The costly part of a highly automated banking process is often the small percentage that refuses to follow the rules. At Goldman Sachs scale, even a thin layer of exceptions can mean thousands of documents, discrepancies, and decisions that require human attention.
The bank plans to deploy Anthropic’s Claude in trade accounting and client onboarding. The model will work alongside existing rules and systems of record, handling document extraction, comparison, and preliminary assessment while analysts retain responsibility for uncertain or consequential cases.
This is a more precise ambition than replacing a complete back-office function. Goldman is targeting the messy boundary where conventional automation stops: identity records that do not match perfectly, documents nearing expiry, fragmented ledger entries, and ownership structures that require context.
Neural Models Fill Gaps Between Fixed Rules
Traditional automation performs well when inputs and outcomes are defined. A rules engine can validate most transactions, but it cannot anticipate every variation in a passport, corporate registration, counterparty confirmation, bank statement, or internal ledger.
Goldman chief information officer Marco Argenti argues that neural networks can make these “micro-decisions” by applying contextual reasoning where rules do not give a clear answer. In client onboarding, an agent can extract entities, identify missing documents, assess ownership information, and trigger additional compliance checks.
In trade accounting, the work involves reconciling figures and text across fragmented sources. Forrester analyst Indranil Bandyopadhyay points to large context windows and instruction-following as useful capabilities for reviewing this material.
The benefit is not that the model becomes the bank’s ledger or compliance authority. It reduces comparison work and sends a smaller, better-described set of exceptions to analysts.
Coding Experience Provides an Operating Template
Goldman’s experience using Claude with Cognition’s Devin agent in software development influenced the expansion. Developers set specifications and regulatory parameters, the agent writes and tests code, and humans review the output.
That pattern transfers well to operations: domain experts define acceptable behaviour, the agent completes bounded tasks, validation catches errors, and authoritative systems retain the final record.
Goldman and Anthropic teams observed workflows with employees before selecting bottlenecks. This is essential because a process diagram rarely captures the informal checks, workarounds, and judgment that keep a regulated operation functioning. Automating the documented step without understanding those practices can simply move work elsewhere.
The bank should measure complete outcomes, including resolution time, false flags, missed discrepancies, analyst review effort, and downstream corrections. A model that processes documents quickly but creates extensive verification work may only hide the cost.
Comparisons should also include the exceptional cases that generate the greatest regulatory and operational exposure.
Auditability Matters More Than Apparent Confidence
Anthropic financial-services head Jonathan Pelosi says Claude is designed to surface uncertainty and provide source attribution. Those features can support an audit trail, but institutions still need to verify how reliably they work in the specific workflow.
High-risk exceptions should retain human approval, and every extracted fact should link back to its source document. Access should be limited to the information required for the case, while prompts, outputs, decisions, and later corrections remain reviewable.
Accounting and compliance platforms must stay canonical. If a generative system can silently alter the record it is supposed to analyse, the division of responsibility collapses.
Goldman’s approach illustrates where generative AI may deliver its strongest banking value. It does not discard deterministic software; it extends automation into the ambiguous residue that deterministic software leaves behind.
The achievement will not be a back office without people. It will be an operation in which machines handle more preparation and comparison, while human attention concentrates on the cases where judgment changes the outcome. In regulated finance, that narrower claim is both more credible and more valuable.
