Artificial Intelligence Adoption, Algorithmic Risk and Financial Performance in Indian Banking: A Literature Review (Zenodo)

Summary

• The review draws on 20 studies identified in a preliminary literature matrix.p.1
• Banks use AI for credit evaluation, fraud detection, customer service and risk management, but face risks such as algorithmic bias, inaccurate predictions, data quality problems and a lack of transparency.p.1
• A model may predict repayment risk more accurately overall while still generating unfair outcomes for certain categories of borrowers.p.3
• A study of large US commercial banks linked AI innovation to better asset quality, but also to higher immediate operating costs and weaker short-term profitability.p.4
• A comparative study of public- and private-sector Indian banks found no statistically significant difference in direct ROA and ROE outcomes attributable to AI adoption alone.p.5
• The reported findings do not establish that AI adoption consistently improves profitability.p.6

Key insights

1. AI adoption in banking covers credit assessment, fraud detection, customer service, document processing and risk management.p.2
2. Model performance should not be assessed solely on overall predictive accuracy but also with fairness metrics and error rates across different groups.p.3
3. Benefits and costs of AI adoption may materialize over different time horizons, so short-term and longer-term financial effects may differ.p.4
4. Findings from US banks cannot be assumed to apply directly to Indian banks.p.4
5. Disclosure-based indicators of AI initiatives should be interpreted with caution because more reported initiatives may reflect better reporting rather than more effective AI use.p.5
6. Bank ownership may not be enough to explain the financial effects of AI.p.5
7. AI adoption, algorithmic risk and financial performance should be treated as related but separate dimensions.p.5
8. A relationship between AI adoption and financial performance does not by itself establish causation, so future studies need stronger research designs and controls.p.6