Queue economics for a fraud and compliance review team

The problem

Every fraud team decides thousands of times a day which alerts get a human reviewer, usually by ranking on a model score.

This project treats a generated 24,000-alert queue as a constrained allocation problem instead. Each alert can be approved, declined or reviewed, and shadow prices on analyst-hours and on the decline-rate budget are solved so both constraints are respected.

Headline numbers

• Against a score-ranked baseline with the same decline budget and six analysts, total cost drops 54%.
• Alert recall falls from 85.1% to 80.0% while dollar recall rises from 76.8% to 93.0%.
• A score with AUC 0.866 overstates risk 2.7x in the 0.0 to 0.1 band, which holds 49% of the queue.
• The fifth analyst returns $2,882 a month and the sixth returns $109.

What it means

Alert recall weights a $12 fraud the same as a $9,000 one, so letting it fall can be the right call. Calibrated probabilities matter for any dollar decision, but they only help when the capacity and decline constraints are priced as well. Sanctions rules stay pinned to human review rather than priced.