09 — Limits & the human loop
Where AI over-promises — and why a human stays in the loop
Three cautions keep the numbers honest. First, most lift figures are self-reported. "+41% CTR," "up to +22% opens," and "+2,361% conversions" come from practitioner and vendor surveys comparing AI-assisted to manual sends, not controlled trials — genuine signals, but ceilings rather than guarantees.Directional Second, open-rate-based AI inherits Apple's ~18% inflation,6 so any model trained on opens is learning from partly-fictional data. Third, deliverability and brand voice still need a human — generic AI copy at volume risks spam signals and sameness.
● The failure mode to avoid
Fully-automated AI send programs with no human review tend to drift toward bland, on-the-nose copy and open-rate-optimized timing that a privacy pre-fetch has already distorted. Keep a person on voice, offer, and click-based measurement; let AI own volume and variants.
The consistent finding across every dataset is that AI amplifies whatever system it's dropped into. On a well-segmented, permission-based, click-measured program it compounds returns; on a batch-and-blast list it mostly automates the sending of email nobody wanted, faster.