There is a single mistake running underneath most of the expensive problems in retail right now, and it is a placement error. The biggest number in a category keeps getting filed under a function, a tool, or a calendar that has neither the mandate nor the instruments to move it. The number wears a costume that tells the org where to send it, and the org obeys the costume rather than the mechanism. So the problem gets owned by people who cannot touch its root cause, and the root cause sits somewhere else, unowned, compounding quietly.
Once you see this pattern, you see it everywhere. Shrink wears a security costume, so it goes to loss prevention. Returns wear a logistics costume, so they go to a 3PL contract. A pricing model wears an optimization costume, so it goes to whoever chases margin. A supply chain twin wears a visualization costume, so it gets graded on how it demos. A clean room wears a privacy costume, so it gets used as an activation engine. Assortment wears a seasonal costume, so it gets locked once a year against a customer who no longer exists. In every case the decision is being made. It is just being made by the wrong owner, on the wrong signal, at the wrong moment, and reported as something other than what it is.
The variance and the budget point in different directions
Start with the clearest example, because the data does the arguing for you. The 2026 Appriss Retail benchmark puts shrink at $90 billion in losses, 73% of which was preventable: employee theft at $26 billion, inventory errors at $19 billion, operational errors at $12 billion, and organized retail crime at $9 billion. Organized retail crime, the thing driving the security spend and every headline, is the smallest of the four buckets. Inventory errors and operational errors together reach $31 billion, more than three times the ORC number, and neither is a crime problem. They are data and process problems. Even employee theft is mostly POS manipulation, a void pattern or refund anomaly sitting in a transaction log, not a person walking out with product in a jacket.
Yet the aggregate lands on loss prevention, a function staffed with investigators, built around cases, and measured on incidents. Hand that function a number and it does what its DNA tells it to do. It hunts theft. Cameras go up, cases get locked, guards get hired. The apparatus points at the perimeter while the variance sits in the ledger. The org that owns the number cannot move it, and the orgs that could move it (supply chain, merchandising, pricing, finance) do not own it. That gap is precisely where the money leaks.
Returns run the identical play. One in five online orders comes back, and the NRF’s 2025 Retail Returns Landscape puts the overall ecommerce rate at about 19 to 20%, with apparel at 20 to 40%, footwear at 17 to 30%, electronics at 8 to 15%, and beauty at 4 to 12%. The return gets filed under reverse logistics, with a cost center and a VP graded on moving a box backward cheaply. But the return did not start in the warehouse. Sizing, fit, and color are the primary reasons for 45% of all returns, 16% are due to damage, and 14% are due to inaccurate item descriptions. Fit, color, and copy are merchandising and pricing decisions, made weeks before the box ever ships. The logistics team is being asked to win a race that started at the price tag, and the return reason, which is the actual signal, dies on the dock instead of reaching the people who set price, fit, and photography.
The same structural error hides inside the tools. A clean room got sold as the privacy-safe fix for signal loss, and 66% of organizations now use one in some capacity. The average company spends around $879,000 on it, and then 39% of organizations struggle to drive actionable insights from what they bought. Match rates run 35% to 65% on a good day. That is defensible as a measurement instrument across parties whose data you have no legal right to see. It stops earning its cost the moment it is used to activate real-time decisions on data you already own, because the tool was built for a batch world and returns capped, aggregated answers on a delay. The job (measurement) and the tool (a clean room) got welded to a second job (activation) they were never designed to serve.
Why the wrong owner keeps getting the number
Three forces produce this pattern, and they reinforce each other.
The first is functional DNA. Every function metabolizes a problem into the shape it already knows. Loss prevention converts shrink into cases. Logistics converts returns into freight. A pricing model converts a customer relationship into a single transaction to maximize. That last one is worth dwelling on, because the mechanism is so clean. A price optimizer has no memory of the customer’s last order and no line item for the goodwill it spends to capture an extra 40 cents today. So it captures the 40 cents and books it as a win, and the tax it just levied on the relationship lands in a different quarter under a different metric with no obvious cause. The function did exactly what it was built to do. That is the problem.
The second force is that the calendar and the demo reward the wrong number. Legacy assortment planning starts with last year, copies the prior range, adjusts for known wins and losses, and applies a growth number. It is fast and familiar and blind to anything that changed. It persists because the planning calendar rewards a stable number you can lock in February over a range that keeps moving and has to be defended. Supply chain twins persist in the same way. Programs chase visual fidelity and node count because that demos well to executives, when the number that actually matters is wall-clock time from question to ranked action, and nobody puts that on the slide. A twin that produces a defensible answer in three weeks is a research paper. A twin that puts a good-enough ranked recommendation in a planner’s hands in twenty minutes, before the next replenishment cycle, is an operating asset. The org optimizes what it can see in the room, not what bends the P&L.
