Goldman Sachs has restricted advanced AI coding tools to its engineers, ending months in which bankers and traders could experiment with them across the firm. Under internal guidance updated in October, tools including Anthropic’s Claude Code and Microsoft’s GitHub are now reserved for software engineers, with quantitative specialists known as strats also exempt, while other staff lose access they had used for months.
What changed
The tools write, refactor and run code, and their use had spread beyond engineering teams. Traders had used them to monitor client trading flows, model market changes and positions, and create written and graphic materials, according to a person familiar with the matter. Some of those traders are frustrated by the restriction, which follows a review of productivity and risk.
Goldman spokesperson Nick Carcaterra framed the move as the end of an experiment, not a rejection of the tools. The bank first made certain AI tools available widely to find the best uses across the firm, he said. After assessing productivity and risk, it is putting them in the hands of those best suited to use them long term.
A sharp reversal from July
The change reverses the bank’s public posture from July, when chief information officer Marco Argenti praised non-technical staff for building software prototypes with AI tools. Those workers passed their creations to one of the firm’s 12,000 engineers to refine, in a workflow often called vibe coding, which let front-office staff turn ideas into working software without an engineering queue.
Goldman is not the first bank to redraw the line. Earlier this year it restricted Claude use by staff in Hong Kong, a location-specific limit. More broadly, rival banks have also moved to control outside AI agents, with JPMorganChase barring them from its systems and Bank of America and Citigroup placing limits on agents’ access to accounts.
Our read: the experiment proved the risk, not just the value
Goldman’s decision reads less like a verdict on coding tools than on who should create software inside a bank. Code written by a trader to track flows can become operational software, with permissions, data access and failure modes nobody formally owns. Engineers are not immune to those risks, but they work inside review structures that front-office experimentation bypassed.
The timing also matters for the coding-tool market. Agents now draft a large share of code changes, and vendors pitch their tools as safe for everyone. Goldman’s pullback suggests the next selling point is not how many employees can code, but how clearly a firm can show who wrote what, who reviewed it, and who is responsible when it breaks.
Why It Matters
For other large employers, Goldman is a template and a warning at once. Letting every employee build with AI tools can surface prototypes fast, but it also creates shadow software that security and compliance teams cannot see. Firms that keep the gain will give staff a safe path to prototype, with engineers owning anything touching real data or money.
For AI vendors, the lesson is about governance, not capability. A tool that works in a demo needs audit trails and review workflows before a regulated firm lets it loose. The risk is real: a Claude agent reportedly ran for two months on a Pentagon network, showing how quickly experimental tools can reach sensitive work before rules catch up.
Background
Wall Street’s AI debate has shifted from whether the tools work to where they belong. Banks are spending heavily on AI while trying to control cost, productivity and cybersecurity risk, and Goldman’s move shows that even a firm that celebrated employee-built software can decide, after testing it, that coding tools belong with the people trained to own the code.
Source: Bloomberg report via PYMNTS and second Bloomberg report
0 Comments