Every reply is classified for intent and sentiment within seconds: positive (interest), neutral (deferral/question), negative (rejection), OOO, and opt-out. Accuracy runs ~95%+; low-confidence classifications escalate rather than guess.
The categories
Positive (buying interest, meeting intent), Question (needs an answer, sentiment neutral), Neutral (not-now, send-info, deferral), Negative (clear rejection), OOO (auto-reply), Opt-out (remove-me). Sub-labels add nuance (positive-scheduling vs positive-curious).
Confidence handling
High confidence → automated handling per the sequence's reply rules. Below the threshold → escalation with the best-guess label attached for one-click confirm-or-correct.
Your corrections teach it
Every re-label you make feeds the classifier. Workspace-specific patterns (your industry's jargon, your prospects' politeness conventions) converge within a few hundred corrections.
Auditing accuracy
Unibox → Settings → Classification shows accuracy on your corrected sample, per category. If a category runs weak (commonly: distinguishing polite-negative from neutral-deferral), tighten its confidence threshold so more of it escalates.