The third force is that raw signal feels like progress even when it is placed wrong. Wire a two-week TikTok spike straight into the buy and you have not modernized merchandising, you have built a slot machine that whipsaws replenishment and torches open-to-buy chasing a trend that peaked before the PO cleared. The failure there is placement, whatever the volume of signal. History is good at structure because structure changes slowly: category role, store clusters, size curves, margin guardrails, the mission each location serves. Live signal is good at deltas inside that fixed structure: which trend expresses deeper in which cluster, which store is moving early and deserves an in-season chase. Feed each type of data into the wrong slot and the result looks sophisticated right up until it starts destroying margin.
The cost that never appears as a line item
The reason this survives audit after audit is that the damage almost never shows up where you would look for it. A stale assortment does not fail loudly on buy day. Nobody gets fired in February. It bleeds across the season through markdowns, dead SKUs, and demand never captured, and in North America understocking alone runs into double digits as a share of inventory value. The number is real and large, and almost no planning team writes it on the whiteboard before locking the buy.
Call it what it is on every steering committee: the Cost of Doing Nothing. It is the expedited freight you ate because the twin could describe the delay but could not test a cheaper response in time. It is the markdown on the wrong 40,000 units. It is the shrink variance you attribute to demand volatility while the average US retailer runs at just 65% inventory accuracy, cascading into phantom stockouts, safety stock carrying costs of 20 to 30% annually, and forecast error you keep blaming on the market. It is the recovery gap on returns, where processing costs $5 to $15 per item but the all-in cost reaches $17 to $29 once freight, depreciation, and labor are counted, and you finance that same merchandising defect thousands of times a season while reporting it as operational efficiency because the box moved cheaply.
Trust behaves the same way. When Instacart quietly ended its AI item pricing tests in December 2025 after several shoppers were shown different prices for an identical product at a single moment, the trigger was a mutiny. Personalized pricing reduces perceived price fairness even for the shopper who receives the lower price, because it violates social norms. The optimizer wins the transaction and expenses the relationship, and the expense lands next year as a repeat rate that drifts down while margin looks fine. Amazon relearned this in 2000 when customers found they were quoted different prices by browser history. Twenty-five years later the lesson is being relearned with better tooling and worse memory. In every one of these cases the CODN is enormous, and it hides because it is booked as something else: as demand volatility, as the cost of ecommerce, as a license fee for a tool the planners already went back around to a spreadsheet and a phone call.
The operating model: route the signal to where the variance lives
The fix is a discipline of placement, and it has four moving parts.
Separate the job from the tool. A clean room is a measurement instrument across parties, defensible only at roughly $2 million and above in spend with genuine multi-publisher complexity and clean inputs. Real-time decisions on data you own belong on a live decisioning layer your agents can actually reach. Keep the clean room for the ledger and give the agents the live wire, because agentic systems decide in milliseconds and cannot run on a batch answer that landed yesterday and covered half your file.
Move the number to the function that owns the mechanism. Shrink belongs with the data and process orgs that can bend inventory accuracy, markdown timing, and POS exceptions, not with the security org counting cameras. Returns belong as merchandising telemetry, with an agentic loop that reads return-reason text at the unit level and hands merchandising a ranked list: this SKU is returning for too small, fix the size curve; this one for not as pictured, reshoot; this one is a genuine price-value gap, discontinue. Stop chasing the portfolio average and diagnose at the SKU.
Run it as a loop with a governor and a human seam. Structure comes from history, deltas come from live signal, and a governor quantifies the financial impact of acting or not acting so decisions align with margin and service goals. The mature deployment is hybrid: the system proposes the optimized range, and merchants approve or override. Signal proposes, humans dispose, guardrails hold the line. Before any of that, put in the foundational process controls, cycle counting, separation of duties at POS, receiving dock discipline, then layer AI on top of clean process.
Optimize for the number that bends the P&L. For a twin, that is wall-clock time from question to ranked action. Pick three disruptions that actually hurt you, a supplier going dark, a regional demand spike, a lane closure, demand ranked actions with margin deltas inside one shift, and kill everything that does not serve that loop. For pricing, make trust the constraint the optimizer runs inside rather than the thing it is allowed to spend. Pick market-responsive pricing with one public price at a moment and no personal data in the pricing function, make the driver legible to the shopper, and track repeat rate by the cohort exposed to the model as a first-class metric.
The retailers who win the next several cycles will be the ones who fixed placement rather than buying more data, a prettier map, or a flashier live feed: who moved each number to the owner who can move it, routed each signal to the point in the decision where it belongs, and put the Cost of Doing Nothing on the board before locking anything. The costume is coming off every one of these problems whether you take it off or not. Better to be the operator who names the mechanism first, while the variance is still yours to engineer.