MYbank Opens Bailing 2.0 AI Agent to 42 Million Small Businesses as AI Enters Core Banking Functions

The example the bank used to demonstrate the front end involved a restaurant owner, referred to as Mr. Wang, who had run one restaurant for two years and had identified a 65-square-meter street-front site for a second. He opened Bailing and typed a single line: he wanted his credit limit raised. The agent did not treat the request as a routine limit increase. It queried him about the new store's location, size, opening date, intended use of funds and how much he could contribute himself, and reconstructed the intent behind the vague request as financing for the second outlet.

Across the exchanges, the agent worked out that the total investment would be about 600,000 yuan, with roughly 450,000 yuan for branding, renovation and equipment and 150,000 yuan for rent, first inventory and working capital. Wang would fund 200,000 yuan himself, leaving a gap of about 400,000 yuan. The agent then combined the existing store's cash flow, industry patterns, the surrounding business district and the opening plan to assess expected revenue and repayment pressure, and asked him to supply the necessary operating and purpose documents. The result was that his available limit rose from 150,000 yuan to 400,000 yuan, drawable in installments in line with the store's progress, with interest-only payments for the first six periods. The process took 10 minutes.

MYbank said the front end amounts to a messaging entry point that could not deliver results without the back-end infrastructure, skills and real-time databases behind it. It has built eight AI workbenches and disclosed four at the conference. The first, Qianliyan, is aimed at risk control that assesses individual customers rather than customer groups. The bank cited the case of a photography chain in Suzhou carrying two risk flags, a poor rating and high debt. A conventional system would have blocked the application, but the tool traced the decision chain and found the owner had expanded the same business to nine stores in six cities over two years, with monthly revenue steady above 700,000 yuan. After strategy specialists verified the evidence and confirmed the risk boundaries, the rejection became a 300,000-yuan limit with pricing and tenor recalculated for the business.

The second workbench, Dinghaizhen, encodes 25 years of approval experience. In one internal test, a woman identified as Ms. Wu, who sells storage devices online, applied through Bailing to raise her limit ahead of expanding purchases. The tool condensed her operating, financing, repayment and guarantor information into a single page within seconds and recommended no increase, citing a guarantor's overdue payment, weak support from peer lenders and several recent loan applications. The human reviewer did not follow the recommendation: the guarantor was repaying on schedule under a restructuring plan, e-commerce businesses are asset-light so weak peer support does not indicate poor trading, and the multiple applications stemmed from buying ahead of rising storage prices. The tool re-reasoned on the new facts, and Wu was approved for 1.8 million yuan. Annotations from the second review flow back into the system so similar cases are judged more accurately later.

The third, Baoliandeng, addresses demand that spikes quickly. In July, after a typhoon crossed Guangxi, a citrus grower named Lao Tan in Wuming district, Nanning, had 20 mu of orchard flooded and fruit knocked to the ground. He did not call or visit a branch; he told Bailing he had been hit by a disaster. The tool cross-checked the conversation, satellite remote sensing, operating cash flow and regional event data, concluded that the damage extended beyond Lao Tan to other growers and to upstream agricultural input dealers and downstream cold-chain logistics operators, and generated deferral and interest-free plans based on past disaster handling and local industry patterns. Within one hour, tens of thousands of customers along the chain had received the plans.

The fourth, Jindouyun, targets research and development. MYbank said its use of AI coding has doubled and that 15 percent of business modification requests are now completed by AI. The bank said this is not the loosely generated "vibe coding" often described, and that code remains subject to financial compliance requirements, standardized review, testing and auditing. It added that faster coding alone is insufficient because requirements clarification, interaction design, system analysis, testing and compliance review can become bottlenecks. Jindouyun has AI generate an interactive high-fidelity prototype during the requirements stage so product, design and engineering can work from the same page, then converts the requirement into a system analysis precise enough to identify affected data, modules and interfaces, before sub-agents proceed in parallel with automated evaluation. A weather-alert feature for an agricultural product called Nongxiaobao, developed after the system linked high-frequency user queries to an entry point buried too deep in the app, reached a customer-verifiable version in two days, against at least a month previously.

On how AI is allowed into core business in a sector where a wrong credit decision becomes a bad loan, the bank described an "Agent First" approach: when designing any task, first ask which parts can be given to an agent, then define where humans make judgments, where permission boundaries lie and who bears final responsibility. The bank said agents handle execution, exploration, analysis and simulation, while people set goals and standards, judge results, correct deviations and carry responsibility. It described Harness engineering to work around hallucinations, context limits and long-horizon task execution, with layered governance of evidence, model output, permissions, tool calls and results so a bug can be halted, taken over, recovered and traced; an evaluation system that verifies and traces capability when models, prompts or skills change; and a knowledge base compiled from scattered documents, emails and staff knowledge plus historical business rules and customer data. At the 2025 conference, MYbank proposed a "New 310" for the AI era: zero-latency interaction, 360-degree perception of users and one-to-one expert service. A year later, the bank said complex requests that took three days now take 10 minutes.

Editor's Summary. MYbank has opened its Bailing 2.0 finance agent to 42 million small and micro business owners and reported AI deployment across risk control, human review, marketing and research and development. The bank's examples show individual credit decisions being reshaped by customer-level analysis and human review, with one restaurant owner receiving a 400,000-yuan limit in 10 minutes. The results remain the bank's own account of internal testing and customer cases rather than independently verified figures.