The return was decided before you shipped the box

Picture the order from the warehouse floor, because that is where we first traced it on a returns engagement for an apparel retailer. Order 8412: three dresses, same style, three sizes. Packed with care, shipped express, delivered in two days. Eleven days later, two dresses came back, worn tags intact, and entered the returns stream: inspection, repackaging, restocking fee the retailer chose to eat, and a refund that unwound most of the order's margin. Standard bracketing behaviour, the merchandising team shrugged. Customers buy three sizes to try them on. The parcel is the fitting room now.

The shrug was the expensive part. Because when we scored a year of orders afterward, order 8412 was not a surprise that unfolded at day eleven. Every signal that predicted its outcome existed at checkout: the same-style-three-sizes pattern, the category's historic fit ambiguity, the customer's own return history, even the size-chart page visits before purchase. The model we later trained put the order's return probability above 80 percent while it was still in the cart.

Most returns are not decided in the customer's hallway mirror. They are decided at checkout, and the box ride is theater.

The retailer's dashboard said returns ran at 23 percent, which was budgeted for and therefore invisible. Scored order by order, the distribution was violently uneven: most orders carried single-digit risk, while a small slice, bracketed sizes, certain fabrics, certain fit-ambiguous styles, first-time customers in those styles, ran past 60. The aggregate number was an average of two different businesses: one that sells products, and one that operates an expensive lending library of dresses.

Per-order risk turns returns from a cost of doing business into a decision variable. Not "should we accept returns," but "knowing this order's risk before it ships, what is the margin-optimal way to fulfill it?"

(The merchandising team's fitting-room metaphor was right, incidentally. The mistake was not pricing the fitting room.)

At the soft end: intervene in the cart. For high-risk bracketing orders, surface the fit information that actually resolves the ambiguity, this style runs small, 84 percent of buyers your height kept the medium. On that engagement, a fit hint shown only to high-risk carts cut bracketing on the targeted styles measurably without touching conversion, which no sitewide banner had ever managed.

In the middle: fulfill differently. An 80-percent-return order does not earn express shipping and premium packaging; route it standard, hold it out of the tightest inventory, and let the promise-date algorithm treat it honestly. At the sharp end: price the option. Free returns for the low-risk majority, a modest return fee or store-credit-only on the highest tier, communicated plainly. The point is not punishment; it is that a uniform returns policy makes your best customers subsidize your fitting-room users.

A returns policy applied uniformly is a subsidy flowing from customers who keep things to customers who borrow them. .// The model behind this is unglamorous gradient-boosted classification on features you already log: basket composition, size multiplicity, category fit-ambiguity, customer history, session behavior. The craft is in calibration and in the economics layer on top, because an 80 percent score only supports a fulfillment decision if 80 means 80, and because every intervention has a conversion cost that must be netted against the return saved. We built that decision layer, score to policy to measured margin, and the measurement loop is the part that keeps it honest: every intervention runs against a holdout so the margin claim survives an audit.

Take last quarter's orders, score nothing, just split them by one feature, same style in multiple sizes, and compare return rates. That single split usually reveals whether you are running a lending library on the side, and how large its balance sheet is. The full model refines the answer; the split starts the conversation.

At TensorLabs we build return-risk scoring into the checkout and fulfillment path, calibrated on your own history, with the holdout measurement wired in from day one. Reply with your category and your blended return rate, and we will point at where the risk concentration almost certainly sits in your mix and what a scored pilot would look like on one style family. The pilot is designed to fit inside a season, not a roadmap.