Collections is largely reactive, with standard dunning after missed payments instead of early risk detection and outreach tailored to the borrower's situation.
Build direction: An early-warning collections agent that scores a mock portfolio, flags likely missed payments 7–14 days early, explains risk drivers, and recommends channel and message treatments.
Expected outcomes:
- Identify at least 60% of accounts that will miss a payment 7–14 days ahead of due date
- Increase cure rate against a control group
- Reduce cost-to-collect by replacing blanket outreach with targeted recommendations
Data inputs: Repayment history, balances, transaction data, behavioural signals, past collections outcomes, and treatment history.
Customer data is ingested from an upstream source — for this hackathon, a mock portfolio file stands in for real data. Each account is evaluated against a predefined risk matrix that produces a risk score and surfaces the key drivers behind it.
The risk score is passed to an LLM alongside a predefined treatment matrix. The LLM uses both to generate a personalised treatment recommendation per account — covering channel, message tone, and timing.
A human-in-the-loop (HITL) review step sits before any treatment is actioned. Reviewers can approve, edit, reject, or escalate each recommendation. A monitoring dashboard provides visibility across the portfolio — scores, treatment status, and outcomes.
Customer signals
-> Early-warning risk scoring
-> Risk driver explanation
-> Personalised treatment recommendation
-> Message or action draft
-> Human approval
-> Outcome tracking
| Step | Purpose |
|---|---|
| Customer signals | Collect repayment, balance, transaction, behaviour, and treatment-history signals. |
| Risk scoring | Identify customers likely to miss payment in the next 7-14 days. |
| Risk explanation | Show the main reasons behind each risk score in simple language. |
| Personalised treatment | Recommend channel, message tone, timing, and treatment path based on the customer case. |
| Human approval | Let a reviewer approve, edit, reject, or escalate the recommendation. |
| Outcome tracking | Measure whether the treatment improves cure rate and reduces unnecessary outreach. |