India’s technology lobby NASSCOM urged regulators to expand sectoral sandboxes for artificial intelligence deployments this week, as Reserve Bank of India supervisors monitor pilot credit-scoring models at non-bank lenders that use large language models to parse alternative data.
The push comes while Electronics and IT Minister Ashwini Vaishnaw prepares a consultation paper on AI safety, deepfake harms, and skilling, signaling a move beyond voluntary industry commitments. NASSCOM executives argued that controlled test environments—similar to RBI’s existing regulatory sandbox for fintech—would let banks and NBFCs experiment with generative models without exposing retail borrowers to unaudited black boxes.
What NBFC pilots entail
People familiar with two supervised pilots said lenders feed anonymised cash-flow and GST invoice metadata into LLM-assisted scoring engines that flag anomalies for human loan officers. Models do not auto-approve unsecured loans above prescribed ticket sizes; instead they produce narrative risk summaries that must be stored for audit.
Regulators worry that hallucinated features or biased training data could slip into decisions affecting micro, small, and medium enterprises already strained by higher borrowing costs after the October monetary policy review.
Industry asks
NASSCOM’s memo to MeitY and RBI proposed shared compute for red-team exercises, standardized incident reporting when models drift, and clarity on liability when a third-party SaaS model misclassifies creditworthiness. Startups said sandboxes would reduce time-to-market compared with seeking one-off no-objection letters.
Consumer groups counter that sandboxes must publish exit criteria and forbid live customer data unless opt-in consent is explicit—a tension Vaishnaw’s upcoming paper is expected to address.
Global backdrop
Indian NBFCs are not alone: global banks are racing to embed LLMs in know-your-customer and collections workflows, but India’s combination of Aadhaar-linked data rules and strict RBI outsourcing guidelines makes domestic pilots a test case for the Global South, policy analysts said.
Until formal rules land, lenders said they are provisioning extra capital for model risk and training staff to treat AI output as advisory, not authoritative—a interim posture NASSCOM hopes sandboxes can codify into best practice.
Data governance questions
Pilots require NBFCs to mask personally identifiable information before prompts reach cloud LLMs, and to store prompts and outputs for RBI inspection. Vendors pitching end-to-end automation must demonstrate that model versions are frozen during a credit decision window so auditors can reproduce outcomes.
Privacy lawyers noted that the Digital Personal Data Protection Act’s purpose limitation rules still apply when alternative data—such as utility payments or e-commerce receipts—feeds models.
Competition with traditional scores
Bureau scores remain the baseline for most unsecured loans; LLM overlays are being tested on thin-file borrowers such as gig workers. NASSCOM argued sandboxes should let lenders compare error rates between traditional logistic models and generative summaries without immediately exposing consumers to denials based solely on AI output.