AI assortment optimization is the practice of deciding which products belong in which stores, using models of demand, substitution and transference instead of a twice-yearly range review in a spreadsheet. It fits the product mix to each store or cluster against local demand, shelf capacity and the category's financial targets, and estimates what happens when an item is delisted and shoppers move to their next choice. The upside is quantified. Industry research cited by Databricks puts a 10% to 20% improvement in supply chain forecasting accuracy at roughly a 5% reduction in inventory costs and a 2% to 3% increase in revenues. For a Head of Category Management, a CDO or a data lead in retail or CPG, the opening question sits before the vendor shortlist. Can your history carry the models, and should you license a tool or build on the platform you already run?
What AI Assortment Optimization Actually Does
Three mechanisms do most of the work. Demand forecasting at the level of one SKU in one store, granular enough to show that a 500 g pack sells out in one catchment and stalls in another. Demand transference, which models where sales go when a product leaves the range, whether the shopper takes a substitute you stock, trades down or leaves with a smaller basket. And store clustering built on observed demand behaviour rather than region or floor area, so a commuter branch and a city centre store can share a range head office would never have paired.
The optimization step then picks the range that maximizes a stated objective, usually category margin or sales per linear metre, under real constraints: planogram space, supplier terms, own-label targets. Shelf capacity binds hardest in physical formats, where localized assortment planning meets shelf space optimization analytics. Our AI-driven assortment optimization services follow that order: forecast, transference, clusters, then the constrained choice.
Why Spreadsheet Range Reviews Break at Modern Scale
Take a chain of 600 stores carrying 30,000 active SKUs, with four seasonal resets and weekly price changes. That is millions of keep, cut or localize decisions a year, so the work collapses into averages. Regional groupings stand in for demand behaviour, and last year's promotion plan gets copied because unpicking it would take a quarter. AI assortment optimization replaces the averages with an estimate per SKU and per store, the level at which the decision is made.
The expensive failure is the wrong delisting. SKU rationalization in a spreadsheet ranks products by rate of sale and cuts the bottom decile, which silently assumes the lost units reappear elsewhere in the range. Often they do; sometimes the cut item was the reason a shopper visited, and the damage surfaces two categories away, in a basket that no longer happens. A transference model puts an error bar on that risk before the planogram is signed. Our guide on moving from assortment chaos to AI-driven category management covers the operating model around that shift.
The Evidence: What Better Forecasts and Assortment Are Worth
Google Cloud's summary of the top ten AI use cases in retail reports that for specialty retailers five of the ten sit in demand planning and merchandising, assortment, inventory and markdown optimization among them. The same write-up notes that AI forecasts granularly enough to cover new and short-lifecycle products, where manual range reviews are weakest.
The value of AI assortment optimization splits into two halves, and the forecasting figures above size the upstream one: a 10% to 20% accuracy gain associated with roughly 5% lower inventory cost and 2% to 3% higher revenue. Databricks builds its accelerator around the mechanism behind that number, since retail demand forecasting pays when it runs at the fine-grained store and item combination rather than at aggregate level, the same unit of decision assortment planning works in.
Read all of it as potential rather than entitlement, available where data supported the models and category teams acted on the output. A recommendation that never reaches a planogram returns nothing.
Data Readiness: What Your History Must Contain
Five conditions decide whether assortment optimization in retail is feasible on your data.
- Transaction-level POS, basket by basket and not weekly totals, because substitution shows only in what shoppers bought together and instead of each other.
- One product hierarchy reconciled across ERP, POS and planogram systems, since a model cannot group products that three systems name differently.
- On-shelf availability, or a credible proxy, so an out of stock is not read as absence of demand.
- Promotion and price history on the same calendar, otherwise the model learns that discounting is the category.
- One to two full seasonal cycles, so a range decision does not rest on a single Christmas.
Theory backs the instinct. A 2025 arXiv paper on learning an optimal assortment policy under observational data studies the constraint directly: choosing an assortment under a multinomial logit model from nothing but historical records of what customers picked. Its central result is a condition the authors call optimal item coverage, meaning the items of the optimal assortment appear often enough in that history. That condition is necessary and sufficient for efficient offline learning, and the paper pairs it with a near-optimal algorithm, Pessimistic Rank-Breaking, that works under it.
The translation is blunt: if a product never sat on the shelf, no model can tell you what it would have done there. Coverage and quality of history set the ceiling on what AI assortment optimization can find without expensive live experiments in trading stores, and no licence removes that ceiling. Readiness work therefore earns its own phase rather than a checkbox in a vendor questionnaire.
