The context
A credit insurer protects its policyholders against the risk that their customers do not pay. When a covered invoice goes unpaid, the insurer indemnifies the policyholder and then works to recover the debt itself. For a global insurer, that means a continuous flow of overdue commercial accounts, across countries, sectors and amounts.
Each of those accounts reaches a fork. It can be pursued through amicable collection, by letter, phone and negotiation, or it can be escalated to lawyers and the courts. Legal action recovers money that would otherwise be lost, but it is slow and expensive, and on the wrong case it costs more than it returns.
The challenge
Escalation decisions were made case by case by experienced collection teams. Their judgement was good, but it was applied one file at a time, under time pressure, and without a consistent way of weighing everything the company already knew: the debtor’s profile and history, the size and age of the debt, the jurisdiction, and how similar cases had ended in the past.
The result was a familiar pattern. Some cases went to legal action that would have settled amicably. Others that justified escalation waited too long. Across thousands of decisions a year, small inconsistencies added up to a large and avoidable legal bill.
The question we were asked to answer was simple to state and hard to solve: for each overdue account, which route gives the best expected recovery after cost?
What we did
We started with the insurer’s own history rather than with a model. Years of closed cases already contained the answer to the question: for each one, the route taken, the cost incurred and the amount recovered. We assembled that history into a single, clean view of every case and its outcome.
On that foundation we built a machine learning model that estimates, for a new overdue account, the likely outcome of each collection route and the cost of getting there. The model does not simply predict whether a debt will be repaid; it compares the routes and recommends the one with the best expected result net of legal cost.
Before the model touched a live decision, we tested it on past cases it had never seen. We compared what it would have recommended with what actually happened, case by case, so the expected saving was measured on real outcomes rather than assumed. Only then did it move into production, where it now informs collection teams at the point of decision. The final call stays with the team; the model makes sure every case is weighed in the same, evidence-based way.
The result
The model now scores every case before an escalation decision is made. Its estimated impact is more than €3 million a year in avoided legal costs, achieved by escalating fewer cases that would have settled anyway, while maintaining recovery on the cases that justify legal action.
Why it worked
- It started from a money question. The objective was expected recovery after cost, not prediction accuracy for its own sake.
- It was proven on history first. Back-testing on past cases turned a forecast into a measured result before any change to live operations.
- It supported people rather than replacing them. Collection teams kept the decision, with a consistent, data-driven recommendation in front of them.
The same approach applies wherever a business makes many similar, high-value decisions: which invoices to chase first, which customers to extend credit to, which claims to investigate. See our decision intelligence practice.