Goldman Sachs Partners with Anthropic to Build AI Agents for Core Operations
Goldman Sachs is working with AI startup Anthropic to build autonomous agents for internal operations, according to a CNBC report the bank confirmed. The collaboration has been in motion for roughly six months, with Anthropic engineers embedded alongside Goldman's teams.
Marco Argenti, Goldman Sachs' chief information officer, said the agents will focus on trade and transaction accounting, client due diligence, and onboarding. The goal is simple: shorten cycle times on high-volume, rules-heavy processes without sacrificing control.
What These AI Agents Will Actually Do
- Trade and transaction accounting: prepare entries, reconcile breaks, flag anomalies for review.
- Client due diligence: gather documents, validate data across systems, monitor changes.
- Onboarding: pre-fill forms, trigger checks, track status, and escalate exceptions.
The systems are being built on Anthropic's Claude model and related tooling. Anthropic's push into enterprise use cases, including products like Claude Cowork, lines up with this kind of back-office automation. Learn more about Claude.
Why Operations Teams Should Care
Argenti told CNBC the tech is expected to materially reduce processing time on core workflows. For operations leaders, that means faster throughput, fewer manual touchpoints, and cleaner audit trails-if the deployment is done with the right controls.
- Throughput: agents can run continuously and in parallel, smoothing spikes in volume.
- Quality: consistent application of rules, with fewer keying errors and missed checks.
- Auditability: structured logs of every step, input, and decision for review.
- Cost: less time spent on repetitive tasks, more time on exceptions and analysis.
How the Collaboration Is Set Up
Goldman Sachs engineers are building alongside Anthropic's technical staff embedded on their teams. That tight loop helps translate policies, controls, and edge cases into agent behaviors that hold up in production.
The bank is still early in development, and a measured rollout is expected. Financial operations require rigorous testing, model validation, and compliance sign-off before agents touch live data.
Risks and Controls to Plan For
- Data protection: strict PII handling, data minimization, and environment isolation.
- Model accuracy: guardrails, confidence thresholds, and automatic escalation on ambiguity.
- Human-in-the-loop: clear approval points for high-risk steps and material changes.
- Exception handling: well-defined workflows, queues, and SLAs for break resolution.
- Monitoring: real-time metrics for latency, error rates, and policy breaches.
- Audit: immutable logs of inputs, outputs, and rationale; easy retrieval for reviewers.
- Governance: documented policies, periodic reviews, and independent model risk checks.
- Integration: API-first design, idempotent operations, and rollback plans.
Timeline and What to Expect Next
Goldman Sachs plans to launch the agents soon, but no specific date was shared. Expect pilot-first deployments, progressive scope increases, and close involvement from risk, compliance, and audit before broader rollout.
Action Plan for Operations Leaders
- Identify 3-5 candidate workflows: high volume, clear rules, measurable outcomes.
- Map controls: define approval gates, exception criteria, and documentation standards.
- Set metrics: turnaround time, first-pass yield, exception rate, and cost per transaction.
- Create runbooks: inputs required, decision trees, escalation paths, and rollback steps.
- Pilot with guardrails: start with synthetic or sampled data; expand by segment.
- Train teams: how to supervise agents, verify outputs, and tune prompts/policies.
- Close the loop: weekly reviews on errors, drift, and new edge cases; update rules fast.
- Coordinate with compliance: model documentation, access controls, and retention rules.
Workforce Impact
Banks consistently say the aim is to augment human work, not replace judgment. Expect capacity to shift from routine processing to exception handling, client support, and control improvements.
If your team is leaning in, upskilling pays off. See our resources for operations and finance teams: AI Certification for Claude and AI Tools for Finance.
Bottom Line
Goldman Sachs and Anthropic are building agent-driven workflows for core banking operations. If they hit the efficiency gains they're aiming for, the playbook will be clear: start small, lock down controls, measure relentlessly, and scale what works.
Your membership also unlocks:
