Tarka
Back to Fundamentals
Intercom
Config

Auto-Tag Intercom Tickets

Auto-tag and categorize Intercom conversations by topic, severity, and sentiment

Instructions

Auto-Tag Intercom Tickets

Programmatically classify and tag Intercom conversations by category, severity, and sentiment so downstream analysis drills can aggregate patterns.

Prerequisites

  • Intercom account with API access (scopes: read conversations, write conversations, write tags)
  • Tags pre-created in Intercom for your taxonomy
  • Conversation data accessible via intercom-conversations-export fundamental

Step 1: Create your tag taxonomy

Use the Intercom Tags API to create a consistent set of tags:

POST https://api.intercom.io/tags
{
  "name": "category:bug"
}

Recommended taxonomy:

  • Category tags: category:bug, category:feature-request, category:billing, category:how-to, category:integration, category:performance, category:access, category:data-loss
  • Severity tags: severity:critical, severity:high, severity:medium, severity:low
  • Sentiment tags: sentiment:frustrated, sentiment:neutral, sentiment:positive
  • Churn signal tags: signal:churn-risk, signal:escalation, signal:repeat-issue

List existing tags to avoid duplicates:

GET https://api.intercom.io/tags

Step 2: Classify conversations with LLM

For each conversation, extract the first customer message body and pass it to an LLM for classification:

POST https://api.anthropic.com/v1/messages
{
  "model": "claude-haiku-4-20250414",
  "max_tokens": 500,
  "messages": [{
    "role": "user",
    "content": "Classify this support ticket. Return JSON only.\n\nTicket text: \"{first_message_body}\"\n\nClassify:\n- category: one of [bug, feature-request, billing, how-to, integration, performance, access, data-loss]\n- severity: one of [critical, high, medium, low]. Critical = data loss or complete outage. High = major feature broken. Medium = workaround exists. Low = cosmetic or question.\n- sentiment: one of [frustrated, neutral, positive]\n- churn_signals: array of any that apply: [churn-risk, escalation, repeat-issue, competitor-mention, cancellation-intent]\n\nRespond: {\"category\": \"\", \"severity\": \"\", \"sentiment\": \"\", \"churn_signals\": []}"
  }]
}

Use claude-haiku-4-20250414 for cost efficiency at volume. Estimated cost: ~$0.001 per ticket classification.

Step 3: Apply tags via API

POST https://api.intercom.io/tags
{
  "name": "category:bug",
  "conversations": [{"id": "{conversation_id}"}]
}

Apply each relevant tag from the classification. A single conversation may receive multiple tags (e.g., category:bug + severity:high + sentiment:frustrated + signal:churn-risk).

Step 4: Batch processing

For historical backfill, iterate over all conversations from the past 90 days:

  1. Use intercom-conversations-export to pull conversations
  2. Filter to those without category tags (avoid re-tagging)
  3. Classify each with the LLM
  4. Apply tags in batch

Rate limit: process ~100 conversations per minute to stay within Intercom API limits.

Step 5: Real-time tagging via webhook

For ongoing classification, set up an Intercom webhook for conversation.created events:

POST https://api.intercom.io/subscriptions
{
  "service_type": "web",
  "url": "https://your-n8n-instance.com/webhook/intercom-ticket-tag",
  "topics": ["conversation.created"]
}

The webhook payload contains the conversation ID. Your n8n workflow fetches the full conversation, classifies it, and applies tags within seconds of ticket creation.

Error Handling

  • If LLM classification fails, tag as category:unclassified and queue for retry
  • If Intercom tag API fails with 404, the tag was deleted; recreate it first
  • Log all classifications for audit: conversation_id, classification result, confidence