---
name: conversational-ai-chatbot
description: Deploy and manage AI-powered chatbots for 24/7 automated customer support, including intent classification, knowledge base retrieval, multi-turn dialogue, and human handoff. Use when setting up chatbots, training NLP models, configuring deflection workflows, managing chatbot analytics, designing conversational flows, or handling escalation rules. Triggers on phrases like "chatbot", "conversational AI", "AI chat", "bot deflection", "NLP training", "intent recognition", "bot handoff", "chatbot analytics", "conversational flow", "self-service chat".
---

# Conversational AI Chatbot Management

Deploy and optimize AI chatbots to deflect tickets, answer common questions, and provide 24/7 support.

## Workflow

### Chatbot Deployment Lifecycle

Trigger: New bot launch; bot retraining; performance optimization:

1. **Requirement gathering**: Identify top support ticket categories (typically billing, account, technical, onboarding); define scope and success metrics (target deflection rate 50–65%).
2. **Intent taxonomy design**: Map 20–50+ intents (e.g., reset_password, check_order, billing_question, feature_request, cancellation, refund); define example utterances per intent (50–100 examples minimum).
3. **Knowledge base integration**: Connect to existing help articles; configure retrieval algorithm (semantic search + keyword); set confidence threshold for auto-response.
4. **Conversation flow design**: Build dialogue trees for common journeys (login issues, order status, billing inquiries); define fallback responses for unrecognized intents.
5. **Human handoff configuration**: Set escalation triggers (3 failed attempts, sentiment threshold, explicit "talk to human" request); define handoff message and context transfer format.
6. **Testing and QA**: Test with 100+ sample conversations; validate intent accuracy >85%; verify handoff quality; A/B test tone variations.
7. **Soft launch**: Enable for 10–20% of traffic; monitor live conversations; track deflection rate, CSAT, escalation rate; iterate weekly.
8. **Full launch**: Expand to 100% traffic; establish weekly review cadence; monthly retraining with new conversation data.
9. **Continuous improvement**: Analyze failed intents monthly; add new intents from support tickets; optimize responses based on CSAT feedback.

### Chatbot Configuration Framework

```
CHATBOT CONFIGURATION — STANDARD TEMPLATE
===========================================

Bot Identity:
  Bot name: [SupportBot / HelpDesk AI / etc.]
  Brand voice: □ Professional □ Friendly □ Casual □ Technical
  Greeting message: "Hi! I'm [Bot Name]. How can I help you today?"
  Language support: English (primary) + [secondary languages]

Intent Classification:
  Minimum training examples per intent: 50
  Intent confidence threshold: 0.75
  Fallback when below threshold: "I'm not sure I understand. Can you rephrase?"

Knowledge Base Integration:
  Source: [Zendesk Help Center / Confluence / Custom CMS]
  Retrieval method: Semantic search (embeddings) + keyword matching
  Results per query: Top 3 articles
  Confidence threshold for auto-display: 0.80

Escalation Rules:
  Trigger 1: Explicit handoff ("talk to human", "speak to agent")
  Trigger 2: 3 consecutive failed intent matches
  Trigger 3: Negative sentiment detected (score < 0.3)
  Trigger 4: Security-related keywords (password, compromised, hack)
  Handoff message: "Let me connect you with a human agent who can help."
  Context transfer: Full conversation log + detected intent + customer data

Analytics & Monitoring:
  Deflection rate target: 60%
  Response time target: < 2 seconds
  CSAT target: > 4.0/5.0
  Escalation rate target: < 30%
  Review cadence: Weekly performance review; monthly retraining
```

