Data lineage is the paper trail behind a piece of data: which source system it came from, what transformed or combined it along the way, and which downstream system, model, or AI answer it ended up feeding. Without it, a company that gets asked what its AI was trained on, or why it answered the way it did, has no way to reconstruct the answer.
This stopped being optional the moment regulation caught up with deployment. The European Union AI Act specifically requires clear data lineage and training-set documentation for higher-risk systems, and an unmapped pipeline, one nobody can trace back to its source, invites real fines regardless of whether anything actually went wrong.
In practice, lineage gets built two ways: upstream, by tagging and tracking data from the moment it’s ingested, and downstream, by logging which user queried what and which retrieved chunks fed which generated answer. Both records need to exist before an incident, not reconstructed after one.