Keep the revenue your payment layer loses.

Three mechanisms lose money at the payment layer, and most brands treat them as unrelated problems with different owners. A false decline is a good customer refused at the door. A chargeback is revenue you already recognised, taken back weeks later, with a fee and a deadline you probably missed. An expired card is a subscriber who never chose to leave. Same layer, same month, three separate exits.

This page is one half of a pair. Payment recovery covers the charge that failed — retry timing, decline-code routing and the dunning sequence after it. This one covers what that page does not: the screening decision before the charge, the dispute that arrives after it, and the card credentials that go stale in between. Halves of one problem, usually built together — but different work, so different pages.

For a subscription brand each of these compounds, because the charge repeats. A one-off store loses one order to a false decline; you lose that order and every one after it, the whole remaining lifetime of a customer whose card was fine. A chargeback is not a one-box problem either: it is the box, the fee, the ratio that fee counts against, and the subscription that ends alongside it.

The fraud filter is refusing customers you want

A screening decision is a classification with two ways to be wrong, and only one of them is ever counted. A fraudulent order that gets through lands in a report with a fee, a reason code and somebody’s name attached. A good customer refused at the door leaves no record at all — no ticket, no dispute, no line item, usually not even a complaint, because they assume the store is broken and buy elsewhere. One error is measured and the other is invisible, so the threshold only ever moves one way. Worse, the two errors are not even priced alike here: a fraudulent order costs you one box and a fee, while a refused subscriber costs you every renewal they would have paid for. The expensive mistake is the one nothing counts. That asymmetry, not the tool, is why the filter ends up refusing people you want.

So rules accumulate — block this country, block a mismatched AVS, block the third order from one address in an hour — each defensible when written, none ever removed. Meanwhile a renewal looks odd in ways a first order does not: same card, same amount, same address, month after month, and then a gift bag shipped to a different state in December. The customer never sees a fraud message. They see a decline, and a decline reads as a broken store.

The information that would settle it is already on the account and no rule is reading it: eleven delivered boxes, two years of tenure, the same card, a support history with no disputes in it. A first-time stranger paying with a new card has none of that. Those are two different populations and they need two different decisions — which is a modelling problem, and it is the one thing a rule list written a country at a time cannot express.

The chargeback nobody contested

Disputes arrive with a clock running and land in an inbox with no name on it. The evidence that wins one exists — the subscription agreement they accepted, the delivery confirmation, the login and IP history, the eleven boxes they took without complaint — but it sits in five systems and takes an hour to assemble, so the case is let go by default. That is an administrative loss, not a commercial one. And plenty of these are not fraud: “I forgot I was subscribed” arrives as a dispute rather than a cancellation, because the dispute form was easier to find than your cancel flow.

A dispute is a classification with a deadline: what kind of case is this, what evidence would answer it, does that evidence exist for this order. All three questions have answers a system can produce in seconds, out of the subscription record, the carrier scan and the message log. That is the part worth automating — the gathering, the typing and the clock — because the hour it takes by hand is the entire reason winnable cases go unanswered. What the issuer reads is still written and sent by a person.

The card that was always going to expire

Every stored card has an expiry date, and on a repeating charge that date is a scheduled failure with a known cause. Cards also get reissued after a breach or renumbered when a bank switches processor. Account and card updater services exist at Stripe and the major processors precisely for this, and are usually either switched off or switched on and never reconciled — so the refreshed credential sits at the processor while Recharge, Skio, Smartrr or Stay AI keeps billing the token it already had.

Where the decision stops being automatic

The system scores an order, types a dispute, assembles its evidence, watches the deadline and reconciles a refreshed card. It does not refund, credit, cancel or re-price anything, and it does not file a case on its own. Two calls in particular stay with a person: blocking a returning subscriber whose history says they are real, and answering a dispute where the honest read is that the customer has a point. Both are cheap for a human to make and expensive for a system to get wrong — a false decline is a customer lost silently, and an aggressive representment against a genuine complaint buys a fee and loses the relationship. Where a wrong answer costs more than a human minute, a human spends the minute.