The AI-Ready Data Playbook

Never boil the whole database. Name the workflow you want to run, then list only the fields and records it will actually read. That list is the cleanup's entire scope.

A whole-CRM cleanup is the project that already failed once. A twenty-field cleanup ships this month. Stale records exit here too, because archiving the dead beats scrubbing them. And every recurring obligation in this kit gets a name in the audit, because unowned cadences quietly stop.

Watch for scope written as table names instead of field names. AI reads fields.

Count it. Duplicate records in scope. Empty required fields. Fields where two systems disagree. Records untouched for a year.

The before-number is the whole point. It makes the cleanup provable, and it is what the re-measure gate at the end reads against.

Write merge rules before merging anything. Which record survives, and for each field, which source wins. Newest wins for titles. The CRM wins for owner. The invoice system wins for revenue.

Rules mean the dedupe never gets done by hand twice, and a new duplicate next quarter meets the same rules instead of a new argument.

Run the field census. For every custom field in scope, who filled it last, when, and what reads it. Unused and duplicated fields are noise, and AI will faithfully read the noise.

Watch for the field someone quietly reports from. The census asks before it deletes.

Cleanup without entry validation refills the mess on a schedule, which is why this step ships its own worksheet. Entry points, field-by-field treatments, and picklist conversions, each with an owner. Required fields only where they earn it, picklists over free text where a picklist is honest.

The test for any new rule is that it makes tomorrow's records cheaper to trust, without making them slower to create.

Fail loud, never plausible. A deterministic check hits a wrong field and throws an error you see. A model hits the same wrong field, adapts, and writes something plausible anyway. Silence is the failure mode.

Three guards, before any tool writes. Loop protection, so enrichment cannot re-trigger itself. A do-not-touch list covering records AND fields, honored at every write path, merges included. And deterministic validation around every AI write, so a schema surprise stops the write instead of bending it.

The before-number must move. Re-run the step 02 counts, with the conflict sample freshly drawn at random, never the records you fixed. The gate is simple. Until the numbers move, no workflow reads these records.

A gate that reads a number cannot be argued with, and it is the difference between a cleanup and a cleanup story.