--- title: "Three minutes, one second, zero humans: inside MYbank's AI microloan engine" date: 2026-10-17 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-mybank-310-ai-microloans language: en --- # Three minutes, one second, zero humans: inside MYbank's AI microloan engine > MYbank, an Ant Group affiliate, runs a fully automated "310" lending model that issues collateral-free loans to millions of Chinese small businesses in minutes—raising real questions about risk, fairness, and what actually scales. A noodle-stall owner in a provincial town opens her phone, taps a few times, and watches a working-capital loan land in her account before the kettle boils. No branch visit, no collateral, no loan officer on the other end. Behind that instant approval sits one of the world's largest live deployments of AI in consumer and small-business lending. ## What the "310" model actually does The headline is a three-number rule: - **3 minutes** to apply, on a phone. - **1 second** to approve. - **0 human beings** in the loop. MYbank (网商银行), an associate of Ant Group, pioneered this "310" model after it was founded in 2015 as one of China's first privately run online banks. It was also the first Chinese bank to run its entire core banking system on the cloud, with no physical branches. That architecture is the quiet prerequisite: there is no paperwork queue to wait in because there is no branch to walk into. ## The data engine behind the speed Speed alone is trivial if the credit decision is wrong. The bank leans on a stack of AI techniques—graph computing, multimodal recognition, and privacy-preserving computation—to score borrowers who often have no traditional credit file. - Risk models draw on more than 100,000 risk-profiling metrics and roughly 100 dedicated credit-risk models tuned for small and micro enterprises, according to a Singapore Management University case study of the bank. - Inputs include transaction and repayment records from e-commerce and payment platforms, smartphone payment histories, registered business profiles, local government records, and insurance data. - For farmers, a satellite remote-sensing layer reads crop growth and field conditions to set a credit line—no site visit required. Like much of China's fintech stack, the scoring layer now sits beside large model (大模型) systems that handle borrower questions in plain language, a direction the bank has publicly described as its next step toward "AI banking." ## How big, and who it reaches The scale is the part worth pausing on. - MYbank reported serving **over 45 million** small-and-micro enterprise clients at the end of 2021, up nearly 30% year on year; industry coverage put the figure above **50 million** by the end of 2022. - More than **80%** of its new 2021 clients were first-time borrowers, and roughly **80%** of its small-business customers had previously been unable to secure a loan from any financial institution. - Loans are small-ticket and short: the average borrowing period is about **3 months**, and **70%** of the interest charged on a given loan was below RMB 100 (≈ US$15 / HK$117). A documented average loan size runs around RMB 34,000 (≈ US$4,760 / HK$37,100). - The per-loan operating cost is reported at about RMB 2.30 (≈ US$0.32 / HK$2.50), against roughly RMB 2,000 (≈ US$280 / HK$2,180) for a traditional micro-loan—a thousand-fold drop that is the real story here. - In 2021, more than **500** major brands such as China Mobile, Haier, and Mengniu used MYbank's supply-chain financing, lifting loan accessibility for suppliers in their chains to about **80%**. ## Why it matters beyond China The number that surprises outside observers is the loss rate. MYbank's non-performing-loan ratio held at about **1.53%** in 2021 and roughly **1.94%** in 2022—comparable to, and sometimes below, traditional corporate lending, despite zero human underwriting. That combination of speed, scale, and contained defaults is why the model gets studied abroad: a University of Zurich executive program and Swiss financial press have explicitly pointed to the 310 approach as something European banks could learn from. Chinese state-owned and private competitors, including China Construction Bank, Tencent, and Ping An, have since launched similar automated small-business lending. ## The model's transferable lesson The transferable lesson from the 310 engine is not the brand but the data architecture behind it. Traditional underwriting assumes a paper trail — a credit bureau file, collateral, a relationship manager. MYbank inverted that: it treats the borrower's digital footprint inside payment and e-commerce ecosystems as the credit file, and lets alternative-data scoring stand in for the missing documents. That reframing is what other markets find instructive, because thin-file borrowers exist everywhere, not only in China. The caution is symmetric: the same data that makes inclusion possible also makes exclusion automatic for anyone who falls outside the dominant platforms. Whether the model scales abroad depends less on the AI than on whether a country's financial ecosystem produces the behavioral signal the scoring needs — a reminder that the technology is the easy part and the data fabric is the constraint. ## Honest limitations - The default and cost figures above are **bank-reported**. At this automation scale there is little independent audit of how models behave across a full credit cycle, especially in a downturn. - **Algorithmic exclusion is real**. Credit rests on behavioral and ecosystem data, so businesses that live outside the dominant payment platforms can be scored as invisible. - **Correlated risk**. When many lenders share similar alternative-data assumptions, a shock can propagate quietly across the system. - **Regulation is moving**. Chinese authorities have capped comprehensive microlending costs (around 18% in normal performance) and tightened scrutiny of fintech credit; the model operates inside a shifting rulebook, not a stable one. ## What readers can do now - **If you run a small business in China**, your payment and e-commerce footprint may already qualify you for instant credit—but read the all-in APR against the regulatory cost ceiling that applies to microlending before borrowing. - **If you invest or analyze**, treat "AI lending scale" as a starting metric; demand NPL, write-off, and vintage data before crediting efficiency. - **If you build fintech**, the transferable asset is the data pipeline—alternative data plus alternative scoring—not the brand name on the app. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the 310 model's real achievement is not speed but turning thin-file borrowers into a statistically manageable risk pool. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-mybank-310-ai-microloans Free to quote with attribution and a link to the original.