---
name: outreach-response-classifier
description: Classifies inbound sales or partnership outreach responses into intent, stage, topic, urgency, sentiment, and routing category while preserving the message evidence and detecting opt-outs, support requests, security issues, and ambiguous replies. Use for inbox triage, lead-response analysis, campaign measurement, and response queues.
license: Apache-2.0
metadata:
  adlass.categories: "operations/inbox-triage"
  adlass.industries: ""
  adlass.tags: "outreach,classifier,response-triage,intent,lead-routing,opt-out"
  adlass.adaptation: "mapping"
  adlass.source: "original"
  adlass.version: "1"
---

# Outreach response classifier

## Purpose

Turn a batch of outreach replies into consistent, reviewable classifications that downstream owners can act on. The classifier preserves the respondent’s wording, separates intent from sentiment, and highlights replies that must not be treated as ordinary sales interest.

## Scope

Classify each message for response intent, lifecycle stage, product or topic, urgency, sentiment, language, confidence, suggested queue, and safety flags. Detect positive interest, clarification, objection, referral, not now, unsubscribe, wrong person, out of office, support issue, security or privacy concern, bounce, and unclear responses when those labels exist in the taxonomy. **Excluded:** replying, sending follow-ups, changing consent records, making legal determinations, or assigning an owner not present in the routing rules.

## Data basis

- Message table: message ID, thread ID, timestamp, sender, recipient, subject, body, campaign, account, contact, and existing CRM identifiers.
- Label taxonomy, routing matrix, opt-out policy, language list, sentiment definitions, and escalation rules.
- Prior labelled examples only when they are explicitly approved as a calibration set.
- Optional `classification_scope` input for campaign, date range, or label version.

## Result

Create a classification sheet with one row per message: IDs, primary intent, secondary labels, stage, topic, sentiment, urgency, language, confidence, evidence span, queue, opt-out flag, escalation flag, and rationale. Add a summary document with counts, ambiguous clusters, policy-sensitive replies, and label-quality observations.

## Quality criteria

- Every message receives exactly one primary intent and only allowed secondary labels.
- The rationale quotes a decisive phrase or points to a message span; empty, quoted, and automated replies are handled explicitly.
- Opt-out or do-not-contact language is flagged even when the message also expresses interest.
- Confidence is reduced for sarcasm, multiple intents, language mismatch, thread contamination, or missing body text.
- Counts by campaign, intent, and routing queue reconcile to the classified message count.

## Instructions

Classify the response, not the original outreach copy. Use the latest human-authored message in a thread unless the taxonomy says otherwise, and retain thread IDs for audit. Safety and consent flags take precedence over commercial intent. Never infer urgency from an unread marker or sender seniority. Keep sentiment descriptive rather than diagnostic. When no label fits, use the approved “unclear” value and explain which taxonomy boundary failed.

## Adapt before use

- Supply the label taxonomy, examples, routing matrix, opt-out wording, and escalation policy.
- Map message, thread, campaign, account, contact, and CRM identifiers.
- Define confidence bands, language handling, precedence rules, and allowed output queues.