The Data Readiness Checklist Every AI Initiative Should Pass First

Ownership and definitions

Every critical field needs an owner who can say what it means and when it changes. Where two systems disagree on a definition — a customer, a trade, a policy — that ambiguity will surface as model error later.

• A named steward exists for each source system.
• Business definitions are documented and agreed across functions.
• Conflicting identifiers have a documented resolution rule.

Lineage and reproducibility

You should be able to trace any figure back through every transformation to its source record, and re-run that path to get the same answer. Without reproducibility, model debugging becomes guesswork and audit conversations become uncomfortable.

Quality signals you can monitor

Completeness, freshness, uniqueness, and range checks should be automated tests, not periodic reviews. Treat a failed data test the way you treat a failed build.

• Null and duplicate rates tracked per table.
• Freshness alerts tied to the business refresh cadence.
• Schema changes detected before downstream jobs run.

Access, privacy, and controls

Decide which data can leave the boundary, which must be masked, and which model providers are approved. Retrofitting these controls after a pilot is far more expensive than designing them in.

Refresh cadence versus decision cadence

If the business decides daily and the data lands weekly, no model will fix the gap. Align pipeline frequency with the decision it supports before scoping anything else.