I have spent this year in enough retail and CPG steering committees to notice that the same conversation keeps happening under three different names. In one room it is a forecasting RFP. In the next it is a trade promotion pilot. In the third it is a retail media measurement stack. The vendors are different, the slides are different, and the sponsors sit in different functions. The underlying mistake is identical.
Every one of these programs is buying intelligence at the top of the stack while the value is trapped at the bottom. The organizations funding a smarter forecasting model, a flashier personalization engine, or a prettier measurement dashboard have quietly convinced themselves that the constraint is the quality of the thinking. From the forward-deployed seat, the constraint is almost always something duller: the signal feeding the thinking, the baseline the thinking is measured against, and the absence of any wire that carries the output into a decision. Fix those three things and a mid-tier model outperforms a genius one. Skip them and the genius model produces confident, expensive, well-formatted mistakes.
This essay is about that single failure and the operating model that ends it.
The pattern: three programs, one broken layer
Start with demand planning, because it is the cleanest example. Leaders walk in shopping for transformer architectures and probabilistic ensembles, and almost none of them have a model problem. They have a signal problem. The point-of-sale feed lands two or three days late, so the engine forecasts against a past it cannot fully see. The promotional calendar lives in a merchandising spreadsheet the model never ingests, so a forty-percent-off event registers as unexplained noise. As I argued in Demand forecasting has a signal problem, not a model problem, you cannot model your way out of a bad input layer. Feed an advanced model dirty, lagging, partial signal and it hands you back a more articulate wrong number.
Now move to trade promotion. CPG brands spend fifteen to twenty-five percent of revenue on trade, and a Tellius analysis citing McKinsey puts the share of US promotions that lose money at seventy-two percent. That is the largest, leakiest, most data-rich workload in the enterprise, and it sits mostly untouched by AI. The reason is not intelligence. The blocker is that the organization cannot define a defensible baseline. Without a clean read on incrementality, you cannot separate a genuinely profitable promotion from a sugar high that pulled forward demand you would have captured anyway. The math is not the hard part. The baseline is.
Then look at retail media. Skai and Stratably reported that only fifteen percent of marketers rate themselves very or extremely effective at measuring retail media performance, and zero percent named access to data as the barrier. The pipes are connected. The platforms are sharing. Seventy-five percent still name incrementality as their hardest measurement problem, and only twenty percent are good at both measuring incrementality and applying it to a decision. As I laid out in The omnichannel measurement stack that actually closes the loop, the break is in the gap between knowing and acting.
Three programs, three functions, three vendor categories. In each one the top of the stack is fine and the bottom of the stack leaks. The demand model is smarter than its inputs. The promotion analysis is smarter than its baseline. The measurement dashboard is smarter than the decision it can trigger, which is to say it triggers none at all. Retail has industrialized the production of insight and forgotten to wire it to anything.
Why the wrong layer keeps winning the budget
If the constraint is so consistent, why does the money keep flowing to the wrong layer? Three forces explain almost all of it.
The first is that the wrong layer demos beautifully and the right layer does not. A personalization engine produces a tidy lift chart for the steering committee. A forecasting vendor can show you a benchmark accuracy number in a thirty-minute call. A measurement dashboard renders a SKU-level sale stitched back to an ad exposure and it looks like proof. Meanwhile the work that actually matters is data engineering: instrumenting clean capture, closing the latency gap, standardizing grain, building a defensible incrementality baseline, and constructing an action layer that moves spend inside the platforms. That work is unglamorous. It photographs poorly in a board review. It wins anyway, and it keeps losing the budget to things that look better in the room.
The second force is that the broken layer is politically owned by no one, or by three functions that disagree. Trade promotion is the sharpest case. It is messy, buried in a spreadsheet swamp, and split across sales, finance, and marketing, none of whom want to open it in a board review. So it gets left alone while the AI capital funds the safer, photogenic work next door. The same fragmentation shows up in measurement, where Amazon measures at SKU level, Meta by audience, and Google by keyword, and no single owner is accountable for reconciling all three to one human being.
The third force is the most dangerous, because it hides the cost. These failures are silent. There is no outage and no error log when a forecast runs on lagging signal. There is just a stockout that did not have to happen and a markdown that did not have to be taken. There is no alarm when seventy-two percent of promotions lose money, because the answer only arrives after the quarter closes, when it no longer matters. There is no red light when a retail media network buys sales you would have won anyway. The failure produces a number that is quietly off, and a quietly wrong number is the hardest thing in the enterprise to fund a fix for.
There is one tell that cuts through all of this, and leaders consistently misread it. In planning, it is the override rate. When planners systematically overrule the system, leadership blames adoption or change management. The planners have learned, correctly, that the model is running on signal they trust little. They are protecting the business from a polished output.. Treat the override rate as a signal-quality gauge rather than a resistance problem, and you find the broken layer immediately.
