happy-technologies-llc/detect-conversation-type

Conversation Type Detection and Classification

Overview

This skill detects and classifies the type of conversation occurring in ServiceNow engagement channels (Virtual Agent, chat, email, portal) to enable intelligent routing and handling:

• Inquiry: Information-seeking questions about services, policies, or status
• Request: Actionable service requests or catalog orders
• Complaint: Expressions of dissatisfaction requiring service recovery
• Feedback: Constructive input about services, processes, or experiences
• Escalation: Urgent issues requiring immediate attention or management involvement
• Troubleshooting: Technical problem-solving requiring diagnostic steps
• Follow-up: Continuation of a previous conversation or existing ticket

When to use: When building Virtual Agent topic routing, when enriching interaction records with classification metadata, when automating conversation handoff decisions, or when analyzing conversation patterns for service improvement.

Prerequisites

• Roles: admin, sn_customerservice_manager, itil, or virtual_agent_admin
• Plugins: com.glide.cs.chatbot (Virtual Agent), com.glide.interaction (Agent Workspace Interaction)
• Access: Read access to sys_cs_conversation, sys_cs_message, interaction tables
• Knowledge: Virtual Agent topic design, conversation flow concepts, sentiment analysis basics
• Related Skills: csm/sentiment-analysis for sentiment scoring, csm/chat-recommendation for response suggestions

Procedure

Step 1: Retrieve Conversation Messages

Fetch the conversation history for classification.

MCP Approach:

Tool: SN-Query-Table
Parameters:
  table_name: sys_cs_message
  query: conversation=<conversation_sys_id>^ORDERBYsys_created_on
  fields: sys_id,body,direction,sys_created_on,typed_by,message_type
  limit: 50

REST Approach:

GET /api/now/table/sys_cs_message
  ?sysparm_query=conversation=<conversation_sys_id>^ORDERBYsys_created_on
  &sysparm_fields=sys_id,body,direction,sys_created_on,typed_by,message_type
  &sysparm_limit=50

For interaction-based conversations:

Tool: SN-Query-Table
Parameters:
  table_name: interaction
  query: sys_id=<interaction_sys_id>
  fields: sys_id,short_description,type,channel,state,opened_for,assigned_to,opened_at,work_notes
  limit: 1

Step 2: Analyze Initial User Message

The first user message is the strongest signal for conversation type. Extract:

Intent Signals:

| Signal Type | Examples | Classification | |-------------|----------|---------------| | Question words | "How do I...", "What is...", "Where can I..." | Inquiry | | Action verbs | "I need...", "Please set up...", "Can you create..." | Request | | Negative sentiment | "This is unacceptable...", "I'm frustrated..." | Complaint | | Suggestion language | "It would be nice if...", "Have you considered..." | Feedback | | Urgency markers | "URGENT", "This is critical", "Need help NOW" | Escalation | | Problem description | "It's not working", "I'm getting an error..." | Troubleshooting | | Reference to prior | "Following up on INC...", "As discussed..." | Follow-up |

Step 3: Apply Classification Rules

Score the conversation against each type using a weighted analysis:

=== CONVERSATION CLASSIFICATION ===
Conversation ID: CS0045678
Channel: Virtual Agent (Service Portal)
User: Jane Smith (Engineering)
Initial Message: "I've been waiting 5 days for my laptop and no one
  has contacted me. This is the third time I've had to follow up.
  I need this resolved today or I need to speak with a manager."

Classification Scores:
| Type            | Score | Signals Detected |
|-----------------|-------|------------------|
| Complaint       | 85%   | Negative sentiment, dissatisfaction, wait time mention |
| Escalation      | 75%   | "speak with a manager", urgency ("today") |
| Follow-up       | 60%   | "third time", reference to prior interaction |
| Request         | 30%   | "need this resolved" (implicit service need) |
| Inquiry         | 10%   | No information-seeking language |
| Feedback        | 5%    | No constructive suggestion |
| Troubleshooting | 5%    | No technical problem described |

PRIMARY: Complaint (85%)
SECONDARY: Escalation (75%)
COMPOSITE: Complaint with Escalation Request

URGENCY: High
SENTIMENT: Negative (frustrated)
PRIORITY ACTION: Route to human agent with escalation flag

