Goldman Sachs speeds up AI agent adoption and reshapes how its developers work

Goldman Sachs is no longer limiting itself to in-house AI bots for individual tasks. According to CIO Marco Argenti, the bank is reworking its approach to AI adoption amid the rapid rise of agent platforms and new development tools.

From pilots to a system

A year and a half ago, the AI conversation at Goldman Sachs centered mostly on internal tools and isolated automation use cases. Now the focus is broader: the bank sees AI as an infrastructure layer that should be built into teams' daily work, rather than live as a collection of experimental services.

The arrival of agent platforms such as Claude Code has raised expectations. The business wants more than model answers: it needs systems that can run chains of actions, work with code and speed up real processes.

For a large bank, that means stricter architecture requirements. You cannot simply open access to a new model and wait for results. The bank has to determine where an agent is allowed to act on its own, what data it can reach, how its actions are recorded and who is accountable for the outcome.

So AI adoption at the bank looks less like a quick launch of a trendy feature and more like the gradual assembly of a controlled platform with rules, logging and internal restrictions.

How development is changing

A separate topic is AI coding. Tools that help write, check and rewrite code are already visibly changing the work of developers and engineers. It is not only about speeding up routine tasks but about a different allocation of time: less effort goes to boilerplate code, first drafts and searching for standard solutions, and more to review, task definition, architecture and verifying what the agent generated.

For a bank with a large engineering staff, this is not a cosmetic optimization but a change to the production loop.

• Prototypes of internal tools appear faster
• Engineers increasingly work as reviewers and task-setters for agents
• Good documentation, tests and code standards become more valuable
• Errors shift from writing code to verification, security and access control

But the gains do not come automatically. If a team has weak tests, tangled code or unformalized development rules, AI coding starts replicating chaos as fast as it replicates useful solutions. That is why large companies are reviewing their processes in parallel: where human-in-the-loop is mandatory, which changes can be proposed automatically, how to check code security and how to measure real productivity rather than the number of generated lines.

This is also why adopting such tools quickly runs into the discipline of the team.

Data and regulation

The hardest part of scaling AI in a bank is not the interface or the choice of model, but data. For an agent or assistant to be useful, it needs access to internal context: documents, systems, policies, source code, transaction history. But this is exactly where the financial industry faces its strictest constraints.

Every new use case runs into data classification, access separation, retention requirements, audit and explainability of decisions. The more powerful the agent, the higher the cost of a mistake and the tighter the perimeter must be.

Hence the regulatory question. It is not enough for a bank to show that a model is fast and convenient. It must show that its use is reproducible, controlled and compliant with internal and external requirements. In real operation, that means action logs, limits on the use of sensitive data, vetting of model providers and clear escalation paths for when the AI makes a mistake or goes beyond what is permitted.

For Goldman Sachs, AI adoption is at once an engineering project, a risk management project and long-running work with compliance.

What this means

The main takeaway from Goldman Sachs' position is simple: the era of the "chatbot for experiments" is ending. The next stage is AI as managed corporate infrastructure, where value comes not from individual demos but from the combination of agents, data, control processes and a new role for the engineer.