csat-sentiment-triage-automation

CSAT & Sentiment Triage Automation Skill Guide

This skill governs the integration of machine learning and LLM-driven sentiment evaluation to prioritize frustrated customers, flag churn indicators, and trigger automated escalations.

1. Sentiment & Churn Triage Architecture

[ Incoming Customer Message / Survey ]
                  |
                  v
[ Fast Sentiment & Intent Classifier ]
  |-- Valence: Negative / Neutral / Positive
  |-- Urgency Score (1 - 10)
  |-- Churn Risk Triggers: "cancel", "refund", "unacceptable", "lawyer"
                  |
                  +---> High Urgency / Severe Negative (Score >= 8)
                  |       |
                  |       v
                  |     [ Immediate Manager Slack Alert & P1 Queue Boost ]
                  |
                  +---> Mild / Standard Query
                          |
                          v
                        [ Normal Agent Assignment / AI Auto-Draft ]

2. Production Implementation (Python & Pydantic)

import json
from typing import Literal
from pydantic import BaseModel, Field


class SentimentTriageResult(BaseModel):
    sentiment: Literal["VERY_NEGATIVE", "NEGATIVE", "NEUTRAL", "POSITIVE"]
    urgency_score: int = Field(ge=1, le=10, description="1=low, 10=immediate crisis")
    churn_risk: bool
    detected_issues: list[str]
    suggested_action: Literal["ESCALATE_TO_MANAGER", "PRIORITY_SUPPORT", "STANDARD_QUEUE", "OFFER_RETENTION"]
    executive_summary: str


def build_triage_prompt(customer_message: str, account_tier: str) -> str:
    return f"""You are an expert customer experience triage intelligence agent.
Analyze the following customer message and account context:

Customer Account Tier: {account_tier}
Message Content:
\"\"\"{customer_message}\"\"\"

Provide an objective assessment in JSON matching the schema:
- sentiment: VERY_NEGATIVE | NEGATIVE | NEUTRAL | POSITIVE
- urgency_score: 1 to 10
- churn_risk: true if the customer expresses intent to leave, cancel, or switch competitors
- detected_issues: list of core issues (e.g., "downtime", "billing error", "rude support")
- suggested_action: ESCALATE_TO_MANAGER | PRIORITY_SUPPORT | STANDARD_QUEUE | OFFER_RETENTION
- executive_summary: 1-sentence summary for the receiving agent
"""


def process_sentiment_webhook(triage_data: SentimentTriageResult, ticket_id: str):
    """Executes action based on structured sentiment classification."""
    if triage_data.urgency_score >= 8 or triage_data.churn_risk:
        # Boost ticket priority and notify escalation Slack channel
        print(f"[URGENT ESCALATION] Ticket {ticket_id}: {triage_data.executive_summary}")
        # Integration logic: emit webhook / update CRM