Goldman's AI Agents Are Coming for Trade Ops

Goldman Sachs has had Anthropic engineers sitting inside its offices for six months, co-building autonomous AI agents for two specific functions: trade accounting and client onboarding.

This was not a pilot or a proof of concept. Actual engineers from an AI company, embedded in one of the world's largest investment banks, building agents that will do work currently done by people.

Goldman CIO Marco Argenti told CNBC the bank was "surprised" at how capable Claude was at tasks beyond coding. Specifically, parsing large data volumes while applying rules and judgment. The agents are expected to launch "soon."

Let that land for a second.

Why trade accounting and onboarding?

These are not glamorous functions. They are not the parts of banking that make headlines or attract talent. Trade accounting is the work of reconciling transactions. Matching what was supposed to happen with what actually happened, across hundreds of thousands of records, in multiple currencies, across counterparties who use different systems and different conventions.

Client onboarding is the work of vetting new counterparties. Pulling documents, checking beneficial ownership structures, screening against sanctions lists, cross-referencing with adverse media, and producing a compliance file that satisfies regulators.

Both are manual, repetitive, high-volume, and error-prone. Both require reading unstructured documents and applying judgment. Both have been stubbornly resistant to traditional automation because the inputs are messy and the rules have exceptions.

This is exactly the kind of work where AI agents represent a genuine step change. These are not chatbots or copilots. These are autonomous agents that can execute multi-step tasks without someone holding their hand.

The gap this opens

Goldman's agents will work. The technology is clearly there. What matters is what happens when Goldman can onboard a new trade counterparty in hours instead of weeks, and reconcile a day's transactions before the operations team has finished their morning coffee.

The answer: Goldman's cost-to-serve drops. Its capacity to handle small and mid-size transactions improves. Its compliance throughput increases. And every bank that has not made this investment falls further behind.

JPMorgan is spending nearly $20 billion on technology in 2026, a $2 billion increase from last year. Its internal OmniAI platform runs over 400 production use cases. Deutsche Bank and Goldman are both deploying AI for trading surveillance.

These are not experiments. This is a structural investment cycle by the largest banks in the world. And the gap between institutions that invest at this scale and those that do not is about to become the defining competitive dynamic in trade finance.

What this means for the rest of us

Nobody at the big banks will say this out loud, but agentic AI favours scale.

If you are Goldman Sachs, you can afford to embed Anthropic engineers for six months. You can afford the compute costs. You can afford to build proprietary training pipelines on your own transaction data. You can afford to get it wrong a few times before you get it right.

If you are a mid-tier trade finance bank in the Gulf, or a specialist commodity lender in Singapore, or a regional bank in West Africa doing LC processing? You cannot.

This does not mean smaller institutions are doomed. It means the path to AI adoption in trade finance will look different depending on where you sit. The large banks build custom. Everyone else will need to buy or partner. And the vendors who serve this market are only beginning to appear.

The practical implications for trade finance:

LC processing times compress. When an AI agent can check documents against UCP 600 requirements, cross-reference the LC terms, and flag discrepancies, the 3-5 day document examination cycle starts looking unnecessary. Goldman will not be the bank that takes five days to examine your documents.

KYC bottlenecks ease. For some. Onboarding a new trade counterparty currently takes weeks to months at most global banks. AI agents that can pull, parse, and verify corporate documents across jurisdictions could collapse that timeline. But only for banks that deploy them.

The cost of compliance becomes a competitive weapon. Today, compliance is a cost that every bank bears roughly proportionally. When AI makes compliance dramatically cheaper at scale, the banks that deploy it fastest gain a structural cost advantage that compounds with every transaction.

The workforce question

There is an obvious follow-up that trade finance professionals are already asking: what happens to the people who currently do this work?

Goldman has not said. Nobody has. But the direction is clear. When David Solomon talks about an AI "supercycle" and the bank embeds AI company engineers to automate operations functions, the trajectory is not towards hiring more operations staff.

The World Economic Forum noted in February that "banking enters the agentic era." The phrasing is telling. They did not call it the digital era or the automation era. They called it the agentic era. AI systems that act autonomously rather than merely assist.

For trade finance practitioners, the immediate priority is not existential worry. It is understanding which parts of your work are most exposed. Document checking, data reconciliation, compliance screening, and report generation are the first functions to be automated. Relationship management, deal structuring, risk judgment on novel situations, and regulatory negotiation are the last.

The middle ground, the work that requires some judgment but follows established patterns, is where the disruption happens fastest. That is most of trade operations.

Goldman just showed everyone the direction. The rest of the industry gets to decide how quickly to follow.

-- Tamara