The Dollarized Confusion Matrix: How to Set AI Thresholds That Actually Make Sense

AI thresholds become defensible when every false positive and false negative carries a dollar cost. The optimal cutoff falls out of the cost asymmetry, then rolls into expected cost per item and an autonomy level: disable, human review, or spot-check. The useful move is simple: price the errors first, then let the confusion matrix tell you where the model belongs.