AI Data Foundation Check

Every AI roadmap I review gets the same first check: whether the data could go in a board deck unchecked.

If the answer is no for humans, it’s no for models. The model scales the error.

My check before any AI proof-of-concept, scored 1-5 each:

1. Completeness. Are the core entities (customers, orders, usage) actually all there, or does each system hold a partial view? Models trained on the CRM’s idea of a customer inherit its blind spots.
2. Freshness. Is there an SLA anyone enforces? An agent confidently quoting last Tuesday’s stock levels is worse than no agent.
3. Agreement. Do finance and product get the same number for the same question today? If not, the AI picks one, at random, with confidence.
4. Ownership. When answers go wrong, is there a named person for the data underneath, or only for the model?

Score below 12 and the honest move is a few weeks of foundation work first. That’s usually cheaper than one failed pilot, and every later AI project inherits it.

Teams that skip the check don’t skip the work. They do it mid-project, under a deadline, with a disappointed sponsor watching.

Would your customer table survive question 3 today?