Audiences AIThe return-risk score never leaves the warehouse
Walk into the fulfilment building of a mid-sized D2C brand and go past reception, past the packing benches, to a desk near the dispatch bay. This is where orders get looked at before they leave. The screen shows pin codes with poor delivery records and first-time cash-on-delivery buyers, plus a column the storefront never sees: a return-risk score on every order. When the score is bad, the desk acts. Perhaps a confirmation call before the parcel is packed. Perhaps a hold until the customer prepays. The desk exists because a returned order is among the most expensive events in the building, and the people closest to the trucks learned to see returns coming.
The desk earns its keep. In published festive-quarter industry data, return-to-origin on cash-on-delivery orders has run near 58 percent against under 15 percent on prepaid. Operators who gate dispatch on that score report double-digit reductions within a season, same catalogue, same customers, entirely through decisions made before dispatch. The figures are industry-reported ranges, and none of them is a claim about our own client work.
Three floors up
Now take the stairs to the marketing floor. A different screen shows the same order, and here it reads as pure good news. The CRM logged an order-placed event and fired the brand's best material at it: the thank-you flow, the cross-sell offer, loyalty enrolment, a replenishment reminder queued for next month. Nobody on this floor is careless. By every instrument they watch, courting a fresh buyer at the moment of purchase is exactly the job.
The wiring is identical in markets where cash on delivery barely exists. An American apparel shopper who orders three sizes to keep one looks, at checkout, like a high-value customer. Serial returners sit near the top of the file right up until the credits post. Wardrobing, the outfit worn once with the tags tucked in, reads as loyalty until the label comes back. The trigger is universal: marketing systems treat order placed as the finish line, and a meaningful share of placed orders never turns into kept revenue.
The number that never rides the elevator
Here is what strikes us in every version of this building we walk. The risk score exists. It is accurate enough that the warehouse bets real freight money on it daily. And the CRM has never once asked for it. Logistics commissioned the score to protect shipping cost, so it lives in the order-management system, scoped to the department that paid for it. The marketing stack was wired years earlier to read commerce events, and order placed is the richest event it knows how to hear.
So a decision falls between the floors: how much should the company invest in this order before the order proves real? Ops holds the information and no mandate over marketing spend. Marketing holds the budget and no sight of the risk. The decision still gets made daily, by default, in favour of full investment in every order, including the ones the ground floor has already flagged. This is what it looks like when a company has data on every floor and no decision layer: two competent departments and an unowned decision sitting in the gap between them.
When the score travels
Connecting the floors is mostly plumbing, an API call and a field mapping, and the plumbing is the easy part. The operating changes are where the money is.
- Gate the post-purchase sequence on risk tier. High-risk orders get delivery-focused messages, confirmation and tracking, and the celebration waits for the delivery scan. For those orders, delivered is the real order placed.
- Move loyalty accrual and cross-sell incentives to delivered status. A discount redeemed against an order that comes back is marketing spend with a negative return, paid out to the least profitable behaviour in the file.
- Report every flow against delivered revenue rather than booked revenue, so sequences compete on sales that stood.
- Feed outcomes back. When a well-timed confirmation nudge changes the survival rate of a risky cohort, the score gets a new input and the courier bill gets smaller.
None of this requires new modelling work. The score in the warehouse already survived the hardest test available, which is a logistics manager spending freight money on its say-so. What it requires is a place where order context and customer treatment meet before the treatment fires, with someone accountable for the trade.
That meeting place is what we build. Interpret, our retail and D2C product, reads the context a stack already holds, return risk included, and decides which treatment an order has actually earned. The brand's most expensive attention goes to sales that will stand, and the risky ones get a quiet confirmation nudge before any courtship. If your ops team scores orders and your flows have never heard about it, the two floors of your building are overdue an introduction.
Interpret is our decision system for Retail & D2C. See how it reads order context →