Data lineage and provenance record where data came from, how it was transformed, and what depends on it. This is useful when teams need to explain, audit, or safely refresh datasets that feed reports and AI systems.
Whilst lineage diagrams look complete in governance decks, in practice critical paths are often undocumented until something breaks. For example, an embedding job silently reads a superseded export and answer quality drops with no owner to blame. We often recommend tracing gold outputs first — tables and published artifacts — and enforcing breaks when upstream trust fails, before mapping the entire estate.
Lineage answers “what produced this?” Provenance answers “can we trust and attribute it?” Together they sit under lakehouses, feature tables, and any corpus you feed to models. Without them, refreshes and schema edits become silent regressions. Practical systems record job IDs, table versions, model/prompt versions, and the downstream artifact that depends on them. Start on the critical path (gold tables and published outputs); expand upstream once that path is visible and enforced.