Goldman Sachs is expanding its use of generative artificial intelligence, rolling out systems built on Anthropic’s Claude model to support trade accounting and client onboarding operations.
The move, reported by American Banker, reflects a broader shift among major financial institutions toward using AI to streamline back-office functions. While banks have experimented with generative AI in research, coding, and customer support, applying it to operational workflows such as compliance checks and reconciliation marks a more recent development.
Penny Crosman
Moving Beyond Knowledge Work
Large banks have already integrated AI into knowledge-based tasks. JPMorgan Chase provides employees with access to large language model tools for research and data analysis. Bank of America uses its Erica assistant to handle internal HR and IT queries. Meanwhile, both Goldman and Citi have deployed AI tools to help software developers write and test code.
Goldman’s latest initiative goes deeper into operational infrastructure. The focus is on tasks that traditionally require large teams to review documents, reconcile transactions, and ensure compliance with regulatory standards.
Tackling the “Edge Cases”
In banking operations, many processes are rules-based. Systems collect data, validate it against internal and external records, and generate required documentation. In theory, conventional software already automates much of this work.
But according to Goldman’s Chief Information Officer Marco Argenti, even if automated systems handle the majority of transactions, a small percentage fall outside defined rules. At the scale of a global bank, those exceptions can number in the thousands.
Know-your-customer (KYC) checks offer a clear example. Minor discrepancies in identity documents or near-expiry paperwork can trigger manual review. These so-called edge cases require contextual judgment rather than simple rule matching.
Argenti argues that neural networks like Claude can help resolve these micro-decisions. Instead of replacing existing systems, generative AI operates alongside them, narrowing the number of transactions that need human intervention. The goal is to shorten resolution times while preserving oversight.
From Coding Assistant to Operational Engine
Goldman’s experience using Claude internally for software development helped shape its broader AI strategy. Developers work with a Claude-based system integrated with Cognition’s Devin agent. Engineers define specifications and compliance constraints, the agent generates code, and humans review the output. The system also runs automated tests.
This setup has reshaped developer workflows, increasing productivity and accelerating project timelines. Encouraged by these results, Goldman extended similar AI capabilities to operational teams.
For trade accounting and onboarding, project leaders studied existing workflows with domain experts to identify bottlenecks. The resulting AI agents review documents, extract relevant entities, determine whether additional paperwork is required, analyze ownership structures, and initiate compliance checks when needed.
These document-heavy tasks often involve comparing fragmented data across internal ledgers, counterparty confirmations, and bank statements. By automating extraction and preliminary assessment, the AI reduces time spent on repetitive comparison work.
Balancing Automation and Oversight
Industry analysts emphasize that AI operates within a controlled layer rather than replacing core systems. Indranil Bandyopadhyay, principal analyst at Forrester, notes that accounting and compliance platforms remain the official systems of record. Claude functions within the workflow layer, handling extraction and comparison while human analysts review exceptions.
Jonathan Pelosi, Anthropic’s head of financial services, says Claude is designed to surface uncertainty and provide source attribution, creating an audit trail intended to reduce the risk of inaccurate outputs. Analysts stress that human validation remains critical, particularly in regulated environments.
Argenti rejects the idea that AI systems are inherently easier to deceive than people. He points out that social engineering exploits human vulnerabilities, while AI can identify subtle anomalies across large datasets. Still, Goldman’s approach combines automated scrutiny with human judgment rather than relying on AI alone.
A Measured Shift in Banking Operations
The deployment underscores a cautious but deliberate shift in how banks apply generative AI. Instead of flashy customer-facing experiments, institutions are targeting operational efficiency, accelerating document processing, and reducing exception-handling time in high-volume workflows.
For global banks operating under strict regulatory frameworks, the balance is clear: AI can enhance productivity and expand operational capacity, but human oversight and established systems remain central.
As financial institutions continue refining these tools, the question is no longer whether AI belongs in banking operations, but how deeply it will integrate into the workflows that keep global finance running.
