Propensity Scoring Agent AI

Invoked when: daily data pipeline completes before Collections Head AI generates the day's work-list

• –Reads the available signals for each delinquent account and synthesises them into a propensity score: CBS payment patterns (recent partial payments are a positive signal; a clean miss with no prior delinquency is different from a chronic pattern); CRM contact response history (accounts that have responded to WhatsApp before are more likely to respond again; accounts with repeated non-contact are harder to reach); bank balance signals via AA where consent is active (a positive balance on the day of scoring is a direct capacity signal); and bureau refresh data (new tradelines at other lenders may signal financial stress or competing obligations).
• –Weights signals differently by DPD band: for DPD 0–30, bank balance and recent payment behaviour are the dominant signals. For DPD 31–90, contact response history and PTP compliance rate become more predictive. For DPD 90+, the propensity to settle rather than to pay in full requires a different signal set — the agent applies the appropriate weight configuration per bucket.
• –Produces a plain-language propensity narrative for each scored account — 2–3 sentences explaining why the account scored where it did. The narrative is the signal behind the rank, and it is the narrative the collections caller needs to have a productive conversation. A rank without a narrative is a number without context.

Invoked when: propensity scores are complete and the Collections Head AI requires a ranked list for work-list generation

• –Ranks accounts within each DPD bucket by propensity score — ensuring the highest-propensity accounts are contacted first within each bucket. Also applies a secondary ranking criterion: exposure size. Two accounts with similar propensity scores are ranked by outstanding exposure — the higher-exposure account is prioritised because the recovery value of success is greater.
• –Identifies the accounts that are approaching bucket boundary dates — DPD 27–29 accounts that will move to the mid-bucket, DPD 87–89 accounts that will classify as NPA in the next 1–3 days. These near-boundary accounts receive a priority flag regardless of their propensity score, because the cost of missing a recovery opportunity in the final days before bucket transition is higher than for accounts with more time remaining.
• –Flags accounts where the propensity score has changed materially since yesterday's run — a large upward movement (bank balance improved, payment received at another lender) is an opportunity signal worth acting on immediately; a large downward movement (bank account drained, new delinquency on bureau) warrants a same-day review for escalation.

Invoked as part of the prioritisation run — channel recommendation generated alongside each account's rank

• –Recommends the optimal first-contact channel for each account based on its prior response history: an account that has responded to WhatsApp in the past 30 days receives a WhatsApp recommendation; an account that has never responded to WhatsApp but has responded to IVR calls is recommended for IVR first. Where no prior response history exists, defaults to WhatsApp as the lowest-friction first contact.
• –Flags channel constraints: accounts where the borrower has unsubscribed from WhatsApp (do not contact via WhatsApp), accounts where the phone number is inactive or disconnected (flag for address update before contact), and accounts where TRAI DND status prevents automated messaging (route to human caller only).
• –Recommends human caller contact for accounts flagged as requiring empathy or judgment — prior hardship indicators, accounts the Early Warning Agent AI has flagged as stress-signalled, or accounts where the propensity score narrative indicates financial complexity that an automated channel cannot handle.