Head of Credit Policy AI

Invoked when: periodic policy review due, regulatory update issued, or portfolio stress observed

• –Reads the current policy document (retrieved via RAG) alongside the triggering input — new RBI circular, portfolio MIS, or a specific proposal — and identifies exactly which clauses need to change and why.
• –Drafts the revised policy section in the same format, language register, and clause structure as the existing document. Tracks what changed using a clean before/after diff so the committee sees only the delta.
• –Checks every proposed change against the full RBI Master Directions and NBFC regulatory corpus for conflicts, gaps, or compliance requirements that the change might create.
• –Does not decide what policy should be. It drafts what was asked, surfaces the tradeoffs, and flags regulatory risks — the policy committee makes the call.

Invoked when: MIS shows segment deterioration, new product segment proposed, or annual limit review

• –Reads segment-level portfolio data — approval rate, NPA rate, average ticket, vintage default curves — and identifies which segments are performing inside risk appetite and which are not.
• –For underperforming segments: analyses what the current limits allow, where defaults are concentrating (ticket size band, geography, tenure), and proposes specific limit tightening with the rationale.
• –For new segments: uses comparable segment performance and regulatory guidance to propose a starter limit framework — ticket size, LTV, FOIR, tenure, sector caps — calibrated conservatively until data accumulates.
• –Cannot run a statistical model. Reasons qualitatively over the data provided and clearly labels recommendations as judgment-based, not actuarially validated.

Invoked when: new product launch, bureau partner change, or approval rate / NPA rate outside target band

• –Reads bureau score distribution data for the target segment — applicant score spread, approval rates by score band, and NPA rates by score band — and identifies where the current cut-off sits relative to the risk-return curve.
• –Evaluates multi-bureau strategy: when to use CIBIL vs Experian vs CRIF as primary, how to handle thin-file applicants with no bureau footprint, and what alternate signals to apply (GST vintage, banking vintage, trade reference).
• –Proposes cut-off adjustments with explicit tradeoffs stated: a 10-point cut-off increase typically reduces approval rate by X% and reduces expected NPA by Y% — based on the score-band data provided.
• –Does not access live bureau data autonomously. Bureau analytics must be exported and passed to it. It reasons over the data; it does not run the query.

Invoked when: exception requested pre-sanction, or monthly exception log review due

• –For individual exceptions: reads the proposed exception, identifies which policy clause is being overridden, checks the exception log for how often this type of exception has been granted before, and drafts a structured exception memo with the business justification, risk assessment, and recommended conditions.
• –For monthly audits: reads the full exception log, categorises exceptions by type, frequency, authoriser, and loan outcome — and surfaces patterns that indicate policy gaps (a clause being systematically over-ridden suggests the clause itself needs review).
• –Tracks exception-to-outcome linkage: where loan performance data is available, it identifies whether exceptions that were granted are performing in-line with, better, or worse than standard approvals — feeding policy refinement.
• –Does not approve exceptions. It prepares the case and the governance trail. Human authorisers decide and their decisions are logged for the audit record.