Rethinking Tabular Foundation Models for Decision Making with Enterprise Data: Ontology Renormalization for Transferable Representations

The next frontier for enterprise AI will include flexible, pervasive systems that translate enterprise-level objectives into automated decision-making optimized for those objectives. Building such systems at scale requires foundation models designed for enterprise data, yet enterprise tables are fragmented across clients, campaigns, and operational units, with heterogeneous and often non-overlapping schemas, while private enterprise tables are generally inaccessible as training data to general-purpose foundation models. We address this challenge through $\textit{ontology renormalization}$: rather than forcing heterogeneous schemas into a shared feature vocabulary, we map each row into a fixed set of $G$ persistent, objective-aligned groups. We pretrain a tabular backbone with self-supervised grouped tabular JEPA (GT-JEPA) and study the framework in healthcare engagement, where new programs frequently face severe cold-start and sparse feedback. On proprietary program tables, leave-one-program-out evaluation with a frozen backbone and label-only hold-out shows that GT-JEPA improves engagement-propensity ranking relative to the same grouped encoder without SSL, providing evidence that the learned representation transfers across programs within the same enterprise and can effectively support downstream decision-making.