Goldman Sachs and Intuit are not running experiments. They are running operations. When two organizations of that scale and risk profile deploy autonomous SI/AI agents inside their core business processes, it is worth pausing to understand what is actually happening, and what it means for every organization watching from the sidelines.
Let me be precise about what an SI/AI agent is, because the term is used loosely. An agent is not a chatbot. It is not a model that answers questions. An agent is a system that perceives its environment, makes decisions, takes actions, and observes the results, in a loop, without a human approving each step. When Goldman deploys an agent to handle parts of its software development lifecycle, or when Intuit uses agents to navigate customer support workflows, they are not asking a model for suggestions. They are delegating execution.
That distinction matters enormously for how you think about governance, risk, and organizational design.
The Goldman Sachs pilot, as reported, focused on code generation and review, agents that could write, test, and flag issues in software without a developer approving every line. The productivity numbers cited were significant. But the more interesting signal is not the productivity gain. It is the fact that Goldman was willing to put agents into a workflow where errors have real consequences. That is a statement about their confidence in the control layer, not just the model.
Intuit's deployment focused on customer-facing workflows, agents that could navigate complex tax and financial questions, escalate appropriately, and resolve issues without routing everything to a human agent. Again, the productivity story is interesting. But the governance story is more important. Intuit operates in a regulated environment. Deploying agents in that context requires a level of auditability and explainability that most organizations have not built.
Here is what I think these pilots actually tell us.
First, the organizations that are moving fastest on agents are the ones that already had strong data infrastructure and governance practices. Goldman and Intuit did not deploy agents into chaos. They deployed them into environments with clear data ownership, well-defined processes, and existing audit trails. If your organization does not have those things, agents will not fix them, they will amplify the dysfunction.
Second, the human-in-the-loop question is more nuanced than most frameworks acknowledge. Neither Goldman nor Intuit removed humans from the process. They changed where humans intervene. Instead of approving every action, humans set the parameters, monitor outcomes, and handle exceptions. That is a fundamentally different operating model, and it requires different skills from your team.
Third, the failure modes of agents are different from the failure modes of traditional software. Traditional software fails predictably. An agent can fail in ways that are hard to anticipate because its behavior emerges from the interaction between the model, the tools it has access to, and the environment it is operating in. Building robust agent systems requires thinking carefully about what happens when things go wrong, not just what happens when they go right.
For organizations that are not Goldman Sachs or Intuit, the question is not whether to deploy agents. The question is where to start and how to build the foundation that makes agent deployment safe and effective. That means investing in data quality, process documentation, and governance infrastructure before you touch the agent layer. It means being honest about which processes are actually well-defined enough to delegate to an autonomous system. And it means building the organizational muscle to monitor, evaluate, and course-correct agent behavior over time.
The Goldman and Intuit pilots are not a green light for everyone to deploy agents everywhere. They are a proof point that agents can work in high-stakes environments, if you build the right foundation first. That is the lesson worth taking.