Goldman Sachs Deploys Anthropic’s Claude AI for Trade Accounting and Client Onboarding

Goldman Sachs is moving forward with the deployment of Anthropic’s Claude AI model in key operational areas, including trade accounting and client onboarding. According to a report in American Banker, this initiative is part of a larger trend among major banks adopting generative artificial intelligence to improve efficiency and streamline operations.

The focus is not on customer-facing services, but rather on back-office functions, areas traditionally handled by large teams performing tasks such as document review, reconciliation and compliance checks.

Expanding AI Use in Banking

Many major financial institutions are already using generative AI in various knowledge-based tasks:

• JPMorgan Chase provides employees with access to large language models for information retrieval and data analysis.
• Bank of America uses its Erica assistant to answer internal technology and HR-related questions.
• Citigroup and Goldman Sachs both use AI tools to assist developers with coding tasks.

However, what’s new is the shift toward applying generative AI to operational workflows like trade accounting and know-your-customer (KYC) processes.

Automating the “Edge Cases”

In banking, many operational processes are rules-based. These processes typically involve:

• Collecting data
• Validating information against internal and external databases
• Compiling required documentation

While traditional software has automated much of this work, challenges remain.

Marco Argenti, Goldman Sachs’ Chief Information Officer, explains that even if a rules-based system handles the majority of transactions, a small percentage fall outside predefined rules. At large scale, even a small percentage can result in thousands of exceptions requiring manual review.

For example, in KYC identity verification, small discrepancies or documents nearing expiration can create edge cases that require judgment.

According to Argenti, neural networks are better suited to handle these micro-decisions because they apply contextual reasoning where rigid rules fall short. In this model, generative AI does not replace rule-based systems — it enhances them.

The result is fewer cases requiring manual intervention and faster resolution of exceptions.

Lessons from AI in Software Development

Goldman’s decision to expand Claude into operations was influenced by its earlier success in software development workflows.

Developers at Goldman use a version of Claude integrated with Cognition’s Devin agent. In this setup:

• Human developers define specifications and regulatory constraints
• The AI agent generates code
• Humans review and validate outputs
• The agent also runs tests and validations

This approach changes developer workflows by introducing AI agents that operate under clearly defined instructions. The outcome has been increased productivity and faster project completion.

Applying Claude to Trade Accounting and Client Onboarding

Before implementing Claude in operational roles, Goldman and Anthropic teams observed existing workflows alongside domain experts to identify bottlenecks.

The AI agents now:

• Review documents
• Extract key entities
• Determine if additional documentation is needed
• Assess ownership structures
• Trigger additional compliance checks when necessary

These tasks are document-heavy and require contextual judgment. By automating document extraction and preliminary analysis, AI reduces the time analysts spend comparing and cross-referencing information.

Why Claude Fits These Workflows

Indranil Bandyopadhyay, principal analyst at Forrester, explains that trade accounting reconciliation involves comparing fragmented data from:

• Internal ledgers
• Counterparty confirmations
• Bank statements

Accurate extraction and matching of figures and text are critical. Claude’s large context window and ability to follow structured instructions make it suitable for these tasks.

Similarly, client onboarding requires parsing passports, corporate registration documents and cross-referencing multiple sources. AI’s strength in structured data extraction and inconsistency detection makes it effective in reducing manual workloads.

Importantly, accounting and compliance platforms remain the official systems of record. Claude operates at the workflow layer, handling extraction and comparison, while human analysts manage exceptions.

In regulated industries like banking, this division of labor is considered operationally valuable.

Managing Risk and Ensuring Oversight

Jonathan Pelosi, head of financial services at Anthropic, states that Claude is trained to surface uncertainty and provide source attribution. This helps create an audit trail and reduces the risk of hallucinations.

Bandyopadhyay emphasizes that human oversight remains essential. Systems must be designed to detect errors early in the workflow.

Argenti also challenges the idea that AI systems are inherently easier to deceive than humans. He argues that social engineering often exploits human vulnerabilities, while AI can detect subtle anomalies at scale. Still, he stresses the importance of combining automated scrutiny with human judgment.

This hybrid model allows banks to increase operational capacity without proportionally increasing staffing levels - even while acknowledging known challenges in AI implementation.

The Bigger Picture: AI in Banking Operations

In the banking sector, generative AI is emerging as a tool for:

• Accelerating document processing
• Reducing exception handling time
• Increasing throughput in high-volume workflows

At the same time, human oversight and established systems of record remain central to operations.

Rather than replacing existing infrastructure, AI like Claude operates as a workflow enhancement layer, improving efficiency while keeping compliance and accountability intact.