Retail ai earns most: bu ne anlama geliyor
Retail AI investment concentrates on forecasting and replenishment. Both are worth doing. Neither is where the largest decision sits.
Assortment decides what is available to sell before any forecast applies. A perfect replenishment system keeps the wrong product in stock efficiently.
Why assortment resists analysis
Replenishment has a clean feedback loop: order, receive, sell, measure. Assortment does not. Delisting a product produces no counterfactual — you never observe what it would have sold, and you cannot easily separate the customers who switched to another line from the ones who stopped coming.
That absence of feedback is why assortment stays judgemental long after replenishment has been automated.
The three questions that are answerable
Transferable demand. When a product is out of stock, what do customers buy instead? Stockout periods are a natural experiment that most retailers already have in their data and rarely use. Products with high transferability are cheaper to delist than their sales rank suggests.
Incremental range. Does adding a line grow the category or split it? Measurable by comparing stores that carry it against matched stores that do not, at category level rather than product level.
Local fit. Which products perform differently by location once you control for store size and footfall? The residual is the part of assortment that should vary locally, and it is usually smaller than the number of variants stores actually request.
The delisting decision is the one with money in it
Ranking products by sales and cutting the tail is the common approach, and it is wrong for a specific reason: it ignores whether demand transfers.
A slow product whose buyers will switch to another line in the same store is cheap to remove. A slow product that is the only reason a segment visits is expensive to remove at any sales rank. Both look the same in a sales ranking.
Space is the constraint that makes it real
Assortment without a space constraint is a wish list. The decision that matters is what to carry given fixed shelf, and that turns every addition into a displacement.
Systems that recommend additions without naming what they displace are producing suggestions, not decisions. The displacement is the part the category manager is accountable for.
What good looks like operationally
A recommendation that states the product, the store cluster, the product it replaces, the expected category effect and the confidence in that estimate — reviewed by a category manager who can override with a recorded reason.
This is slower than an automated ranking and it is the version that survives the first quarter, because the person accountable for the category can see the argument and argue with it.