The tax you cannot see on a dashboard
The reason this matters at the P&L level is that the cost of leaving the broken layer alone stays in motion. It compounds.
Consider the scale. LatentView puts global retail inventory distortion, the combined drag of overstocking and stockouts, at roughly 1.73 trillion dollars a year. That cost lives in the inputs feeding the model and in the disconnect between a forecast and the decision it should drive. Every quarter spent evaluating algorithms while the signal layer leaks is a quarter that distortion keeps compounding.
Trade promotion is the most quantifiable version of this tax in the entire AI portfolio. If trade is twenty percent of a five-billion-dollar revenue line, that is a billion dollars in annual spend, and if seventy-two percent of it is unprofitable, a meaningful fraction of that billion is funding programs that destroy value. The Tellius work notes that organizations applying measurement-first AI typically see a ten to fifteen percent improvement in trade ROI simply by identifying and killing the underperformers. On a line worth a fifth of revenue, that is a margin event, and it recurs every planning cycle you delay it. The cost of leaving trade alone for four more quarters is a billion-dollar line operated on faith while the AI budget funds a chatbot.
Retail media carries the same compounding structure. Every cycle you measure without acting, you re-fund non-incremental spend you have already proven is non-incremental. You let identity gaps misattribute a store sale to a channel other than the one that earned it. You widen the distance between the brands building control systems and the brands admiring reports.. The fifteen-percent effectiveness number is not a measurement curiosity. It means the market has accepted measurement that cannot act as the ordinary price of doing business, and that acceptance is the most expensive line item nobody writes down.
The strategic point is that this tax is asymmetric. When a competitor fixes the broken layer first, they get a structurally lower cost of doing business, and they get it before you notice, because their advantage is invisible on your dashboard too.
The operating model: signal in, baseline, action out
The fix is the same in all three domains, which is what makes it an operating model rather than a project. It has three parts, and they run in order.
The first part is signal. Before anyone evaluates an algorithm, complete a data audit and treat the input layer as the primary engineering problem. In forecasting, that means wiring POS, promo, pricing, weather, web, and external demand signals into a single timely feed at the right grain. In measurement, it means identity resolution first, with deterministic matching on captured email, phone, and authenticated customer IDs as the spine and probabilistic as the backup. If a customer browsing on mobile and buying in-store reads as two strangers, every downstream number is fiction. Signal quality is the variable that actually determines whether anything above it delivers, so it goes first.
The second part is the baseline. The most common way AI programs fail is by automating a decision the organization could never defend in the first place. An agent pointed at a baseline you cannot defend just makes a bad decision faster. So before automation, build the measurement layer the automation will reason over. In trade, that means getting incrementality right so you can separate real lift from pulled-forward demand. In retail media, it means treating incrementality as distinct from attribution and running holdouts and controls as always-on standard practice rather than a once-a-quarter lift test. Attribution tells you which touch got credit. Incrementality tells you what would not have happened without the spend, and only the second question justifies a budget. Given that only twenty percent of brands are good at both measuring incrementality and applying it, this is where most of the competitive ground is still open.
The third part, and the one almost everyone skips, is the action layer. Identity and a defensible baseline produce a signal. The action layer turns that signal into a change in the world on a cadence the market actually moves at. In planning, the forecast output has to flow directly into a replenishment decision, because a forecast that does not trigger an action is a research project. In trade, an agentic layer can run continuous post-event analysis, compute true incrementality against the modeled baseline, flag the losers inside the cycle rather than three weeks later, and recommend reallocation before the next planning window locks. In retail media, re-allocation should be a rule the system executes, so that when incrementality says a network is buying sales you would have won anyway, the budget is already moving before the readout deck gets formatted.
The test for whether you have built this or merely bought another report is blunt. If your measurement layer cannot move a dollar of spend on its own, you have a thermometer. A thermometer reads the temperature. A thermostat changes it. The whole operating model exists to convert the first into the second.
What to build in the second half of 2026
The organizations that internalize this stop treating forecasting as a model-shopping exercise and start treating it as a signal-engineering discipline. They stop deploying AI where it photographs well and start deploying it where the money is leaking, which in most CPG P&Ls is the trade line. They stop buying prettier dashboards and start building systems that behave like control loops.
The practical move for a CDO, CMO, or VP of Transformation reading this in June 2026 is to change the first question on every roadmap review for the rest of the year. When a vendor wants to talk architecture, ask what signals the system consumes, at what latency, at what grain, what baseline it reasons against, and what decision the output triggers without a human in the middle. If the answer to that last question is a chart, you are looking at a slower mirror, and you should redirect the budget to the layer underneath it.
The market has spent a decade getting very good at producing insight and almost no time getting good at spending it. The retailers who close that gap in the next few cycles will win because their signal is clean, their baseline is defensible, and their systems already moved the money while everyone else was still formatting the deck. Build the loop. The intelligence was never the hard part.