China’s financial institutions are moving AI agents from demonstrations toward bounded business work, with examples spanning cross-border settlement, credit due diligence, research, trading support, and enterprise risk management. The activity described around the fifth Global Digital Trade Expo in Hangzhou, held September 23–27, 2026, is not evidence that every bank or financial firm has completed a broad rollout. It does show where institutions are testing practical uses, and where governance is being built alongside them.
A 21世纪经济报道 (21jingji / 21世纪经济网) report, whose page credits the China Business Journal, described selected deployments and discussions at the event. The examples point to a measured shift: institutions are using large models and intelligent agents to handle parts of established workflows, while retaining controls around sensitive financial decisions.
ICBC Zhejiang maps agents onto settlement and due diligence
The clearest operating example comes from the Industrial and Commercial Bank of China’s Zhejiang Branch. Its ICBC Zhiyong system is being used to explore desktop intelligent-agent business partners. The primary report says the branch had deployed more than 30 localized application scenarios across more than 10 business areas.
One cross-border settlement application, digital-intelligent customs clearance, was described as digitizing the chain from information assessment through regulatory filing. That framing matters because the system is associated with a defined sequence of work rather than an open-ended promise that an agent can manage an entire banking relationship without supervision.
The report attributed a 50% improvement in due-diligence efficiency to the branch’s due-diligence report assistant. That figure applies to the named use case. It is not an independently audited result, a bank-wide measure, or evidence that other financial institutions would achieve the same outcome.
Separately, ICBC’s official product page describes the broader platform. The ICBC platform page identifies 工银智涌平台 as an application-development platform containing a financial-industry large model enhanced through training on substantial financial data, alongside a large-model development platform. ICBC says the platform reflects experience from more than 20 business lines and more than 200 scenarios.
The official page lists functions including automated drafting of due-diligence, credit-approval, and anti-money-laundering reports. It also describes data queries, knowledge-base business questions and answers, document summarization, editing, expansion, and multilingual translation. The page lists those functions but does not supply a measured productivity result.
The distinction between the branch’s deployment figures and the official platform description is important. The first provides a snapshot of localized applications and a stated efficiency result. The second describes the product’s capabilities and business experience. Neither establishes a sector-wide adoption rate.
TradeMate and intelligent branches widen the banking test bed
The reported examples extend beyond ICBC. Hangzhou Bank was described as focusing on branch-hall construction and interaction through a new generation of lightweight equipment. Its technical route covers the operating system, hardware peripherals, security controls, and a third-party ecosystem. The report said the bank was integrating terminal and data intelligence to expand the time and space of branch services, but it did not provide a quantified outcome or a completed rollout measure.
That approach places AI in a physical service environment rather than only in back-office software. The report describes equipment, operating systems, hardware peripherals, security controls, and a third-party ecosystem. It does not establish customer numbers, revenue, or satisfaction gains, so the significance remains an implementation direction rather than a measured business result.
Zheshang Futures offers a different model. The company built TradeMate, a dedicated financial AI assistant described as a super AI agent. The report said TradeMate connects to the company’s professional data sources and links data, analysis, and automation workflows. It can automatically perform data processing and analysis for research, trading, and strategy work.
TradeMate should not be read as an autonomous trading system. Zheshang Futures chairman Hu Jun said the company embeds research results into basis trading and over-the-counter options and develops hedging and risk-management solutions for businesses. The sourced description supports assistance with data, analysis, and workflow automation. It does not support claims that TradeMate independently makes investment decisions, places trades, manages client assets, or operates without human review.
These examples center on fitting models into institutional workflows rather than simply adding a chatbot to a financial website. EastFrontier’s report on Citic Baixin Bank’s software pipeline covers a separate bank and does not substantiate the deployments described here.
Risk controls are advancing with deployment rules
The primary report attributes several risk-control capabilities to AnJie Broad Law Firm. It says stress testing and early-warning models can measure changes in portfolio exposure; transaction-fraud models can score transactions in real time; and fund-network analysis can identify suspicious transactions and assist with suspicious-transaction reports.
Those descriptions are capabilities or application possibilities, not verified outcome metrics. No reduction in defaults, fraud losses, or money-laundering incidents is established by the supplied sources. The regulatory material emphasizes governance architecture, data compliance, model risk, explainability, and lifecycle controls.
That concern is reflected in the regulatory timetable described in the report. In June 2026, the National Financial Regulatory Administration issued guidance on the safe use of artificial intelligence in banking and insurance, identified as Jinfa [2026] No. 8. In September, the NFRA Shanghai Regulatory Bureau issued measures calling on institutions to improve AI governance architecture, include AI in digital-transformation planning, build AI talent, and strengthen data-compliance systems.
The Shanghai measures also call for controls on access to high-risk application scenarios, stronger outsourcing-risk management, improved model-risk management and explainability, and attention to hallucinations, black-box behavior, and large-model security vulnerabilities. These are guidance and regulatory measures, not a claim that the underlying texts are new statutes or that every institution has already implemented each element.
China’s banking AI story is consequently less about a single breakthrough than about operational placement. ICBC Zhejiang Branch is testing localized agents in settlement and due diligence. Hangzhou Bank is exploring intelligent equipment for branch services. Zheshang Futures is connecting an assistant to specialist data and research workflows. At the same time, regulators and legal advisers are emphasizing controls that cover data, models, vendors, access, and explanations.
EastFrontier’s separate report on Chinese banks and AI loans provides additional sector background, but it does not establish the deployments above. The examples in this article remain bounded: ICBC Zhejiang describes local applications, Hangzhou Bank describes branch equipment, and Zheshang Futures describes a specialist assistant, while the report also details governance and risk controls.