Shinhan Securities Built a 500-Agent AI Workforce on Google Cloud in Three Months, Citing a 24× Speedup on Derivatives Contracts

Shinhan Securities, one of South Korea's largest retail brokers, has built a 「next-generation AI agent platform」 running on Google Cloud that today deploys more than 500 internal agents across the firm, the two companies said on August 18. The platform, which uses Google Cloud's Gemini Enterprise agent product alongside a Shinhan-built 「hybrid orchestrator」 tailored for financial workflows, has already automated roughly 70 percent of the manual review work in OTC derivatives processing — collapsing a representative contract review from two hours to five minutes per case.

The Architecture: Rules First, Models Second

Shinhan did not simply point Gemini at its back office. The firm built a hybrid orchestrator that routes money-movement workflows — wire transfers, position changes, settlement instructions — through deterministic rule logic, and only invokes large language models for tasks that require advanced reasoning, like contract analysis. That split, the company said, is what lets it pass South Korea's 「innovative financial service」 approval and the cloud-service-provider safety assessment without giving up the model capability that made the project worth building in the first place.

Why a Broker Cares About Governance

The agent count itself is striking — more than 500 production agents in a single broker — and almost as interesting is the governance model Shinhan used to get there. Each department designated an 「AX coordinator」 who owned the agent roadmap for that team; the central AX Division then consolidated the long tail into a single Google Cloud-hosted platform so the agents could call internal financial systems through one integration surface. Shinhan said the architecture cleared cloud usage reporting and internal security due diligence in roughly three months end-to-end.

The Representative Workflow

The OTC derivatives use case is the most concrete. Staff previously had to open 50-page unstructured contracts attached to inbound emails, manually review them, and enter more than 50 individual data items into the firm's internal systems — work that took three to four people per case. With the agent platform, the workflow is triggered from the work-system inbox, the agent selects the email, parses the contract, validates the data, and writes it back to the system of record. Per-case processing time fell from two hours to five minutes, Shinhan said, and the firm now plans to push the share of back-office work done by agents past 70 percent by year-end.

What It Means for Enterprise AI

Shinhan's deployment is one of the largest single-customer Google Cloud Gemini Enterprise rollouts in financial services to date, and it lands in the same week Microsoft shipped five additional Claude capabilities inside Microsoft Foundry hosted on Azure and Google separately published a reference implementation showing how its Customer Support & Returns Agent can be hardened against prompt injection. The throughline is that the next production agents are not single-prompt LLM calls; they are rule-plus-model hybrids with governance on top, and the platforms that ship both halves fastest will capture the regulated-workload budget.