Stripe's Smart Retries are genuinely good at the problem they were built to solve. Given a failed charge, the model picks a retry moment informed by patterns across an enormous number of transactions — time of day, day of week, signals about when similar charges have succeeded. For the large class of failures that are really about timing, that is exactly the right tool, and merchants who enable it recover payments they would previously have lost to a fixed schedule's bad guesses. Any honest account of Stripe dunning starts there: the native machinery is not the enemy, and replacing it wholesale would be a step backwards.
The model's blindness is equally specific, though. It sees timing patterns. It does not see that this subscriber's card was replaced last Tuesday and the new number has never been presented. It does not see that this subscriber opened the billing email four times and clicked nothing, which usually means the ask itself is unclear rather than the moment being wrong. It does not see that this account has failed three renewals running for three different reasons, a pattern that says the payment method itself needs replacing, not retrying. Timing optimisation applied to a credential problem produces beautifully scheduled failures.
That distinction is the organising idea in stripe dunning: keep the model doing what it is good at, and build the missing layer for the cases where the variable that matters is not when the charge fires but what has changed since it last fired. In practice that layer is small and unglamorous. Listen for card updates and fire immediately when new details land, rather than at the next modelled slot. Treat an updated payment method as an event, not a schedule entry. And when a failure's code says the credential is dead — expired, lost, stolen, closed — stop spending retries entirely and spend an email instead.
There is a measurement trap here that catches careful teams. Smart Retries raise the overall recovery rate, and the rise can hide the composition underneath it. Timing failures recover better and better while credential failures recover at the same flat rate they always did, and because the timing pile is usually larger, the blended number improves and everyone moves on. Splitting the recovery rate by failure reason, even crudely, is the only way to see whether the remaining losses are a scheduling problem or a casework problem. The two piles need different budgets, different messages, and different definitions of done.
The email side deserves the same scepticism as the retry side. Stripe's built-in notifications are functional, but they arrive as processor correspondence, and subscribers have been trained by years of phishing warnings to be wary of exactly that shape of message. A recovery email that comes from the business itself, names the product the subscriber is losing, and asks for one action tends to be acted on at rates the generic notice does not reach. The retry model cannot fix that either. It was never asked to.
None of this argues for turning the native machinery off. It argues for knowing precisely which job it does. Timing is one variable in a failed payment, and for a meaningful minority of failures it is not even the important one. The merchants who recover the most from Stripe are the ones who let the model schedule the timing cases and refuse to let it improvise answers for the rest — which is the division of labour practised at Swipe Again: the platform retries on its clock, and every other failure gets read as the individual case it is.
Top comments (0)