An AI Mission for Returns Processing

Every returned item passes through a decision that is really a pricing decision, made in a few seconds by someone with no prices in front of them. The aggregate of those decisions, repeated across a season, is where most of the recoverable value in returns quietly disappears.

A returns processing station is one of the least ceremonious places in any large company. Someone opens a carton, pulls out a mid-priced household appliance, and turns it over twice. The outer box is crushed at one corner, the accessories look present, there is no visible damage to the unit itself, and the scanner shows a SKU, an order number, and a reason code that a customer selected from a dropdown three days ago: not as described. The station has four destinations — back to sellable stock, onto a cart bound for refurbishment, into a liquidation pallet, or into the recycling stream — and the person standing there has perhaps twenty seconds to choose one, because a few hundred more cartons are waiting behind this one and throughput is what the station is measured on. The item looks fine, so it goes back to stock, and three weeks later it comes back through the same door, returned by a second customer who found a component missing from the box. By then the company has paid to ship it twice, handle it twice, and refund it twice, and whatever the unit was worth on the first pass is mostly gone.

Nothing about that sequence involves anyone doing their job badly. The person at the station made the most reasonable inference available from the only evidence they had, which was what the item looked like from the outside. But the choice in front of them was not a sorting choice at all. It was a valuation choice — what is this specific unit worth in each of four different channels, right now, net of what it will cost to move it into each one — and not one of those four numbers was anywhere on the screen. The apparatus of the modern returns operation is built to make that decision fast, and almost none of it is built to make it correct, because correctness would require information the station has never had.

Disposition is a pricing decision made without a price

Consider what each of those four paths is actually worth, and how little of it holds still. Restocking recovers close to full value, but only if the unit is genuinely resellable and only if demand for that SKU still exists at the price it was carried at. A seasonal item put back on the shelf after its season has passed is not recovered value; it is a markdown scheduled for later, plus the carrying cost of holding it until then. Refurbishment recovers a good fraction of value when the defect is small and the parts are on hand, and destroys value when the refurbishment center is backed up for weeks or when the one component that fails on this model is the one on extended lead time. Liquidation converts the unit to cash immediately at a discount that depends on what else is in the lot and what the buyers are paying this month. Recycling recovers material value and, for some categories, is genuinely the highest-value option once the cost of touching the item again is counted honestly.

Four numbers, all of them real, all of them moving — with the season, with on-hand inventory for the SKU, with the current backlog at the refurbishment bench, with the composition of the pallet being built that week, with the freight cost of the lane the item would travel. Against that, the typical operation sets a disposition matrix: a grid keyed to category, price band, condition grade, and reason code, written by a thoughtful person some months ago, approved, printed, and laminated. The matrix is not stupid. It is a frozen approximation of a calculation that never stops changing, and it was probably right on average on the day it was written. What makes it expensive is that its errors are asymmetric and nearly invisible. When it sends a unit to liquidation that could have been refurbished and sold, no one ever learns what the alternative would have paid. When it restocks a unit that was going to fail on the next customer, the loss shows up later, in someone else's metric, attached to a different order number. The feedback that would correct the matrix is severed from the decision the matrix made, so the matrix stays laminated and the losses compound in silence. It does not help that the condition grade the whole grid hinges on is itself a twenty-second judgment by eye, collapsing enormous variation into three or four buckets that were designed to be assignable quickly rather than to be accurate.

The information existed the whole time; it never reached the dock

What is genuinely striking, once you go looking, is how much of what the station needed already sat inside the company. Current on-hand quantity and recent sell-through for the SKU live in the merchandising and inventory systems. Open backorders — the case where a returned unit is worth more than full retail because a customer is waiting for one — live in the order system. The refurbishment center's queue depth, parts availability, and historical cost per repair for this model live in a service platform or, often enough, a spreadsheet on someone's laptop. Realized recovery rates by channel live in finance, warranty status and serial-level service history somewhere else again, and the contract terms governing what the liquidator pays for which categories in a PDF in a procurement folder. The item-level evidence is right there at the station: photographs, weight against the expected weight of a complete unit, the accessory list, the customer's own free-text description of what was wrong, which is frequently more informative than the dropdown code they were forced to pick alongside it.

None of that reaches the person holding the item, and the reason is not secrecy or bad architecture in any single system. It is that assembling those fragments into one number, per unit, in twenty seconds, has never been something a human being could do, so the operation substituted the one input that is available instantly and free: how the item looks. That substitution is the whole problem in miniature: the company is not short of information about what its returned goods are worth, it is short of any mechanism that delivers the information to the moment of decision, and a piece of knowledge that arrives after the decision has been made is indistinguishable, economically, from a piece of knowledge nobody had.

This is also why the last two decades of returns software changed less than it promised. Returns management systems automated the mechanical steps beautifully — issuing the authorization, generating the label, tracking the inbound, triggering the refund — and left the judgment exactly where it was, on a person at a table with a scanner. Automating the paperwork around a decision does not improve the decision, and much of what is now marketed as a fix inherits the same limitation: Gartner has predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, pointing among other things at "agent washing," older rule engines relabeled without the underlying capability changing. A disposition rules engine with a modern interface is still the laminated matrix; it fires the answer it was told to fire, and it cannot tell you that the answer stopped being right in March.

Valuing the unit at the moment it is scanned

The alternative is not a better grid. It is treating each returned unit as a small valuation task that has to be completed in the seconds the dock allows, which means something has to gather the evidence, reach into the systems that hold the moving numbers, reason about the tradeoff, and produce a recommendation with its reasoning attached — all before the person at the station has finished turning the item over. That is a coordination job of the kind that has recently become tractable: read the item's identity and observable condition from the scan and the photographs, pull live inventory and demand for the SKU, check whether a backorder exists, query the refurbishment queue and the parts position for that model, price the liquidation channel against current terms, weigh the handling cost of each path, and route accordingly — while escalating the cases that genuinely deserve a person, the high-value units, the safety-relevant and regulated categories, and the units where two channels come out close enough that the judgment call is real. This is what an AI Mission actually is in an operational setting like this one: not a chatbot at the dock, but specialist agents working across the systems that hold the pieces, with human-in-the-loop gates on the decisions that carry consequence, which is the operating posture platforms like StudioX's are built around and the same shift a growing body of work on the move toward autonomous operations has been documenting across other back-of-house functions.

The second-order effect matters as much as the first. A system that makes the disposition decision also, unavoidably, creates a record of why it made it, and that record can be checked later against what the unit actually sold for. Suddenly the severed feedback loop closes: the operation can see which calls were wrong, in which categories, under which conditions, and the valuation model stops being a laminated artifact and starts being something that learns from its own results. The same record points upstream, because when returns of one model cluster around the same missing component, that is a packaging problem wearing a returns costume, far cheaper to fix at the source than to adjudicate one carton at a time.

So the useful way to see a returns operation is not as a logistics function that happens to make sorting decisions. It is the last pricing desk in the enterprise — thousands of small, irreversible valuations a day, on real assets, in a live market — and it is the only pricing desk in the company staffed by people who are handed no prices and given twenty seconds. Once you hold it that way, the metric on the wall looks wrong. Units processed per hour measures how fast the desk clears its book, and says nothing at all about whether it traded well. The number that ought to be on the wall is value recovered per decision, and the reason nobody has been able to put it there is that, until the information could arrive at the moment of the scan, it was never a number anyone could have computed.