Most failed AI projects were failed data projects wearing a better suit.

Read enough internal post-mortems and the pattern stops being interesting and starts being tedious. The model was fine. The pipeline was fine. The definition of a customer was not.

This is an old finding from the data warehouse era, rediscovered by a new generation with better tools and the same organisational chart.

The specific failures

A field that three systems populate with different conventions. A status code whose meaning changed in 2019 and was never backfilled. A rule that exists only in the head of a person in the operations team who applies it by eye.

An automation project meets all three in its first fortnight and usually responds by building exceptions rather than by resolving the underlying ambiguity.

Why it persists

Because resolving it requires a decision that crosses departmental boundaries, and no automation budget is authorised to make that decision. The technical team can see the problem precisely and cannot fix it.

What changes it

Naming an owner for each critical data definition, at a level senior enough to settle disputes, and treating that appointment as a prerequisite for the automation project rather than a workstream inside it.

It is less exciting than a model launch. It is also the difference between a pilot and a production system.