The tail you cut was holding the shelf

Every assortment review ends in a cut list. The category team ranks the range by velocity and margin, draws a line, and rationalizes the tail below it. The logic is clean and the arithmetic is right, and it still delists products the business needed on the shelf. A SKU that ranks in the tail nationally can be the second-best seller in one retailer's cluster, the only entry the brand has in a fast-growing channel, or the price-ladder step that makes the tier above it look worth the money. Cut the tail as a single ranked list and some of what you cut was load-bearing.

The default assortment question, and the one most SKU rationalization tools are built to answer, is "which SKUs should we drop?" A model ranks the range on national sales-out, flags the slow movers, and the category team takes a delist action. Distribution thins, a retailer's set loses its anchor, and the range looks tidier while revenue quietly leaks. The model was fine. The unit of analysis was wrong.

Why national velocity is the wrong unit of analysis

A brand's range doesn't perform as one list. It performs as a grid of SKU by store cluster by retailer, and the same SKU sits in a different place on every row. A regional analytics lead at a global CPG company put the stakes plainly:

That is the assortment decision stated correctly. The question is which SKUs belong where, against what the competition stocks, given the role each brand plays. A national velocity rank can't answer it, because it has already averaged away the regional and retailer variation where the answer lives. The SKU you'd cut on the national list is the one holding distribution in the region you're losing.

zoom_in

The decision runs on a quarterly clock; the shelf doesn't

Assortment is reviewed on a category-planning cadence, often quarterly, sometimes at the annual line review. The inputs arrive slower still: syndicated panel share lands weeks late, retailer point-of-sale comes in its own format on its own delay, and internal shipment data tells you what you sold in, not what sold through. By the time the category review reconciles all three and answers the question, the shelf reset it was meant to inform has already happened. The team is optimizing the last planogram, not the next one.

What category teams actually want is to ask the question in the week they're making the decision. Instead of waiting to rebuild the category review from scratch to learn a brand's dollar share by channel, they want to query it the morning the retailer asks. That gap between the pace of the decision and the pace of the answer is where assortment value leaks, and it doesn't show up in any single quarter's numbers.

Assortment constraints are the hard half

Here is where a ranked cut list breaks against reality. A category-operations lead at a global beauty company framed the real decision:

"Do we care if we're losing assortment at a given retailer" is a question about constraints. Every range sits inside rules a ranking ignores: planogram slots are finite and a delist doesn't always free the space you want; retailer joint-business-plan commitments oblige you to carry SKUs that don't earn their spot on velocity alone; portfolio roles mean the value anchor and the premium halo each hold their place for reasons that aren't this quarter's units; and supply constraints cap what you can add even when the data says add it. A recommendation engine that doesn't carry those rules will propose the highest-expected-value cut every time, and that cut breaks three commitments the category team can never execute.

The unlock is an assortment model that knows the constraints of the range it's editing, and only surfaces adds and delists that survive them.

What a category-intelligence stack actually has to carry

Most assortment tools carry one or two of these. A system that produces executable range moves carries all six.

Cluster-level velocity. The same SKU ranks differently in every retailer's store cluster. Generic tools rank on national sales-out and average the clusters away.

Distribution and void detection. A slow national SKU can be the only anchor holding a channel or region. Generic tools see the low rank, not the distribution it's holding.

Substitution and halo. Delist a SKU and demand moves somewhere, or the tier above it loses its step. Generic tools treat each SKU in isolation and lose the volume it displaces.

Portfolio role and price ladder. Brand roles and ladder steps govern which SKUs are load-bearing. Generic tools optimize units and cut the step that made the premium tier sell.

Retailer commitments as a hard input. JBP obligations and planogram slots decide what's executable. Generic tools propose delists the account team is contractually holding.

A short feedback loop. Assortment and share drift; last quarter's rank is already stale. Generic tools rank once at line review and never update in-cycle.

Why generic AI assortment optimization can't do this

An assortment recommendation is a constrained decision where the constraints are tacit. The cluster-level velocity lives in your retailer point-of-sale and panel data, not in any model's training set. The portfolio roles and price ladders live in a category playbook and a category manager's judgment. The substitution effects live in cross-SKU correlations only a model trained on your range will surface. The retailer commitments live in JBP documents and the account team's inbox.

A general-purpose model can write a memo about category management. It can't rank your range by store cluster, encode your planogram and JBP constraints, simulate where volume goes when you delist a SKU, and tell you which products to cut and which to protect. That takes a system built on your data, editing your range, inside your rules. It's one of the clearest AI agent use cases in CPG: an assortment agent scoped to the data the category function already owns, deciding a question the business already reviews.

The category agent, built on your range-review rules

A.Team doesn't drop in an assortment SaaS. We build the category-intelligence stack on your own retailer point-of-sale, panel, and shipment data, cluster it the way your accounts are actually merchandised, and encode your range-review rules, the portfolio roles, price ladders, planogram limits, and JBP commitments, as hard constraints, so the model only surfaces adds and delists the category team can execute. The assortment agent watches cluster-level velocity, share, and voids, and flags a range move with its reasoning, the distribution at risk, and a substitution estimate. A person approves it; the agent never edits the range on its own. We prove it on a 90-day lighthouse: one category, one retailer or region, a static dataset to a working model first, then live connectivity.

To be precise about proof, because the honest shape matters here. What's validated today is that category-operations teams at a global beauty company and at a global CPG company are running exactly these assortment questions against live systems A.Team built on their own data. What is not yet a published number is a booked delist-and-reallocate outcome; that captured figure is the next milestone, not a result we're claiming now. The adjacent lever gives a sense of the scale in the same data foundation: on a related engagement the pricing model surfaced nine figures of opportunity in a single market, though pricing is a different lever from assortment and we're precise about the difference. The assortment loop runs on the same owned-data foundation and the same person-in-the-loop discipline; what it hasn't done yet is post its own headline number, and we'd rather say that than borrow one.

The tidy cut list is the easy version of assortment. The store-cluster, constraint-aware one is where the distribution you didn't mean to lose gets protected. Edit the range like a portfolio, not a ranking.

See how the planning intelligence system works →

A.Team AI Solutions builds intelligence systems for Fortune 500 consumer brands. The engagements referenced are anonymized to role and business unit.