AGTNEWS
Goldman is running autonomous AI agents inside compliance and onboarding, with humans only reviewing the output. The shift moves the decision from the analyst to the model, and the supervisor's evidence has to follow.
Goldman Sachs has begun running autonomous AI agents inside two workflows where the bank has historically paid humans to be the slowest, most careful part of the process: trade and accounting support, and client vetting and onboarding. The agents open cases, pull the supporting evidence, and draft the action; a human reviewer signs off after the fact. Goldman CIO Marco Argenti called the configuration a "digital co-worker for many of the professions within the firm that are scaled, are complex and very process intensive" in a Feb. 6, 2026 interview with CNBC.
In a traditional compliance workflow, the analyst reads the underlying document, forms a judgment, and writes the case file. The case file is the artifact the regulator, the internal auditor, and the bank's own second line of defense can reconstruct. In the configuration Goldman is describing, the model reads the document, drafts the case file, and the human reviewer checks the draft. The decision has moved up a level; the audit trail has to follow.
That is the part neither Goldman nor its peers have explained on the public record. No go-live date has been disclosed. The agents remain in what Argenti described to CNBC as "early stages," with a launch "soon." Reuters confirmed the collaboration on the same day, and the Observer profile of Argenti described Anthropic engineers as embedded with the bank's teams for roughly six months, co-developing the agents. The model is Anthropic's Claude, layered behind a routing interface that can call other language models, running in Goldman's private cloud with encryption and access controls.
The strongest counterargument is also the one the bank is leaning on: the human reviewer is still in the loop. A reviewer who reads the underlying document and rejects bad drafts is not a rubber stamp, and a senior officer with sign-off authority remains the named accountable person under U.S. bank-supervision rules. The configuration is closer to model-assisted work than to model-decided work, and U.S. bank supervisors have not publicly objected to the former.
The harder question is what the reviewer has to show they checked. In the old configuration, the analyst's notes were the audit trail. In the new one, the reviewer is asked to attest to a draft they did not write, on evidence the model selected and organized. The bank has not said what evidence the reviewer must produce to demonstrate substantive review, and the supervisor has not said what level of review will satisfy an exam finding. That gap is the audit story, not the launch.
Goldman's prior work supports the seriousness of the underlying bet. The bank launched its GS AI Assistant in January 2025 to roughly 10,000 staff for summarization, proofreading, and code translation, and expanded the rollout in a June 23, 2025 memo. The bank also ran what CNBC called the first major autonomous coder pilot on Wall Street in 2025, using Cognition Labs' Devin to write and ship production code under human review. The compliance deployment is the same shape of pilot extended into a regulated function.
The productivity case is narrower than the launch language suggests. A Goldman analysis reported by Fortune on March 3, 2026 found no overall relationship between AI use and productivity at the bank, and a roughly 30% boost in two specific use cases. A Forbes update on Aug. 6, 2026 said the agentic push has since expanded into software engineering at scale. The figures are scope-bound and self-reported, and they show the tool is working in the places it has been deployed, not that it has remade the bank.
The watch items are concrete. First, whether the Federal Reserve and the Office of the Comptroller of the Currency publish any public statement on the evidence a human reviewer must produce when a model drafts a regulated decision. Second, whether any of the other G-SIBs - JPMorgan, Citigroup, Bank of America, Morgan Stanley, Wells Fargo - disclose a comparable deployment or treat Goldman's pilot as a go signal. Third, whether the headcount disclosures in Goldman's 2026 annual report show a measurable shift in the affected functions, which would be the first independent signal that the agent is doing the work, not just drafting it. The launch language is settled; the audit story is the one to watch.