Goldman Sachs gave 10,000 employees, roughly a quarter of its entire workforce, secret access to an internal AI tool for over a year before anyone outside the company knew it existed.
That quiet pilot phase is worth noting on its own. While competitors were announcing AI initiatives in press releases, Goldman spent a year running a closed test with a meaningful chunk of its actual staff before deciding it was ready for a real rollout.
When they did launch publicly in mid-2025, they didn't ease in. The GS AI Assistant went to all 46,500-plus employees worldwide, firewalled entirely behind Goldman's own infrastructure specifically so no sensitive client or trading data could ever leak into an external model. CIO Marco Argenti, hired from Amazon, built it as a simple chat interface, but with something more sophisticated running underneath.
Here's the architecture decision that actually matters. Goldman built a multi-model system with role-based behavior, meaning a banker's version of the tool is tuned differently than a research analyst's or a developer's, and access to the most capable, most expensive models is reserved specifically for genuinely hard problems, while routine tasks run on smaller, faster models. That's a deliberate cost and capability tradeoff most companies skip entirely by giving everyone the same generic access.
