Early Warning Agent AI

Invoked when: daily CBS pull detects a payment event (bounce, partial payment, late payment), weekly bureau refresh is received, or GST compliance data is updated

• –Monitors a configured set of stress signals across every active account: ECS bounce frequency and recency, DPD creep (account moving from 0 DPD toward 30 DPD), bureau score change since origination and since prior month, GST filing compliance and turnover trajectory, overdraft utilisation trend, average monthly bank balance trend via Account Aggregator, and new credit enquiries or new tradelines at other lenders indicating loan-seeking behaviour.
• –Reads each signal in the context of the account's history — a single ECS bounce from a borrower with a 36-month clean repayment history is a different signal from a single ECS bounce from a borrower who missed a payment 8 months ago and was subsequently regularised. The agent reads the account trajectory, not just the current event in isolation.
• –Weights signals differently based on their demonstrated predictive value by segment: for MSME borrowers, GST filing lapse and bank balance trend are stronger early-warning signals than bureau score changes (which lag by 30–60 days). For salaried borrowers, ECS bounce and bureau enquiry activity are more predictive. The signal weighting is configured by segment in the signal threshold corpus and applied at runtime.
• –Distinguishes between signals that are likely seasonal or structural for the borrower's business type — a construction contractor who always shows low bank balances in June due to monsoon project delays is not a stress signal; the same pattern from a retail borrower in a peak season is. Where the account profile includes sector and business type metadata, the agent applies that context. Where it does not, it flags the ambiguity rather than applying a uniform interpretation.

Invoked when: weekly batch run processes all active accounts to update the health score register

• –Produces a structured borrower health score for every active account — not a numerical PD estimate, but a five-dimensional health assessment across repayment behaviour (on-time payments, bounce history, DPD trajectory), financial health (bureau score trend, GST turnover trend, bank balance trend), leverage position (new credit taken since origination, OD utilisation), business health for MSME (GST filing regularity, sector macro indicators), and engagement quality (responsiveness to prior collections contact if any).
• –Compares each account's current health score against its origination health profile — what the account looked like when the loan was sanctioned — to identify accounts where the health trajectory since origination is significantly negative, even if the current absolute score does not yet breach a trigger threshold. A borrower who looked strong at origination but whose every metric is moving in the wrong direction is more concerning than a borrower with a weak absolute score who has been stable.
• –Produces a health score that is comparable across accounts and across time — so the collections team can run the portfolio against a consistent metric week over week and immediately see which accounts have deteriorated since last week's run, which have recovered, and which are stable but at a level that warrants monitoring.
• –Does not produce calibrated default probabilities. The health score is a qualitative composite of observable signals, not an actuarially validated model. It tells you which accounts are showing deterioration signals; it does not tell you the statistical probability that each account will default in the next 90 days. Those are different questions requiring different instruments.

Invoked when: weekly health scoring run is complete and the collections pre-alert list needs to be generated for the coming week's collections activity

• –Reads the health score register and the collections CRM history — which accounts have been contacted before, what the outcome was, and what the account's payment behaviour has been since — and produces the pre-alert list: accounts that are in the Watch or Amber tier and should be contacted by the collections team this week, before they enter the DPD bucket and become a formal collections case.
• –For each pre-alert account, produces a plain-language account health brief — 3 to 5 sentences explaining what the signals are, what they suggest about the borrower's current financial situation, and what the collections team member should aim to establish in the contact call. The brief is designed to be read in 30 seconds and give the caller enough context to have a meaningful conversation, not a data dump that requires interpretation before dialling.
• –Recommends the appropriate pre-alert action for each account: a proactive relationship call for Watch-tier accounts (the goal is to understand the borrower's situation and provide early support, not to pressure for payment), a structured repayment reminder call for Amber-tier accounts where payment is approaching, and an immediate manager escalation for Red-tier accounts where the signal combination suggests imminent default risk.
• –Does not communicate with borrowers directly, initiate automated payment reminders, or push notifications through any channel. The pre-alert is a list and brief for the human collections team to act on. Every borrower contact is initiated and conducted by a human collections officer using the briefing the agent has prepared.