Poor Enterprise Data Governance Can Sink AI Before Inference

Summary

The article argues that enterprise AI projects often fail because the data layer lacks consistent meaning, lineage, and freshness, rather than because of model quality or computing capacity. It describes a common situation in which two customer systems use different identifiers and fields, contain contradictory mappings, and record timestamps inconsistently. Human employees may know the historical exceptions and system failures, but an AI system cannot reliably connect the records when their relationships were never explicitly captured. The author says embeddings, vector search, and retrieval-augmented generation cannot automatically reconstruct missing semantics. The article applies this problem to entity resolution, data lineage, and time-to-live controls: incorrect mappings can attach information to the wrong customer, broken pipelines can make stale data appear current, and expired inventory information can be presented as real-time. In the author’s view, AI amplifies fragmented, duplicated, stale, or semantically inconsistent data rather than cleaning or reconciling it. The article also argues that more capable AI increases the need for people who understand undocumented business history, unreliable systems, and the intent behind fields. It recommends fixing lineage, cleansing, unification, standardization, TTL discipline, entity resolution, and semantic consistency before investing further in enterprise AI, RAG, or inference workflows.