Molecules, Merchants & Models: How Structure-First AI Makes Promotion Intelligence Trustworthy

The presentation, "Molecules, Merchants & Models: How Structure-First AI Makes Promotion Intelligence Trustworthy," describes a promotion intelligence system built on the Databricks Lakehouse. Retail promotions interact through substitution, complementarity and spillover effects that merchants intuitively understand, but most tools fail to explain.Our system models item and promotion relationships as reusable graph structures—Promo Molecules—learned from over 70 billion rows of transaction data. These structures explicitly surface demand shifts, including substitutes, halo effects and cannibalization. We apply Explainable AI (XAI) at the relationship level—exposing deterministic, merchant-validated signals.Foundation models then reason probabilistically to answer natural-language “what-if” questions. This combined approach enables explainable, merchant-trusted promotion decisions at scale, without manual overrides or black-box predictions.