A 27% default rate, and every approval argued case by case
I built the decisioning framework the business now approves customers against, and got three functions to approve against the same thresholds.
Zension Technologies (Zaam), Dubai · device subscriptions, UAE
Context
The programme sold devices on subscription, so every approval was a bet on twenty-four months of payments rather than a single transaction. Default was running at 27%.
Approvals at that point were case by case. Risk had an opinion, finance had a different one, and commercial had a target. Each individual decision was defensible. The portfolio they added up to was not.
What I did
- Built a weighted scoring model from the portfolio's own history: 128 subscriptions and 455 behavioural data points, covering credit score, plan value, card and payment behaviour, and bank and emirate risk, with knockout rules for the cases that should never pass regardless of score.
- Wired in the external checks that were already available but were being read inconsistently, through AECB and Credicide, so the same evidence produced the same answer every time.
- Backtested it before anyone had to trust it. The model accepted roughly a third of applications while excluding every customer already identified as a first-payment default.
- Found where the risk actually lived. Low-value plans defaulted at more than ten times the rate of high-value ones, and above a certain credit score there were no defaults at all in the sample. That reframed the argument from "approve less" to "approve differently".
- Built country and area risk sheets and demographic scoring alongside it, refreshed quarterly, because a model built once starts drifting the moment the market moves.
- Set explicit thresholds and an escalation path, then took them through finance, risk and commercial until all three signed the same set.
Subscriber default rate, with approval volume protected rather than cut.
Subscriber default rate
Same portfolio, same commercial targets. The framework changed which applications were accepted, not how many.
What I would tell you in an interview
The win was not the model. Any competent analyst can build a scoring sheet. The win was that finance, risk and commercial ended up approving against one agreed set of thresholds, so the argument moved from every individual customer to the policy itself. Cross-functional agreement beat the analytical fix, and it is the part that took longest.
Where this transfers
Any business making repeated approval decisions with money at stake has this problem: lending, subscriptions, insurance, credit terms, vendor onboarding. The mechanics differ, the failure mode does not. Decisions made one at a time, by whoever is available, produce a portfolio nobody designed.