AI on bad data accelerates uncertainty. Reports disagree; models still advise fluently—harder to challenge. One-time scrubbing resets dirt in two weeks. Missing piece is minimum loop: dirty data found, definition owners, permission boundaries, fixes back to source documents.
Do not launch AI first. Make one core table trustworthy. Trustworthy means disputes have an owner—not perfect rows.
What Smart on Dirty Data Does
One customer, many codes—model mixes overdue and good. Two output definitions—prediction fails in the meeting. Permissions open—sensitive fields in prompts; compliance stops the project. Intelligence amplifies dirty data.
Outsourced one-time clean is worse—no owner after; tables revert; business stops believing any AI conclusion.
What Minimum Loop Looks Like
- Pick one operating metric—define source, refresh, owner.
- Daily quality rules—nulls, duplicates, cross-system mismatch—below threshold blocks dashboards.
- Fixes return to source documents—no "clean table only in warehouse."
- Field-level permissions—export and training default minimum; expansion needs approval.
The XYN digital intelligence system lands business actions in configurable apps—definitions and permissions follow documents. When data quality is poor, spin the loop first. Then talk AI. Otherwise intelligence is expensive guessing.
