Goldman Sachs moves to a new phase of AI: Marco Argenti on agents

Goldman Sachs is revisiting its AI strategy as agent tools grow quickly and corporate use cases mature. Marco Argenti, the bank's chief information officer, is talking not just about internal experiments but about how the new wave of AI is changing software development and everyday processes inside a financial institution.

What has changed in a year and a half

A year and a half ago, Goldman Sachs mostly talked about building its own internal AI tools and where it was trying to apply models without releasing them into a broad production environment. Since then the market has accelerated sharply: generative models have become more accurate, corporate use cases clearer, and agent systems, which do not merely answer a query but can carry out chains of actions, have moved to the foreground.

This changes the conversation inside companies too: from "let's give it a try" to "how much of the work can we hand to a machine right now."

For a bank like Goldman Sachs, this shift matters for more than AI fashion. Where value was once often measured by the quality of chat answers, the focus is now on employee productivity: how much faster a developer writes code, an analyst assembles a draft, or a team finds the information it needs inside its own systems. In this sense, agent platforms such as Claude Code are becoming not a standalone toy but a new interface to corporate work, where both results and control are required.

Where the bank is looking

From the account of the new conversation with Argenti, Goldman Sachs' focus is shifting from the mere fact of adoption to the quality of use. It is no longer enough for the bank to have internal AI services. What matters now is understanding which tasks they handle better than classic software, where they genuinely save time, and where they demand too much human review. For a large financial organization this is a practical approach: any automation is valuable only when it fits existing processes, security requirements and team accountability.

• Internal tools for employees
• Support for development and work with code
• More complex chains of actions instead of single queries
• Assessment of real usefulness, not just answer quality

Why agents matter

The main point of interest is how the bank evaluates the agent approach specifically. This is no longer only about an assistant that summarizes documents or answers questions from a knowledge base, but about systems capable of performing multistep tasks. For engineering and product teams this is especially visible: such a tool can read code, suggest changes, look up dependencies and handle part of the routine work without constant switching between windows. But the higher the autonomy, the sharper the questions of trust, action logging and the human's right to the final decision.

That is why Goldman Sachs' experience is interesting not as a story about "a bank that is also adopting AI," but as a signal of the market's next stage. When such organizations discuss not demo pilots but working scenarios and the effect on specialists, it means the technology is being tested for usefulness. Even if some experiments do not deliver immediate results, the shift itself matters: corporate AI is gradually turning from a set of chatbots into an infrastructure layer that helps write, search, check and perform actions inside a company.

What this means

For the whole corporate AI sector, this is a marker of maturity. The conversation around AI is moving from abstract promises to measurable efficiency: where agents save hours of work, how they fit into sensitive processes, and which tasks can be entrusted to them without losing control. If this approach takes hold in banks, it will quickly become a benchmark for other large international companies.