Customer Case Study · Risk, Fraud & Collections
HMC helped the housing finance company's risk and collections teams move from reacting to overdue buckets to acting on early signals: an early-warning score for every live loan, link analysis that exposes fraud rings across applications, a collections plan built on propensity and team capacity, and GenAI outreach that is checked against fair-practice rules before it is sent.
Figures indicative; validated metrics available on request.
Risk was measured in days past due, fraud checks looked at one file at a time, and collections treated every customer in a bucket the same way.
Stress showed up as a missed EMI. Salary delays, bounces and new borrowing that came before it were not tracked together.
Each application looked clean on its own. The shared phones, devices, valuers and employers only showed across files.
Agents worked lists by bucket and amount. Customers who would self-cure got calls; customers who needed a plan got pressure.
Recovery messages and calls must follow fair-practice rules on timing, tone and privacy. Checks were manual and sample-based.
Explainable models the credit and collections teams can read, with GenAI used for outreach inside clear guardrails.
Select a step to see what happened.
Back-tested past slippages against the data available months earlier.
Deliverable: Signal inventory and baseline roll rates.
Every score shows its reasons.
Deliverable: Model and policy design signed off by risk and compliance.
One workspace for risk, fraud control and collections.
Deliverable: Working tools and daily dashboards.
People decide; the tools prepare.
Deliverable: Pilot results and playbook.
Models monitored and refreshed.
Deliverable: Model governance and roll-out plan.
Each module below runs live on a synthetic loan book of 480 home loans and 46 applications. Click, drag and play with them.
Optional. Every ✦ button already works on built-in templates. Add an OpenAI key to make them live: ask the portfolio in your own words, draft in eight languages, fix blocked drafts, read customer replies and role-play calls.
The key goes only from this browser to OpenAI and stays in memory unless you tick "remember". Only the synthetic data on this page is sent.
Each dot is a home loan. Play the months and watch signals build before accounts fall overdue. Switch signals off to see what each one adds, and move the threshold to match team capacity.
Synthetic portfolio. Scores use repayment, bureau, bank-statement and contact signals only; no protected attributes.
Applications are linked when they share a phone, device, bank account, address, unverified employer, valuer or DSA. Strong links form rings. Drag nodes, switch link types off, and click any application for its reasons.
Synthetic applications. Flags send files for verification by the fraud control unit; nothing is rejected automatically.
September's overdue and high-risk accounts, placed by ability and willingness to pay. Switch to the plan view to see each customer flow into a channel, and compare today's bucket-first allocation with propensity-led allocation.
Per cycle: a tele-caller handles 15 accounts, a field agent 6 visits and a counsellor 5 restructuring cases. Digital nudges have no limit.
Pick a customer. Draft a message in their language and check it against fair-practice rules, read what their reply really says, or rehearse the call with the AI playing the customer.
Pick a sample reply or type one, then press Analyse.
Synthetic customers. In production an LLM drafts within these templates and guardrails, and every draft passes the checker before it is sent. Placeholders in [brackets] are filled by the system.
Acting on early signals changed how accounts move between buckets, and the score concentrates future slippages at the top of the list.
Accounts per 10,000 moving between days-past-due buckets
Cumulative gains: share of slippages caught by working the riskiest accounts first
Figures indicative; validated metrics available on request.
The month before a missed EMI usually tells the story. Score it, explain it and act before the bucket changes.
A clean file can sit inside a dirty network. Check what applications share, and know which links are harmless.
Match the channel to the customer, help those in hardship, and check every message for fair practice before it goes out.