It works on four fronts: predicting approval likelihood before the request leaves, optimising message content and formatting per issuer, routing to the acquirer or rail most likely to approve, and choosing retry timing based on recovery probability.