SaaS Tool or Custom Build on Your Data Platform
Both answers are legitimate and the honest comparison has no villain. Packaged assortment optimization software encodes a proven category process, ships planning screens your managers can use next month, and spares you the upkeep of a forecasting stack. A custom build earns its keep when the process, the data or the licence economics do not fit that mould. The signals below sort most cases.
| Decision signal | Favours a SaaS tool | Favours a custom build |
|---|---|---|
| Category process | Standard range review, familiar KPIs, few exceptions | Rules specific to your format, franchise or supplier model |
| Data platform | History scattered across systems, no platform of your own | A lakehouse in production with POS and supply data landed |
| Model ownership | Vendor IP is acceptable; you rent the logic | Models and features are assets you want to own |
| Cost curve | Predictable at your SKU, store and user counts | Licence cost scales with those counts faster than value |
Most large retailers land in between, and the hybrid is undersold: a tool for the long tail of categories where the standard process works, a build on your own platform for the few that carry the P&L. Whether AI assortment optimization arrives as a licence or as code, it has to sit beside the pricing, promotion and supply data it depends on. Teams already on a lakehouse usually build the category management platform there instead of shipping extracts nightly into somebody else's cloud.
From Pilot to P&L: Proving Assortment AI Works
Set the baseline before the first model runs: category margin, stock turn and on-shelf availability for the categories in scope, over a window long enough to contain a promotional cycle. Without it, every later conversation becomes an argument about what would have happened anyway.
Then run the change as a test rather than a rollout, with pilot stores against control stores matched on format, catchment and trading pattern, not on convenience. The measure that decides an assortment case is sales retained after delisting: how much volume from removed SKUs the remaining range absorbed. Margin and stock turn follow. Name the confounders in advance, because range rarely changes alone. If AI-driven dynamic pricing in retail runs in the same window, the read on both blurs unless the store groups are split with that in mind.
An AI assortment optimization pilot earns its next budget on the P&L line, not on the model card. A steering committee shown a four-point gain in forecast accuracy learns nothing about whether the range decision made money, and the habit teaches the organization to reward model work over shelf outcomes. Report the financial result, keep accuracy as an engineering diagnostic, and state the uncertainty honestly.
What a Technical Delivery Partner Does That a Tool Doesn't
AI assortment optimization is a data engineering programme with a model at the end of it. Most of the effort lands before anything is trained: reconciling hierarchies, repairing availability signals, joining promotion history to the right price, then building pipelines that keep it current. Tool vendors assume that layer exists; a delivery partner builds it, and the same pipelines go on to serve replenishment and pricing.
Ownership is the second difference. Models built on your platform stay in your repository, retrainable by your team and portable if you change vendor. A category manager who understands why the model wants to cut an SKU will argue with it, and adoption depends on that recommendation reaching the workflow already used for range reviews, not a separate portal.
DS Stream works as a technical delivery partner across AI solutions for retail and CPG; our Lorenz Polska category management case study shows the pattern, with category management analytics replacing spreadsheet-bound work in a live FMCG business. If you are weighing a licence against a build, or suspect the data is the real blocker, talk to our retail AI team about a readiness assessment on one or two categories.
Frequently Asked Questions
Do we need a data lakehouse before starting assortment optimization?
No. AI assortment optimization needs transaction-level sales history, a reconciled product hierarchy and promotion records you can join to both; a pilot can run from a well-governed warehouse or a curated extract. A lakehouse pays for itself once assortment, pricing, replenishment and supply models share features, because maintaining four copies of one history is where the cost accumulates.
How is assortment optimization different from demand forecasting?
Forecasting predicts what a product will sell if you stock it. AI assortment planning uses those forecasts to decide which products to stock at all, which forces it to model what happens to demand when something is removed. That second half, substitution and demand transference, is what a forecasting engine on its own does not answer.
Can AI recommend assortment for new products with no sales history?
Yes, with wider error bars than for established lines. The usual approach describes the new item by attributes such as brand, pack size, price tier and flavour, then borrows the demand curve of analogues that share them. That works well where attribute data is dense and badly where the product resembles nothing on the shelf. Expect a better read on where to list the item than on how much it will sell.
How long before AI assortment optimization shows measurable results?
Data condition drives the answer more than modelling time. Where history is clean, a pilot on one or two categories can give a first measurable read within a couple of planning cycles, since the result has to appear on the shelf and one reset is not a trend. Where hierarchies and availability data need repair, that groundwork dominates the timeline, and any supplier quoting a fixed date before seeing your data is guessing.
.webp)
