work around its information-access limits and compliance safeguards.

Deutsche Bank keeps the tools with professional developers. More than 15,000 of them use GitHub Copilot and Google Gemini Code Assist, saving an estimated 1.5 to 2.5 hours a week each, and they remain responsible for the code they commit. JPMorgan said in July it wants staff to match each task to a suitable model, saving the priciest for harder work, and it has already shut outside AI agents out of customer accounts over the risk of rogue agents.

That is the real fight on Wall Street. Nobody is arguing about whether AI writes code. The question is who gets to let it touch the bank's systems.

It matters to the vendors too. AI coding companies are selling the idea that anyone can build software from plain language, a pitch that grows their market far beyond engineers. Goldman's answer suggests banks may instead pour their spending into specialists and in-house platforms.

Some offers on this page may be paid placements or contain affiliate links.