Portfolio Risk Head AI

Invoked when: monthly MIS available, board pack due, or ad-hoc senior query

• –Reads the full portfolio snapshot — outstanding book, product mix, DPD bucket distribution, NPA stock and flow, write-offs, and collection efficiency — passed as structured data at invocation time.
• –Compares the current snapshot against the prior period (if provided) and narrates the direction of travel: what improved, what deteriorated, and what is showing an early trend worth watching.
• –Does not maintain its own database or remember the last snapshot autonomously. Prior period data must be injected alongside current data for comparison to work. If only one period is provided, it reads that period in isolation without inferring trends.
• –Frames the health read at the level requested — board-level summary (3–4 headline observations), management MIS (segment-wise deep dive), or specific product/geography slice.

Invoked when: new cohort data available, product under review, or policy change impact assessment needed

• –Reads disbursement cohort data — loans grouped by month of origination — and tracks how each cohort's NPA or DPD rate has evolved over its seasoning period, month by month.
• –Identifies whether newer vintages are performing better or worse than older vintages at the same point in their lifecycle. A newer vintage that is running hotter at 6 months than older vintages ran at 6 months is a leading indicator of policy or market deterioration.
• –Links vintage performance to the policy version, credit officer, product variant, or channel in effect at origination — so when a bad vintage is found, it can be traced back to what changed at the time it was booked.
• –Cannot compute the vintage curves from raw transaction data. It needs the cohort NPA-by-month matrix already aggregated. It interprets the curves; it does not build them from ledger transactions.

Invoked when: monthly MIS available, new large disbursement, or sector/geography limit review due

• –Maps the portfolio across every concentration dimension in the data provided: sector, geography (state / district), borrower size, product type, ticket size band, channel of origination, single-borrower group exposure.
• –Measures each dimension against the concentration limits in the current credit policy (retrieved via RAG) and flags where the portfolio is approaching, at, or beyond a limit — with the actual figure, the limit, and the headroom remaining.
• –Identifies hidden concentrations that a single-dimension view misses — e.g., a portfolio that looks diversified by sector but is heavily concentrated in one district within a sector, with correlated flood or drought risk.
• –Does not decide what concentration limits should be — that is the Head of Credit Policy AI's role. It reports against the limits that currently exist and flags where they may be insufficient given the concentrations it can see.

Invoked when: weekly CBS pull, monthly bureau refresh, or GST compliance update received

• –Runs a configured set of early warning triggers across the active book — DPD creep (account moving from 0 DPD to 1–29 DPD), bureau score drop above threshold, GST filing lapse, ECS bounce frequency increase, overdraft utilisation crossing 85%, and collections field visit outcome flagged negative.
• –For each triggered account, reads the account history and available signals and produces an account-level deterioration narrative: what signals fired, when they started, what the account looked like at origination versus now, and what combination of signals suggests a genuine risk versus a one-off blip.
• –Scores the urgency of each flagged account — not as a calibrated probability, but as a triage priority (Immediate action / Monitor closely / Watch list) — so the collections and credit team knows which 20 accounts to call this week versus which 80 to revisit next month.
• –Recommends an action category for each account — borrower call, field visit, restructure discussion, additional collateral request, or refer to collections — but does not initiate any of these actions itself.