Master data management returns wearing an AI badge. The job description has not changed.

Master data management spent a decade as the slide everyone skipped. It is back on the agenda, and the reason given is that the models cannot work without one agreed definition of the things they reason about.

What happened is a rehabilitation, not a discovery. Master data management is the discipline of producing one trusted, semantically consistent view of the entities a business runs on, such as customer, product, and supplier, across systems that disagree about them. It was unglamorous, so it was deferred. Now the same definition is reintroduced as a precondition for model accuracy, which is the polite way of admitting it was a precondition for accuracy all along.

It matters because the failure mode survived the rebrand intact. A golden record exists to consolidate conflicting records from multiple systems into one version a business can act on. When five systems hold three definitions of "active customer," a model trained or grounded on that spread does not resolve the conflict. It averages it, then reports the average with composure.

What it reveals is the oldest pattern in the catalog: the organization would rather buy a capability than assign a definition. Matching and deduplication now arrive with machine-learning trim, useful for the mechanical work of linking records. It does not decide which record is correct. That is a stewardship decision, and stewardship has an owner or it does not exist.

What to watch is the order of operations. If the data-quality program is scoped after the model ships, the definition is being treated as a downstream cleanup task rather than the spec. The reveal is plain: the technology did not create the need for one definition of "customer." It just made the disagreement faster, and gave it a confidence interval.