EXPRESS: Governed Agentic AI for Retail Baskets: A Consumer World Model with Inventory-Aware Actions

Retail recommendation systems increasingly operate as real-time decision engines that must personalize suggestions while respecting operational constraints such as inventory availability, category rules, and promotion policies. This is especially challenging in basket-based retail because transactions are set-valued and the observed checkout order is operational rather than behavioral. We study a governed agentic AI system for basket recommendation in which a Bayesian consumer world model serves as the agent's internal state. The model maintains calibrated beliefs over latent shopping profiles and updates them online as items are observed while representing basket context in an order-invariant way. The agent then selects recommendation slates under explicit governance, combining interpretable control levers (e.g., profile-context trade-off, bounded exploration) with operational guardrails and feasibility masking (e.g., in-stock status, category, promotion eligibility). Using large-scale grocery transaction data, we evaluate the framework in both offline next-item prediction and an operations-coupled simulator with inventory and promotion dynamics. The agent achieves higher hit rate than a non-agentic variant as well as various strong item-item baselines under a common holdout protocol. In the simulation, ranking accuracy changes little, yet the agent delivers substantial gains in revenue and inventory productivity by steering demand toward feasible complements and enabling controlled substitution when constraints bind. This highlights an operations insight: under binding feasibility, decision quality can improve through constraint-aware substitutions and inventory coupling even when conventional ranking metrics remain unchanged. Overall, the results show how basket-aware demand models can be deployed as governed, agentic policies that coordinate personalization with operational objectives.