Goldman Sachs is moving its use of artificial intelligence beyond conversational assistants toward agentic AI systems capable of performing multi-step tasks across regulated banking workflows. The development marks a shift from AI tools that primarily assist employees to systems that can independently execute defined operational activities under human supervision.
The bank first introduced its GS AI Assistant in January 2025 to approximately 10,000 employees, initially supporting activities such as summarisation, proofreading and code translation. By June 2025, the assistant had been expanded across a wider group of knowledge workers. According to the report, Goldman Sachs subsequently began developing autonomous AI agents with engineers from Anthropic.
By February 2026, Goldman Sachs CIO Marco Argenti indicated that autonomous agents were approaching production deployment. The bank is targeting areas characterised by repetitive, documentation-heavy processes, particularly trade break resolution, client due diligence and onboarding packet preparation.
The emerging agentic system is designed to retrieve information from multiple systems, organise supporting evidence and draft required documentation. Human employees remain responsible for reviewing exceptions and approving critical actions, particularly where regulatory or financial consequences are involved.
Goldman Sachs is also using a multi-model architecture in which different AI models can be selected according to the requirements of a particular task. The report states that Anthropic’s Claude is being used for some of the new agentic applications, while the bank has previously experimented with Devin, an autonomous coding agent.
Given the sensitivity of financial information, governance and security remain central to the deployment. The report says Goldman Sachs is using private cloud environments, access controls and encryption to protect regulated data. Agent activity is also logged, while validation layers are intended to identify unsupported outputs or hallucinations.
Human approval continues to form an important control layer. For example, agent-generated actions affecting financial records are subject to review before posting, while the bank is also using continuous red-team testing to identify potential weaknesses and edge cases.
The bank views AI agents primarily as capacity multipliers rather than direct replacements for employees. Automation of document gathering, reconciliation and routine checks could allow staff to shift towards exception management, client engagement and higher-value analytical work.
The move reflects a broader transformation in financial services, where banks are exploring agentic AI for functions that traditionally require substantial manual processing. However, the source notes that important details, including precise production dates, measurable efficiency gains and cost savings from the rollout, have not yet been publicly disclosed.
For regulated financial institutions, Goldman Sachs’ approach highlights the growing importance of combining AI automation with auditability, human oversight, validation controls and model-risk governance as autonomous systems move closer to core banking operations.