Step 4: Detect Multi-Intent Conversations

Some conversations contain multiple intents. Identify all:

MCP Approach:

Tool: SN-NL-Search
Parameters:
  query: "laptop delivery delay complaint escalation"
  table: sys_cb_topic
  limit: 10

Analyze message-by-message for intent shifts:

Message 1: "I need to reset my password" -> Request
Message 2: "Also, the VPN has been slow all week" -> Troubleshooting
Message 3: "And when will the new laptops be available?" -> Inquiry

Classification: Multi-intent conversation
Primary: Request (password reset - most actionable)
Secondary: Troubleshooting (VPN performance)
Tertiary: Inquiry (laptop availability)

Step 5: Map Classification to Routing Rules

Determine the appropriate handling workflow based on classification.

Routing Matrix:

| Classification | Channel | Priority | Route To | |---------------|---------|----------|----------| | Inquiry | Virtual Agent | Low | Knowledge search -> FAQ topic | | Request | Virtual Agent | Medium | Catalog item topic -> Fulfillment | | Complaint | Live Agent | High | CSM queue -> Service Recovery | | Feedback | Async | Low | Feedback collection -> Survey | | Escalation | Live Agent | Critical | Manager queue -> Priority handling | | Troubleshooting | Virtual Agent | Medium | Diagnostic topic -> IT Support | | Follow-up | Live Agent | Medium | Original assignee -> Context resume |

MCP Approach:

Tool: SN-Query-Table
Parameters:
  table_name: sys_cb_topic
  query: nameLIKEcomplaint^active=true
  fields: sys_id,name,description,queue,priority,active
  limit: 5

Step 6: Enrich the Interaction Record

Update the interaction or conversation record with classification metadata.

MCP Approach:

Tool: SN-Update-Record
Parameters:
  table_name: interaction
  sys_id: <interaction_sys_id>
  data:
    u_conversation_type: "complaint"
    u_secondary_type: "escalation"
    u_urgency: "high"
    u_sentiment: "negative"
    u_classification_confidence: "85"
    u_routing_recommendation: "csm_escalation_queue"

REST Approach:

PATCH /api/now/table/interaction/<interaction_sys_id>
Body: {
  "u_conversation_type": "complaint",
  "u_secondary_type": "escalation",
  "u_urgency": "high",
  "u_sentiment": "negative"
}

Step 7: Create Downstream Records Based on Type

Automatically create appropriate records based on classification.

For Complaints -- Create CSM Case:

MCP Approach:

Tool: SN-Create-Record
Parameters:
  table_name: sn_customerservice_case
  data:
    short_description: "Complaint: Laptop delivery delay - 3rd follow-up"
    description: "<conversation summary>"
    priority: 2
    contact: "<user_sys_id>"
    category: "complaint"
    u_complaint_type: "service_delivery"
    u_source_conversation: "<conversation_sys_id>"

For Requests -- Create Catalog Request:

Tool: SN-Create-Record
Parameters:
  table_name: sc_request
  data:
    requested_for: "<user_sys_id>"
    short_description: "Password reset request via chat"
    description: "<conversation context>"
    u_source_conversation: "<conversation_sys_id>"

Step 8: Handle Escalation Routing

When an escalation is detected, trigger immediate routing.

MCP Approach:

Tool: SN-Update-Record
Parameters:
  table_name: interaction
  sys_id: <interaction_sys_id>
  data:
    state: "transferred_to_agent"
    assignment_group: "<escalation_queue_sys_id>"
    u_escalation_reason: "Customer requested manager involvement"
    u_escalation_priority: "high"
    work_notes: "AI Classification: Complaint with escalation request. Customer has followed up 3 times regarding laptop delivery delay. Sentiment: Negative/Frustrated. Routing to escalation queue."

Step 9: Track Classification Accuracy

Monitor and validate classification decisions over time.

