Large financial institutions are accelerating experiments with generative AI, and Goldman Sachs is now scaling the anthropic claude platform across several back-office workflows.
Goldman Sachs moves generative AI into the back office
Goldman Sachs plans to deploy Anthropic’s Claude model in trade accounting and client onboarding, positioning the rollout as part of a broader push among large banks to use generative AI for efficiency gains. The initial emphasis is on operational processes that sit in the back office and historically relied on large teams handling document review, reconciliation, and compliance checks.
Several banks already apply generative AI to knowledge work. JPMorganChase gives employees access to a large language model suite for information retrieval and data analysis. Moreover, the Bank of America uses its Erica assistant to answer internal technology and human resources questions. Citi and Goldman both rely on AI tools to support developers with coding tasks, highlighting that early deployments focused more on research and software development than operations.
However, the American Banker report notes a newer trend: using generative AI for operational activities such as trade accounting and know-your-customer (KYC) checks. This marks a shift from purely analytical use cases toward automating transaction-heavy workflows that directly affect daily banking operations.
Automating the edge cases in KYC and reconciliation
Many automatable banking processes are rules-based, involving data collection, validation against internal and external databases, and the creation of required documentation. In theory, traditional software already handles much of this work. However, Marco Argenti, Goldman’s chief information officer, argues that even if a rules-based platform resolves most cases, a small percentage of transactions fall outside predefined parameters and create thousands of exceptions at scale.
He cites identity verification in KYC compliance as a typical example. Minor discrepancies in client records or documents close to their expiry date can generate edge cases requiring human judgment. Moreover, these exceptions tend to cluster in high-volume environments, making manual review expensive and slow.
Argenti says neural networks can tackle these micro-decisions because they apply contextual reasoning where fixed rules are missing or ambiguous. In this setup, generative AI augments existing rules engines rather than replacing them. Operational gains arise from shrinking the share of cases that require manual intervention, which in turn shortens the time needed to resolve exceptions and improves straight-through processing.
Lessons from AI-assisted software development
Goldman’s earlier work with Claude for internal software development informed its decision to extend AI into other operational domains. Developers at the bank use a version of Claude combined with Cognition’s Devin agent to support programming workflows. In this process, human engineers define specifications and regulatory constraints, the agent generates code, and developers then review and refine the output.
The Devin agent also runs code tests and validations. Argenti describes this setup as a structural change to developers’ workflows, with AI agents operating under clearly defined instructions. Moreover, the combination of specification-driven coding and automated testing has increased developer productivity and shortened project completion times.
This experience convinced Goldman that AI agents can safely handle tightly scoped tasks within a regulated environment, as long as responsibilities are clearly split between humans and systems. That said, the human review layer remains central, particularly when outputs have regulatory or risk implications.
From coding to document-heavy operational workflows
For trade accounting and client onboarding, Goldman and Anthropic project leaders first observed existing workflows with domain experts to locate bottlenecks. The implemented AI agents now review documents, extract entities, determine whether additional documentation is necessary, assess ownership structures, and trigger further compliance checks where appropriate. These tasks are typically document-heavy and require individual judgment, making them suitable for AI-assisted decision support.
By automating extraction and preliminary assessment, the agents cut the time analysts spend on manual comparison work. However, they do not replace final decision-making. Instead, they present structured data and suggested next steps, allowing specialists to focus on complex or high-risk cases rather than routine file handling.
Indranil Bandyopadhyay, principal analyst at Forrester, explains that reconciliation in trade accounting requires comparing fragmented data across internal ledgers, counterparty confirmations, and bank statements. A typical workflow depends on accurate extraction and matching of figures and text from multiple documents. Here, anthropic claude is positioned as a way to handle this document-intensive matching step at scale.
Why Claude fits reconciliation and onboarding use cases
Bandyopadhyay notes that Claude’s ability to process large context windows and follow detailed instructions makes it well suited to complex reconciliation workflows. For client onboarding, analysts must parse passports and corporate registration files, then cross-reference all sources. Moreover, the need to interpret unstructured documents adds complexity that traditional rules-based tools struggle to manage efficiently.
In this environment, AI’s capacity to extract structured data, highlight inconsistencies, and flag missing documents offers a strong fit. The result is reduced overall workload for analysts and a faster onboarding cycle for clients, while still maintaining the governance standards required in banking.
Crucially, Bandyopadhyay emphasizes that accounting and compliance platforms remain the canonical systems of record. Claude sits in the workflow layer, responsible for extraction and comparison, while human analysts handle the exceptions that the code surfaces. In his view, the operational value in heavily regulated sectors such as banking lies in this division of labor rather than in full automation.
Risk management, uncertainty and human oversight
Jonathan Pelosi, head of financial services at Anthropic, says Claude is trained to surface uncertainty and provide source attribution, creating an audit trail that reduces the effect of hallucinations. Moreover, these design choices aim to make AI behavior more transparent to risk teams and regulators by linking outputs to their supporting evidence.
Bandyopadhyay also highlights the importance of human oversight and validation, urging institutions to design systems so that errors are detected early in the workflow. That said, he acknowledges that when properly monitored, AI agents can handle a large share of repetitive checks and comparisons far more quickly than human staff.
Goldman’s Marco Argenti rejects the idea that AI systems are inherently easier to deceive than humans. He argues that social engineering attacks primarily exploit human vulnerabilities, whereas AI models can detect subtle anomalies at scale. However, he reiterates that the optimal setup combines human judgment with automated scrutiny in integrated teams.
Implications for banking operations
According to Argenti, this combination implies a significant increase in operational capacity without proportional increases in staffing, even given the known issues around AI deployment. Moreover, it allows banks to manage growing regulatory and documentation burdens while keeping headcount growth under control.
Across the banking sector, generative AI is emerging as a tool to improve operational performance by accelerating document processing, reducing exception-handling times, and increasing throughput in high-volume workflows. However, the continued need for human oversight means institutions must retain their existing systems of record and governance structures, using AI primarily to streamline the layers that sit on top of them.
In summary, Goldman’s work with Claude and related agents suggests a pragmatic model for generative AI in finance: automate document-heavy, rules-adjacent tasks; surface exceptions clearly; and keep human experts ultimately responsible for critical decisions and regulatory compliance.
Source: https://en.cryptonomist.ch/2026/02/17/anthropic-claude-backoffice-automation/
