Status: Delivered. Described from our own delivery records. Client identity, brand and commercial terms are withheld. Figures, where given, cover the period stated and nothing beyond it.
Context
A healthcare debt-recovery operation with thousands of active accounts and a finite number of agent hours.
Which accounts get worked, in which order, is the single largest determinant of how much is recovered — and it was being decided using a bought-in credit signal that had nothing to do with this operation's own history.
The problem
- Prioritisation rested on an external rating not built for this population.
- Effort was being spent on accounts with a low likelihood of paying, which is expensive twice over: the cost of the call, and the higher-likelihood account not called.
- There was no feedback loop — outcomes were not being fed back into the next day's priorities.
- The measure of accuracy was not being tracked at all, so there was no baseline to improve on.
What we built
An ensemble model
Combining several gradient-boosting and tree-based approaches, scoring every account daily with a probability of payment.
An automated pipeline
Running the scoring on a daily schedule against current data.
A weekly retraining loop
So outcomes feed back into the model and prioritisation adapts as the population changes.
How we delivered it
- 1
Establish the baseline first
Measure how accurate the existing credit-rating approach actually was, because without that number an improvement cannot be claimed.
- 2
Feature engineering
On the operation's own history, with the data team.
- 3
Model development and comparison
Holding out a period for honest evaluation.
- 4
Pipeline
Daily scoring, monitored, with the score written where agents work.
- 5
Feedback loop
Weekly retraining on outcomes, with accuracy tracked over time.
How it was phased
Indicative phasing for work of this shape. Actual duration varies with data quality, access and decision speed.
What the client keeps
- Model code and training pipeline in their own repository
- The feature definitions
- Held-out evaluation results
- The baseline measurement
- Retraining configuration