Financial services · Machine learning · 2019
They asked whether to hire more collection agents
The collections department was out of bandwidth and the client was ready to hire more agents. The useful question turned out to be a different one: how much of the queue needed an agent at all?
What made it hard
- The collections department had run out of bandwidth.
- The client was preparing to invest in additional collection agents, and wanted to know whether technology could help instead.
- Every loan in the book was being worked the same way, in order, with no triage.
What we did
- Built a machine-learning model reading 1,000+ features per loan.
- Sorted every loan into three piles — self-correcting, potentially fraudulent, and collectable — so agents only ever saw the third.
- Routed potential fraud to investigations rather than to collections, where it had been consuming agent time to no purpose.
- Fed outcomes back into the model continuously so it kept re-tuning against what actually happened.
What came out
- 92% prediction accuracy.
- Agent workload halved, with no additional hiring.
- The amount collected rose by around 10% from the same department.
Built with
- Machine learning
- Python
- AWS
Client names withheld by agreement. References available on request.
About to add headcount to a queue?
Before you do, it is worth knowing how much of that queue needs a person at all. That is a short piece of work and the first conversation costs nothing.