The RBI's FREE-AI Committee survey reveals that approximately 21% of surveyed entities are currently using or developing AI systems, with credit underwriting among the leading use cases. However, significant gaps in readiness persist - only 18% of entities using AI maintain audit logs, and only 35% validate models for bias and fairness. As reported by ETCIO, this regulatory direction is tightening with ongoing consultations on AI and ML governance in financial services, alongside provisions in the proposed Digital India Act on algorithmic accountability. The survey findings highlight that lenders who treat governance as a future compliance obligation rather than a current operating discipline are building on uncertain ground.
The Reserve Bank of India's draft Guidance on Regulatory Principles for Model Risk Management (2026) represents a comprehensive approach to governing AI models across regulated entities, from banks to NBFCs. According to reports from The Economic Times, the framework extends prudential thinking to machine learning and requires every entity to maintain a Board-approved framework and three lines of defence. High-risk models must receive Board Risk Management Committee approval, with each inventoried, validated before and after deployment, and monitored for drift. The framework treats a "model" expansively to include AI and ML, ensuring comprehensive coverage across all regulated entities.
India's credibility in AI governance rests on established digital infrastructure including Aadhaar, which provided verifiable digital identity to over a billion residents, and UPI, which processes billions of transactions monthly. According to reports from The Economic Times, these systems anchor Digital Public Infrastructure (DPI) that banks build upon rather than reinvent. The legal foundation is the Digital Personal Data Protection (DPDP) Act, 2023, which makes consent, lawful basis and purpose limitation the ground rules for personal data.
For AI specifically, entities must implement explainability thresholds, guard against hallucination, run red-teaming, and maintain human oversight with 'kill-switch' mechanisms. As reported by The Economic Times, the RBI states plainly that "an entity should not use any model that harms consumer." In lending applications, AI-assisted underwriting has been shown to compress decision timelines from several days to minutes, while alternative data models in pilot programmes have demonstrated 20-30% reductions in default rates compared with bureau-only approaches. According to ETCIO, AI can support contextual underwriting, reduce dependence on limited bureau signals, and make lending more responsive to business behaviour.
A strong AI lending framework should incorporate four practical layers: auditability, human oversight, explainability, and accountability. As reported by ETCIO, auditability requires maintaining versioned logs of model decisions, input variables, and output scores so that any credit outcome can be reconstructed and reviewed. Human oversight means high-impact decisions should have clear escalation and review points, while explainability ensures outputs are understandable to business, risk, and compliance teams using tools like SHAP values to surface specific variables driving credit recommendations. Accountability defines ownership across product, technology, credit, compliance, and leadership, with IBM's 2025 research finding that 97% of organizations lack proper AI access controls and 63% lack AI governance policies.
