Financial agents are becoming an industry in their own right in China. Caiyue Xingchen, a financial AI venture jointly founded by Jiemian Cailianshe — the financial-news operator behind the Star Market Daily (Kechuangban Daily) — and model developer StepFun (Jieyue Xingchen), formally launched FinWork V1.0 at a launch event in Changsha on November 20, according to Star Market Daily and multiple Chinese media reports.
The product is pitched not as a chatbot bolted onto banking workflows but as a full-stack agent workbench for the financial industry. At its entry point sits an AI desktop assistant; underneath is a "three-layer usage, three-layer access" architecture. Individuals get a personal assistant with more than 220 built-in financial skills, financial-data MCP tools and a dozen-plus professional charting capabilities. Departments can codify a star employee's workflow into shareable, reusable skills. The organization layer is an enterprise agent platform with sandboxed execution, behavior monitoring and full-lifecycle management of every agent.
The access layers are aimed at different audiences: an API surface for developers who want fine-grained capabilities, an SDK that packages financial business semantics into modular integrations, and a configuration layer that lets non-technical departments assemble their own workbenches without writing code. The company says the platform's financial search covers market quotes, announcements, research reports and investor-relations sources, with trust ratings assigned per source.
Trust and auditability get the most design attention, which reflects the realities of selling AI into regulated brokerages. All inputs are converted into a standardized evidence-chain format of entity, attribute, relation and value; outputs support cross-checking across event entities; and every key figure and claim can be traced back to its source and the tool calls that produced it. Permissions follow three principles — independent entitlement, minimal entitlement and time-limited authorization — with policy set before execution, sandbox monitoring during, and full-chain audit of tasks and resource consumption afterward.
The deployment numbers, which come from the company and its clients rather than independent audits, are the concrete part: anchor customer Guotai Haitong Securities has reportedly put more than 1,300 agents into production covering more than 260 scenarios, and Sinolink Securities and Industrial Securities have also deployed at scale. FinWork claims a client portfolio analysis takes about three minutes, and a single generated content item costs roughly 0.03 yuan and eight seconds. Alongside the launch, Caiyue Xingchen signed a strategic partnership with ZStack (Yunzhou Technology) to co-develop a financial agent gateway and compute-scheduling scheme, aimed at banks and brokers that require data to stay in-domain, full private deployment and adaptation to China's domestic (Xinchuang) hardware stack.
Why it matters: China's brokerages are moving from AI experimentation to fleet deployment, and the interesting question has shifted from model capability to governance — who audits an agent's output, what data it may touch, and how a regulator can replay its decisions. Industry participants at the event identified compliance, talent, data and operational audit as the four core bottlenecks for scaling financial agents. FinWork's bet is that a news-data-anchored evidence chain plus hard permission controls is what gets an agent past a brokerage's risk committee — a bet Alibaba's Qianwen Work and ByteDance's forthcoming assistant are making in parallel, each with a different home-field advantage.
Comments (0)
Log in to join the discussion
Log InNo comments yet