Goldman Sachs (NYSE:GS) has curtailed access to sophisticated artificial intelligence coding assistants for most of its non-engineering staff, including traders and bankers, following an internal assessment of productivity gains against potential risks.
According to reporting based on internal guidance revised in October, the firm previously permitted broader experimentation with tools such as Anthropic’s Claude Code and Microsoft’s GitHub.
Those systems are now largely confined to the bank’s engineering teams.
The shift has disappointed some traders who had incorporated the assistants into daily workflows.
Individuals familiar with the matter indicated that staff had relied on the tools to track client order flows more rapidly, construct models of market movements and positions, and generate written summaries or visual materials.
The restriction does not extend to “strats,” the specialized employees who combine trading expertise, quantitative analysis, and software development.
The policy change follows a period of more open testing.
As recently as July, Chief Information Officer Marco Argenti highlighted efforts by non-technical employees—including traders and bankers—to prototype solutions through AI-assisted coding, sometimes described as “vibe coding.”
Those prototypes were then handed to one of the firm’s roughly 12,000 engineers for refinement and production use.
Goldman spokesperson Nick Carcaterra stated that the bank makes selected AI tools available so it can evaluate suitable applications across different divisions.
After assessing factors such as productivity and risk management during the trial period, the firm assigns the tools to the groups best positioned to apply them sustainably over time.
The move reflects wider efforts across major financial institutions to define appropriate boundaries for generative AI.
Banks continue to invest heavily in the technology while grappling with concerns over cybersecurity, operational controls, cost efficiency, and measurable output.
JPMorgan Chase, for instance, has encouraged employees to match specific models to particular tasks in order to avoid overusing more expensive systems for routine work such as document summarization.
Goldman has developed its own internal AI assistant that draws on models from OpenAI, Anthropic, and Alphabet.
At the same time, the bank is positioned to benefit commercially from the sector’s growth.
It has been selected as a lead underwriter for Anthropic’s anticipated initial public offering and has collaborated with Nvidia on substantial debt financing arrangements linked to AI infrastructure.Earlier this year, Goldman imposed a separate, location-specific limitation by restricting use of Anthropic’s Claude among staff in Hong Kong.
The latest guidance represents a role-based rather than geographic tightening. In May, President and Chief Operating Officer John Waldron—widely viewed as a potential successor to CEO David Solomon—described Goldman as a “human assembly line” that could become more automated and scalable through digital agents.
The restriction underscores the tension between rapid experimentation with powerful coding agents and the need for controlled, long-term deployment in a highly regulated industry. While engineers retain access to accelerate development, broader staff must now operate within narrower parameters as the firm continues refining its overall AI strategy.