### Intent Taxonomy Design

```
TOP SUPPORT INTENTS — CATALOG (EXAMPLES)
============================================

Account Management:
  - reset_password: "I forgot my password", "Can't log in"
  - update_email: "Change my email", "Update contact info"
  - delete_account: "Close my account", "Cancel and delete"
  - verify_identity: "Verify my identity", "Two-factor setup"

Billing:
  - check_invoice: "Show my last invoice", "Download invoice"
  - payment_failed: "Payment declined", "Card expired"
  - refund_request: "I want a refund", "Charge me back"
  - plan_change: "Upgrade my plan", "Downgrade subscription"
  - billing_address: "Update billing address", "Wrong address"

Technical Support:
  - login_error: "Login not working", "Error 500 on login"
  - feature_help: "How do I use feature X?", "Feature not working"
  - integration_issue: "API not connecting", "Webhook failing"
  - mobile_app: "App crashes on iPhone", "Can't sync on Android"

Product:
  - pricing_info: "How much is Enterprise?", "Compare plans"
  - feature_request: "Can you add feature X?", "Missing feature"
  - documentation: "Where's the API docs?", "User guide"
  - demo_request: "I want a demo", "Schedule a call"

Compliance:
  - data_export: "Export my data", "GDPR data request"
  - privacy_policy: "What's your privacy policy?"
  - security_inquiry: "Is my data encrypted?"
```

## Edge Cases

- **Multi-intent queries** (customer asks multiple things in one message):
  - Example: "How do I reset my password and also change my billing address?"
  - Strategy: Detect primary intent (highest confidence); resolve it first; then detect secondary intent from follow-up; or offer a menu: "I can help with both. Let's start with password reset. Then we'll handle billing."
  - Limit: Handle max 2 intents per turn; if 3+, escalate to human
  - Training: Add multi-intent examples to training data; flag during QA review

- **Low-resource languages** (insufficient training data for accurate NLP):
  - Strategy: Use translation-based fallback (translate to primary language, process, translate response back)
  - Risk: Translation may alter meaning; lower accuracy than native models
  - Minimum: 100+ training examples per intent for reliable performance
  - Alternative: Offer limited "FAQ mode" for low-resource languages (pre-built Q&A, no open-ended NLP)

- **Rogue chatbot responses** (bot provides incorrect or harmful information):
  - Guardrails: Blocklist sensitive topics; require KB-backed responses (no free-form generation); confidence thresholds
  - Monitoring: Daily review of bot conversations by QA team; automated flag for low-confidence responses
  - Response: Immediately disable problematic intent; revert to last known good model
  - Prevention: Use retrieval-augmented generation (RAG) instead of pure generative models; restrict to verified knowledge base content

- **Bot fatigue** (customer gets frustrated with bot after repeated failures):
  - Detection: Track failed attempts per session; negative sentiment trend; keywords like "stupid bot", "useless"
  - Mitigation: After 2 failures, offer human handoff proactively: "I'm having trouble understanding. Would you like to speak with a human agent?"
  - Tone shift: When frustration detected, switch to more apologetic/empathetic tone
  - Prevention: Continuous training on failure cases; expand intent coverage monthly

- **Chatbot for technical debugging** (complex technical issues requiring multi-step diagnostics):
  - Approach: Guided troubleshooting workflow (see guided-troubleshooting skill)
  - Limitations: Bot can handle known error codes and common issues; cannot diagnose novel problems
  - Escalation: Collect diagnostic data (browser, OS, logs); auto-create ticket with full context; escalate to engineer
  - Integration: Product APIs for account checks; log collection via bot-embedded tools

## Integration Points

- **Help desk platforms**: Zendesk, Freshdesk, Intercom, Front — handoff with full conversation context; ticket creation
- **NLP engines**: Dialogflow, Rasa, Microsoft LUIS, Amazon Lex — intent classification, entity extraction
- **Knowledge bases**: Zendesk Guide, Confluence, Notion — article retrieval, semantic search
- **CRM**: Salesforce, HubSpot — customer data enrichment, account tier detection
- **Analytics**: Mixpanel, Amplitude — conversation analytics, funnel tracking
- **Messaging platforms**: WhatsApp Business, Facebook Messenger, Slack, Teams — multi-channel deployment
- **Translation APIs**: Google Translate, DeepL — multi-language support
- **Auth systems**: Auth0, Okta — identity verification for sensitive queries
- **Monitoring**: Datadog, PagerDuty — bot health monitoring, alerting on failures
- **Data warehouse**: Snowflake, BigQuery — historical conversation data for training and analytics