Why LLMs Aren’t a One-Size-Fits-All Solution for Enterprises

Large Language Models (LLMs) are not optimized for the structured, relational data that drives key business outcomes, leading to high failure rates for enterprise GenAI pilots. LLMs struggle with this data as they were designed for unstructured text, treat numbers as tokens, and can hallucinate, making them unreliable for critical workflows. As an alternative, a new architecture called a Relational Foundation Model (RFM) is proposed, which uses Graph Neural Networks and Graph Transformers to reason directly over relational databases. This model can be prompted to make accurate predictions about entities like customers and transactions without the extensive feature engineering of traditional machine learning. The discussion also covers other industry approaches, such as internal foundation models, AutoML platforms, and graph databases.