Goldman Sachs Took Away the AI Coding Tools Its Traders Had Learned to Use

Goldman Sachs has told most of its staff to stop using advanced AI coding tools, after weighing what the tools saved against what they risked. The restriction, reported by Bloomberg on 3 November, has annoyed some of the people who had come to rely on them.

For several months the bank had let bankers and traders use Anthropic’s Claude Code and Microsoft’s GitHub — products built for writing and reviewing software — for their own work. An internal guidance note updated in October now limits those tools to engineers.

The frustration is easy to understand. Traders had been using them to monitor client order flow, build models of market volatility and their own positions, and draft written and chart material. All of that is work a language model is genuinely good at, and none of it requires being a software engineer.

As recently as July, Goldman’s chief information officer, Marco Argenti, was praising what he called “casual coding” by non-technical staff: bankers would sketch a first version with AI and hand it to the firm’s 12,000 engineers to harden.

A spokesperson, Nick Carcaterra, framed the change as triage rather than retreat. The bank opened some tools to see which parts of the business they suited, ran a trial, then weighed efficiency and risk control and gave access to the people who would get the most out of them over time.

That weighing is the hard part, and every bank is doing it in public. Firms are spending heavily on AI while trying to contain cyber risk, cost and the awkward question of what the output is actually worth. JPMorgan said in July that it was steering staff to pick the right model for the job rather than reaching for the most expensive one to summarise an analyst report. Goldman has built its own assistant, drawing on models from OpenAI, Anthropic and Alphabet.

There is an awkward symmetry here. Goldman was lead underwriter on Anthropic’s initial public offering and has worked with Nvidia on large bond deals — so it is at once a customer, a gatekeeper and a beneficiary of the companies whose tools it is now restricting.

The new rule does not apply to quantitative strategists, who straddle trading, research and software development. For everyone else the message is blunter than the industry’s usual optimism: the tools are useful, but not everyone gets to keep them.