MCP Approach:

Tool: SN-Query-Table
Parameters:
  table_name: interaction
  query: u_conversation_typeISNOTEMPTY^sys_created_on>javascript:gs.daysAgo(30)
  fields: u_conversation_type,u_classification_confidence,state,u_agent_override_type
  limit: 500

Calculate accuracy metrics:

=== CLASSIFICATION ACCURACY (Last 30 Days) ===

| Type            | Classified | Agent Override | Accuracy |
|-----------------|-----------|---------------|----------|
| Inquiry         | 245       | 12            | 95.1%    |
| Request         | 189       | 8             | 95.8%    |
| Complaint       | 67        | 5             | 92.5%    |
| Troubleshooting | 156       | 11            | 92.9%    |
| Escalation      | 34        | 3             | 91.2%    |
| Feedback        | 28        | 4             | 85.7%    |
| Follow-up       | 42        | 6             | 85.7%    |

Overall Accuracy: 93.4%
Most Common Misclassification: Feedback classified as Inquiry

Step 10: Refine Classification Rules

Use override data to improve classification accuracy.

MCP Approach:

Tool: SN-Query-Table
Parameters:
  table_name: interaction
  query: u_agent_override_typeISNOTEMPTY^u_conversation_type!=u_agent_override_type
  fields: u_conversation_type,u_agent_override_type,short_description
  limit: 50

Analyze misclassifications to identify patterns and update routing rules.

Tool Usage

| Tool | Purpose | When to Use | |------|---------|-------------| | SN-Query-Table | Fetch conversations, messages, topics, metrics | Primary data retrieval | | SN-Get-Record | Retrieve specific interaction details | Single conversation analysis | | SN-Create-Record | Create downstream records (cases, incidents) | Acting on classification | | SN-Update-Record | Enrich interaction with classification metadata | Recording classification results | | SN-NL-Search | Find matching topics or similar conversations | Topic routing and pattern matching |

Best Practices

1. Classify on first message -- do not wait for multiple exchanges to make an initial classification
2. Support reclassification -- conversations can shift type mid-stream; update classification dynamically
3. Use confidence thresholds -- below 70% confidence, route to human agent for classification
4. Combine intent and sentiment -- a request with negative sentiment may actually be a complaint
5. Handle multi-intent gracefully -- address the most urgent intent first, then secondary intents
6. Preserve conversation context -- pass full classification metadata to receiving agent or workflow
7. Track agent overrides -- use override data as training signal for improving classification rules
8. Consider channel context -- phone calls have different patterns than chat or email
9. Detect language and tone shifts -- escalation often manifests as a shift from neutral to negative
10. Respect privacy -- do not log or classify sensitive personal information in metadata fields

Troubleshooting

| Issue | Cause | Resolution | |-------|-------|------------| | Messages not retrieved | Wrong conversation table or ID format | Check sys_cs_conversation vs interaction table | | Classification always "Inquiry" | Default fallback too aggressive | Lower inquiry threshold, add more signal patterns | | Escalation not detected | Urgency language not in signal list | Add domain-specific urgency phrases to detection rules | | Multi-intent not handled | Only first intent extracted | Implement per-message analysis with intent accumulation | | Routing to wrong queue | Topic mapping outdated | Update sys_cb_topic routing configuration | | Low confidence scores | Ambiguous or very short messages | Request clarification from user before classifying |

Examples

Example 1: Virtual Agent Inquiry Classification

Input: User message: "What are the company holidays for 2026?"

Classification: Inquiry (95% confidence). Route to Knowledge search topic. Suggested KB article: "2026 Company Holiday Calendar."

Example 2: Complaint with Escalation Detection

Input: User message: "I submitted a request two weeks ago and nothing has happened. This is completely unacceptable. I need to talk to someone who can actually help."

Classification: Complaint (88%) + Escalation (80%). Route to live agent in escalation queue. Flag as high priority. Auto-search for existing open requests by this user.

Example 3: Multi-Intent Chat Session

Input: User sends 3 messages: password reset request, question about VPN policy, and feedback about the portal design.

Classification: Multi-intent detected. Primary: Request (password reset, actionable). Secondary: Inquiry (VPN policy). Tertiary: Feedback (portal). Route password reset to IT topic, queue VPN question for knowledge search, log feedback for UX team review.

Related Skills

• csm/sentiment-analysis - Deep sentiment analysis for conversations
• csm/chat-recommendation - Suggested responses for agents
• hrsd/chat-reply-recommendation - HR-specific chat response suggestions
• genai/playbook-recommendations - Match conversation to handling playbooks
• itsm/incident-triage - Incident classification and routing