Journal / Kirill Naumov
AI for Retail Assortment Planning: A Better List Is Not Always a Bigger One
How retailers can apply AI to assortment decisions while balancing customer choice, substitution, availability, margin and store-team judgement.
By Kirill Naumov ·
The assortment question is not simply “Which products sell most?” A retailer has limited shelf space, working capital and attention from the customer. The job is to choose a range that works for a particular customer, channel and location—and to understand what is lost when one item is removed or unavailable. AI can help surface patterns. It cannot make those trade-offs disappear.
Give the model a decision it can actually inform
Start with a category and a specific decision: which variants to carry, how much space to allocate, or what substitute to show when a product is missing. Connect sales to price, promotions, returns, availability, season, store format and online behaviour. A fast-selling item with frequent markdowns may be less attractive than its headline volume suggests. Likewise, a niche item may complete a basket or serve a customer group that a sales-only ranking overlooks.
Separate demand from availability
Observed sales are not the same as customer demand when products go out of stock, are displayed differently or are hard to find. Check whether the model sees the assortment that was actually offered at the time and place of purchase. Keep product hierarchy and variant identity clean. If a blue size-small item is treated as interchangeable with a blue size-large, the resulting “insight” will be very confident and very unhelpful.
Make substitution visible
Removing an item does not mean every one of its buyers switches to the same alternative. Some choose another brand, some choose another size, some leave the category, and some leave the store. A 2026 Operations Research paper using Instacart data studies assortment optimization under stockout risk and replacement options; it is a specific analytical study, not a forecast of results for any one retailer: https://pubsonline.informs.org/doi/10.1287/opre.2024.1074. The operational question is whether your own substitution data is good enough to support the decision.
Keep the person who owns the shelf in the loop
A Management Science field experiment with a retail business examined algorithm recommendations for product assortment and worker adoption. Its authors report that staff willingness to follow advice varied with context and how the algorithm used workers’ forecasts: https://pubsonline.informs.org/doi/10.1287/mnsc.2020.3599. That points to an important design choice: show why the system recommends a change, invite a category manager to explain local constraints, and record overrides as information—not insubordination.
Test the change, not the dashboard
Pilot a controlled change in comparable stores, categories or periods where possible. Agree in advance on margin after markdowns, availability, waste, substitution, basket effects and customer complaints. Track the operational effort required to maintain the range too. If sales move, check whether price, promotion, weather, competitor activity or stock changed at the same time. Expand only when the outcome is worth the additional complexity and the explanation makes sense to the people executing it.
This is a practical framework, not a report of a live client deployment. The [AI for Retail workshop](https://www.kirillnaumovai.com/workshops/ai-retail-november) covers how to choose and test practical AI applications across retail, sales, learning and loyalty.