Analytics-based end-to-end digital pre-approved personal loan for non-salaried customers at State Bank of India
Abstract
This study examines the problem of assessing credit risk for retail banking customers without verifiable income documentation. It focuses on the State Bank of India’s initiative to design a fully digital, pre-approved personal loan for its non-salaried customers. A two-stage unsupervised learning framework using two-step clustering is proposed. We apply artificial intelligence (AI) to transaction cash flows and demographic variables for a large population with diverse financial flows. In the first stage, experts identified and pruned coarse-grained clusters by reviewing delinquency concentrations. In the second stage, stable transaction behaviour patterns were identified through sub-clustering. This resulted in interpretable segments with heterogeneous financial propensities. The methodology was implemented on a massively parallel processing platform for analytics pipeline execution, combining ETL-driven filters, fraud prevention heuristics, and AI deployment on private cloud infrastructure. This approach allowed scalable inference without relying on hyperscale computing or specialised hardware and supported integration with end-to-end digital loan origination and disbursement. From March 2023 to October 2024, the system generated disbursements of ₹84 billion (approximately USD 1.024 billion) across 536,000 accounts, with all sanctions completed without requiring manual intervention. The contribution of this study is three-fold: it demonstrates the viability of unsupervised clustering as an alternative to rule-based scorecards, provides empirical evidence of the scalability of cashflow-based lending for retail customers, and introduces a governance-aligned framework for integrating human-in-the-loop evaluation with AI insights.