Yingju Ma

I study a generative-AI tool providing licensed brokers with reply suggestions at Ant Insurance, China's largest insurance brokerage. A ten-week randomized broker-level pilot precedes a 20-day customer-level experiment involving 164,732 customers and 1,070 brokers. Linked rollout records cover about 1.3 million customers. AI-assisted insurance advisory deepens customer engagement and increases life-insurance purchases. Customers who had shown interest but not bought purchase more life insurance and show persistent premium gains. Revenue gains rise with broker ability, even though more capable brokers adopt AI suggestions less frequently. A model-based decomposition attributes most of the fitted ability gap to differences in implied returns per adoption. Earlier access leaves persistent advantages in broker performance and effort after full rollout. These findings show that generative AI can complement existing human capital and amplify performance differences, while adoption frequency can be a misleading proxy for economic value.

Leveraging data from a major fintech lending provider for SMEs in China and a natural experiment from a temporary and geographically discontinuous interest rate discount policy, we analyze the presence of information asymmetries, selection, and market power. Our findings reveal a decrease in the average probability of loan repayment following interest rate reductions, indicating that selection is advantageous. We explain the effect in a generalized Stiglitz-Weiss model accounting for a broad distribution of borrower risk and returns, supported by empirical evidence. Building on our estimated demand elasticities and selection effects, we assess the welfare implications of information-sharing and pro-competition policies. Our counterfactual analysis reveals that ensuring competition and mandating information sharing across financial providers results in the largest surplus gains. Meanwhile, optimal information disclosure significantly mitigates monopoly distortions by increasing credit supply to riskier borrowers.

Fintech credit through digital payment platforms reaches 500 million Chinese consumers, many of whom hold no traditional bank credit line. We use de-identified administrative data from Alipay covering 60 months of full payment activity, both online and in-person, across payment instruments and consumption categories, for a sample drawn from a near-billion-consumer population. This provides the first integrated view of consumer payment and credit behavior in China. Using the January 2020 Covid-19 outbreak as a quasi-exogenous liquidity shock and a matched difference-in-differences design on 434,398 consumers, we find that fintech credit users sustain 10.2% higher consumption over the following two years. The response is concentrated among consumers facing binding liquidity constraints: low income, less developed regions, and limited external bank credit. The effect is further amplified among less financially sophisticated consumers, suggesting that fintech credit lowers the financial expertise required to access and use credit effectively.

Breakthroughs in artificial intelligence are reshaping credit markets by enabling real-time risk management, not merely better screening at origination. Using proprietary, high-frequency data on a sample of 800,000 small-business borrowers drawn from a large platform lender, covering January 2018 to March 2021, we show that the AI-powered lending system actively adjusts credit access, credit lines, and interest rates as risk conditions change, leveraging real-time big data on borrowers to contain losses in the unsecured credit loans. Employing an instrumental-variables design, we identify a causal effect of platform dependence on default: greater dependence lowers default, consistent with stronger repayment incentives from continued platform participation. These findings reveal a distinctive feature of AI-powered platform lending: it is most effective when algorithmic risk management is integrated with platform access and activity.