636,792 rows → 207,997 sendable contacts. In one day.

Our own performance-marketing group had 453 raw CSV exports from an SMS platform sitting in a folder: 636,792 rows of contacts and campaign history, spread across files in inconsistent formats. Before a single message goes out, three questions have to be answered. Which contacts are actually sendable? Who has already clicked and should be suppressed? And what does the historical data say about what actually works?

The usual answer is a week in spreadsheets: manual stitching, lookups against clicker lists, and mistakes that only surface after a send has already gone wrong. We wanted the answer the same day, and we wanted it verified before we trusted it.

We ran it as an orchestrated AI pipeline, using our standard method: one orchestrator plans the work and splits it up; a fleet of parallel workers each handle a slice at the same time; and a set of independent QA gates check the result before anything is accepted. Our builds typically fan out across 20–45 parallel AI workers under one orchestrator.

The work split into three tracks:

• Parsing: 453 CSV files read and normalized, with per-file logging. Zero failures — every file was either parsed cleanly or flagged, with no silent data loss.
• Dedup and suppression: the rows were de-duplicated and checked against 73,623 known clickers, who were suppressed. What remained was 207,997 sendable contacts out of the 636,792 raw rows.
• Analysis: 278 historical campaigns covering 1,654,905 recipients were run through for delivery and click rates, assembled into a dashboard with no personal data and not a single external request.

Each track had its own check. Four independent gates had to sign off: that every file had parsed, that duplicates were gone, that the clickers were removed, and that the final count reconciled. 4/4 passed, and only then was the output accepted.