Major Chinese State-Owned Bank (anonymized in academic study)

The bank's traditional rule-based credit scoring model heavily relied on credit history to assess creditworthiness, systematically excluding underserved populations who lacked sufficient credit history due to historical discrimination and limited financial access. This created a 'chicken-and-egg' problem: the underserved needed capital to build credit history, but their lack of credit history prevented approval. The bank served over 50 million customers but struggled to safely extend personal loans to more than one million underserved applicants.

The bank developed an AI-enabled credit scoring model that incorporated 'weak signals' — non-traditional data not conventionally used in creditworthiness evaluation — alongside sophisticated machine learning algorithms to improve default risk prediction accuracy. The AI model was deployed alongside the existing rule-based model for one personal loan product over a seven-month period covering nine million applications, with both models used together to make final lending decisions.

The AI model simultaneously increased the approval rate for the underserved population and reduced the default rate for both underserved and regular populations, while also increasing loan utilization levels. Further analysis confirmed the improvement was driven by the AI model's use of weak signals and ML algorithms that enhanced individual-level prediction accuracy, reducing reliance on group characteristics that historically led to financial exclusion. Simulation analysis showed that even simplified versions of the AI model (fewer signals or simpler algorithms) could still produce positive financial inclusion outcomes.

• AI credit scoring using non-traditional 'weak signals' can improve financial inclusion and reduce default risk simultaneously, without a trade-off between access and safety.
• The positive impact is heterogeneous: subgroups missing weak signal data saw smaller but still positive improvements in approval rates, suggesting data availability is a key moderator.
• Simplified AI models retain meaningful financial inclusion benefits, making this approach generalizable to institutions with limited IT capabilities.