The Outbound Reply Triage Playbook

Replies scatter across sending tools and sit in inboxes nobody owns. Capture every one into a single queue, verbatim.

The words are the product. A paraphrased reply has already lost the thing the whole system exists to collect.

Strip signatures, footers, legal disclaimers, and quoted history, so classification reads the human's words and nothing else.

Most misclassifications are the footer's fault. Clean first, classify second.

Six classes, each with linguistic tells and a numeric confidence. Interested. A question or objection. Not interested. Out of office. Routed away, because people leave companies and treating that as rejection burns a live account. And unclassified, because one letter is not evidence, and the schema can say so.

Watch for low confidence. Below the threshold, a person decides the class, and the human verdict with its reason becomes the calibration data.

The sequence stops the moment words arrive. Any reply pauses the sends, deterministically, before classification even runs. The do-not-contact list splits by how the no arrived. Explicit unsubscribe language lands instantly with no model in the path, and an inferred no waits for the human verdict, because the irreversible action never rests on an inference. One list, honored at every send path you named in the guardrails table.

Watch for the tool that keeps sending while a reply sits unread. Nothing burns a domain and a reputation faster.

An out-of-office is timing, not rejection. The contact pauses until the stated return date plus a bracketed buffer, with dates resolved against the day the reply arrived, then resumes where the sequence left off. Unresolvable dates fall to the default pause.

Interested replies and real questions alert the channel with the verbatim thread and the lead's context, each with a bracketed answer-by time. Anger gets its own alert, read by a person the same day, never answered with a template. A person writes every answer.

Watch for automation drafting replies to objections. An objection answered by a machine reads as exactly that.

Every reply gets recorded with its class, its numeric confidence, the human verdict when one was needed, and the verdict's reason in one line, because the reasons are the calibration data. Three same-shaped disagreements are a taxonomy revision waiting to be written.

Supervised first. The taxonomy earns trust pass by pass, and the objection language you collect becomes next month's better messaging.