Problem
A grocery chain with 2,200 stores, about 31,000 items per store and 624 categories must decide for every item, store and week what to promote and at what price. Price, promotion, space and assortment sit in separate systems, so each is optimised locally. There is no clear way to predict units, revenue and adjusted gross profit (AGP) for a given markdown in a given store. Digital, store and personalised offers stack, and merchants reject models they cannot audit.
Example from the article: a merchant asks what happens to Sprite, to the chips next to it and to the basket two weeks later if Coke is pulled from the promo.
Solution in brief
Albertsons runs five stages: sense, predict, decide, explain, act. A custom deep-learning model with an item-interaction graph predicts units and revenue. A mixed-integer optimiser picks the promotion under guardrails. The LLM only phrases the result in merchant language.
Questions this case answers
- How to recommend item-level promotions for each item, store and week by predicting units and revenue, including effects on substitutes and complements, and then optimizing the choice under business guardrails
- How to keep LLM output out of pricing decisions by using the language model only to explain structured model results to decision makers
- How to recommend item-level promotions at store-week level when price, promotion and placement live in separate systems
- How to predict the effect of promoting one item on its substitutes, complements and the basket
Tools and methods
- Attention-based deep-learning model — pre-trained on own transactions; learns how price, promotion and placement of one item affect others.
- Graph neural network, "Promo Molecules" — signed item-to-item relationships: substitute pressure, complement lift, halo, cannibalisation, timing, basket movement.
- Mixed-integer (quadratic) programming, Gurobi — chooses promotion type and depth under guardrails; deterministic output.
- LLM explanation layer (OpenAI ChatKit, Genie) — turns scored interactions into merchant-language text.
- Databricks Lakehouse, Spark, Unity Catalog, MLflow — data, training, governance.
Google or a one-time email code. No password.
