video1.0<iframe src="https://www.loom.com/embed/94ac3555a2a24e40b660d54a74a229d2" frameborder="0" width="1920" height="1440" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>14401920Loomhttps://www.loom.com14401920https://cdn.loom.com/sessions/thumbnails/94ac3555a2a24e40b660d54a74a229d2-9753357a5ae55d35.gif312.142This Loom explains how to decide whether an 18-agent collections floor should scale an AI-assisted triage, and how to test it credibly. The author locked 6 KPIs computed from 1,200 accounts over 6 months, identified data issues and cure-rate noise (36 to 49% month to month), then scored build versus extend versus buy across 8 criteria, where build led (3.6) after weighting data readiness and explainability. A guardrail-gated pilot returned a cure-rate lift and improved time to first touch by about a day and a half with zero guardrail violations, but the control cohort matched baseline and the weekly analysis showed the pilot’s lift could not be distinguished from doing nothing due to a confidence interval containing zero. Despite a go gate based on guardrails, the author recommended against scaling beyond wave 1, citing the earlier phase-one mortality impact of about 12 points.