ai-ready data checklist
Eighteen yes-or-no checks for the data behind one AI use case. Score it before you build — not after the demo.
AI-ready data noun — data a model can use for a specific job without producing confident, wrong answers: resolved, defined, traceable, current, governed and evaluable for that use case. Accuracy is assumed; readiness is checked. Without it, Gartner forecasts most AI projects get abandoned.
Check the road before you drive it.
Pick one AI use case — not your whole data estate — and answer each check for the data that use case needs. Yes: in place and working. Partly: exists, but incomplete, manual or unowned. No: not in place, or nobody knows.
Yes 2 · Partly 1 · No 0
36 points → scored out of 100
Scoring needs JavaScript. The checks themselves work on paper — print this page.
Three readings.
Not ready
Fix the first mile before you build. A model will still answer — fluently, and wrongly — and nothing will alert you.
Pilot, with named gaps
Run a bounded pilot. Write every “partly” and “no” down with an owner and a date, and don’t scale until they close.
Ready to build
The road is paved for this use case. Keep the evaluation set running after launch — readiness decays.
One rule overrides the total: a property scoring zero is a blocker, whatever the overall number says.
Accuracy is not on the list.
Accuracy is assumed, not sufficient. Correct data fails a model all the time: the same customer under four IDs, a field four teams define four ways, a value nobody can trace. Each check here targets one of the six properties of model-ready data.
They map onto the three cracks where the first mile breaks:
| Crack | What breaks | Properties that close it |
|---|---|---|
| Quality | Errors and duplicates born at the source | Resolved · Current |
| Provenance | Origin and consent lost in the join | Traceable · Governed |
| Readiness | Siloed, undefined, unowned, untestable | Defined · Evaluable |
Data readiness, answered.
What is AI-ready data?
AI-ready data is data a model can use for a specific job without producing confident, wrong answers. On this site it means six properties: resolved, so one real-world thing is one thing in the data; defined, so every field has one written meaning and an owner; traceable, so its origin can be reconstructed; current enough for the decision it feeds; governed, so policy travels with the data; and evaluable, so you can tell whether using it worked. Accuracy is assumed, not sufficient.
How do you assess data readiness for AI?
Assess it per use case, not for the whole data estate. Pick one AI use case, list the data it needs, and check that data against each property — for example with the 18 yes, partly or no checks on this page. Score it, find the weakest property, and fix that before you build, because a model will not warn you when its input is not ready.
What is the difference between data quality and AI-ready data?
Data quality asks whether a value is true. AI-ready data asks whether a true value will produce a right answer for a specific model and question. Data can pass every quality check and still not be ready: duplicated across systems, defined differently by each team, untraceable, stale for the decision, or copied into a prototype the governance rules never reached.
Who should fill in an AI-ready data checklist?
The people on both sides of the gap, together: the team building the AI use case and the owners of the source systems that produce its data. The checklist is most useful as a conversation — every “partly” or “no” should end with a named owner and a date.