Improving Enterprise Data Product Discoverability through AI-Driven Metadata Enrichment: A Real-World Case Study

Enterprise data catalogs increasingly rely on metadata to support data product discoverability, yet metadata quality and semantic misalignment remain major obstacles in real-world settings. This study investigates how artificial intelligence techniques can be applied incrementally to improve metadata-driven data discovery under realistic enterprise constraints. We propose a step-by-step experimental framework that progressively introduces business glossary expansion, semantic retrieval using Sentence-BERT, LLMgenerated data product descriptions, synthetic query generation, model fine-tuning, and LLM-based reranking, starting from a traditional keyword-based baseline. The framework is evaluated on a real-world enterprise metadata dataset using consistent, industry-aligned retrieval metrics. The results show that the transition from lexical to semantic retrieval yields the largest observed performance gains, substantially improving both precision and coverage. Metadata enrichment techniques provide incremental benefits, while more advanced approaches exhibit diminishing returns and strong dependence on data quality. The findings further indicate that task-aligned, fit-for-purpose models can outperform larger, general-purpose alternatives in enterprise metadata retrieval. Overall, the proposed framework offers practical guidance for incremental AI adoption in enterprise data discovery and is designed to be adaptable across organizations with different levels of data maturity, resources, and technical expertise.