Can an AI cart fix the holiday substitution problem? Kroger is testing the idea across its family of banners this season, pairing an AI Shopping Assistant that builds a cart from a photo, a link or a typed list with a prepared-foods range sized for small gatherings. The chain runs roughly 2,700 stores, serves more than 11 million customers a day and offers curbside pickup plus delivery in as little as 30 minutes, with on-demand access through DoorDash, Instacart and Uber Eats. As of November 17, 2026, the assistant is live across the group websites and apps, and the read for brands is narrower than the announcement suggests.
The cart is the new shelf audit
Where the cart meets the shelf
When a cart is generated from a recipe, the shopper is no longer browsing. They are accepting a list. That shifts the moment of selection upstream: the SKU either fits the recipe parameters or it is skipped, and skipped quietly. For a brand, being unavailable at that moment looks nothing like being out of stock on a shelf, because there is no empty slot to notice.
This is why the feature matters beyond convenience. It turns product data into a selection filter. Ingredient lists, pack sizes, allergen tags and dietary claims have to be machine readable, or the assistant will route around the product when it scales a recipe up or down. Retailers have spent years cleaning this data for search; a cart builder raises the cost of getting it wrong.
Four fulfilment paths, four failure modes
Kroger routes the same basket through four completion paths, and each one fails differently. In-store fails on findability, curbside fails on timing, delivery fails on temperature and availability, and marketplace partners fail on substitution rules that the retailer does not fully control. A single cart can hit any of them.
| Path | What the shopper sees | What operations must guarantee |
|---|---|---|
| In-store | Aisle and shelf position | Planogram accuracy during seasonal resets |
| Curbside | A pickup window | Pick accuracy and cold-chain timing |
| Own delivery | A 30-minute estimate | Stock confidence at the store level |
| Marketplace | Partner pricing and fees | Substitution rules and refund handling |
The fourth row is the one brands underestimate. On a marketplace, the substitution decision may follow the platform rules rather than the retailer plan, and the resulting basket is recorded against a partner, not the store. Availability insights gathered from those orders are partial by construction.
Prepared food as a basket anchor
The ready-to-heat range serves a different purpose than the assistant. Single-serve sides and sliced turkey are built to reduce waste for smaller households, and they anchor a basket that would otherwise be a handful of ingredients. Pairing them with AI-generated lists is a way to lift the average order without discounting, which is the lever most grocers would rather pull in a season when shoppers say price is the main constraint.
What to measure before the season ends
Two numbers tell you whether this works. The first is completion: what share of generated carts convert to a picked order without edits. The second is attach: how often a prepared item lands in a cart that started as a recipe. Downloads and feature opens are noise by comparison.
For suppliers, the equivalent test is coverage. Take the ten recipes most likely to be uploaded in your category and check whether your SKUs survive the scaling step at two, eight and twelve servings. If pack sizes force a substitution at eight, that is a listing problem you can fix before the next holiday, not a demand problem.
FAQ
Q: Can a cart builder really reduce substitutions?
A: It can reduce the obvious ones. If a recipe is scaled for eight people, the cart asks for enough stock up front. Substitutions still happen when the shelf is empty, which is an availability problem, not a planning problem.
Q: Is this a pricing story or a convenience story?
A: Both, with different owners. Convenience sits with merchandising and digital product. Price perception sits with the shelf, because a cart built on a budget still fails if the pickup total lands higher than the estimate.
Q: What should suppliers prepare?
A: Structured product data. Ingredient lists, allergen tags, pack sizes and dietary claims need to be machine readable, or the assistant will route around the SKU when it scales a recipe.
Q: How should results be judged?
A: On attach rate and completion rate rather than downloads. A cart that converts to pickup without edits is worth more than a feature that is opened once and abandoned.











