--- name: ticket-categorization description: "Automatically categorize incoming support tickets by issue type, product area, and department for proper routing. Use when configuring NLP-based ticket classification, setting up multi-label categorization, defining custom categories, training classification models, or monitoring categorization accuracy. Triggers on phrases like 'ticket categorization', 'auto-classify tickets', 'issue classification', 'ticket tagging', 'NLP ticket routing', 'category model training', 'multi-label classification', 'issue type detection'." --- # Ticket Categorization & Classification Automatically classify and tag incoming support tickets using NLP and machine learning to ensure accurate routing and consistent data quality. ## Workflow ### Phase 1: Category Taxonomy Design 1. **Define primary categories** aligned with support teams: - Technical issues (bug, error, performance, integration) - Account & billing (payment, subscription, refund, upgrade) - Product questions (how-to, feature inquiry, documentation) - Feature requests & feedback - Security & privacy concerns - General inquiries 2. **Define sub-categories** for granular routing: - Map each primary category to specific product areas (API, UI, mobile app, admin panel) - Define severity indicators within categories (data loss > feature broken > cosmetic issue) - Create custom tags for cross-cutting concerns (compliance, VIP, escalation) 3. **Establish routing rules**: - Category → team queue mapping - Priority-based overrides (security issues always escalate) - Language-based routing for multilingual teams ### Phase 2: Model Training & Configuration 1. **Gather training data**: - Export 1,000+ historically categorized tickets from ticketing system - Annotate manually if historical data is insufficient - Balance classes to avoid bias toward high-volume categories 2. **Configure NLP pipeline**: - Text preprocessing: normalization, stop-word removal, stem/lemma - Feature extraction: TF-IDF, word embeddings (BERT, sentence-transformers) - Multi-label classification model training - Confidence threshold calibration per category 3. **Integration setup**: - Webhook or API connection to ticketing system (Zendesk, Freshdesk, Intercom) - Real-time inference pipeline (target: <2 seconds per ticket) - Fallback to manual categorization when confidence < threshold ### Phase 3: Live Classification & Continuous Improvement 1. **Real-time categorization**: - Ingest ticket content: subject line, body, attachments, customer tier, channel - Run classification model → assign primary category + sub-category + custom tags - Store confidence score alongside categorization - Route to appropriate queue based on category mapping 2. **Human-in-the-loop refinement**: - Flag low-confidence predictions for agent review - Capture agent recategorizations as training signals - Weekly model retraining with new labeled data 3. **Quality monitoring**: - Track categorization accuracy by category (target: >90%) - Monitor drift: changes in category distribution over time - Alert on new issue types not matching existing taxonomy ## Templates ### Category Taxonomy Configuration ``` TICKET CATEGORY TAXONOMY ========================= Version: [2.0] | Last Updated: [Date] PRIMARY CATEGORY | SUB-CATEGORIES | ROUTE TO | CONFIDENCE THRESHOLD --------------------------|-----------------------------|---------------|------------------- Technical Issue | Bug/Error | Engineering | 85% | Performance | Engineering | 85% | Integration/API | DevOps | 80% | Data Sync/Import | Data Team | 80% Account & Billing | Payment Failed | Billing | 90% | Subscription Change | Billing | 90% | Refund Request | Billing | 95% | Invoice Question | Finance | 85% Product Question | How-To / Usage | Support L1 | 80% | Feature Inquiry | Product | 75% | Documentation Request | Content Team | 80% Feature Request | Enhancement | Product | 80% | New Capability | Product | 80% Security & Privacy | Data Breach Concern | Security | 95% | Access/Permission | IAM Team | 85% | GDPR/Privacy Request | Legal/Privacy | 90% General Inquiry | Company Info | Support L1 | 70% | Careers | HR | 90% | Press/Media | PR | 95% ROUTING RULES: - Any Security category → immediate escalation to Security Lead - VIP customer + any category → senior agent queue - Confidence < threshold → manual review queue - Attachment detected (screenshot, log) → auto-tag "requires investigation" ``` ### Classification Performance Dashboard ``` CATEGORIZATION PERFORMANCE — Weekly Report ========================================== Reporting Period: [Date Range] OVERALL METRICS: Total tickets processed: 4,823 Auto-categorized: 4,412 (91.5%) Manual review required: 411 (8.5%) Average inference time: 1.2 seconds Accuracy (validated sample): 92.8% ACCURACY BY CATEGORY: Technical Issue: 94.2% [████████████████░░] 312/331 Account & Billing: 96.1% [██████████████████] 289/301 Product Question: 89.7% [███████████████░░░] 267/298 Feature Request: 91.3% [████████████████░░] 198/217 Security & Privacy: 97.8% [██████████████████] 45/46 General Inquiry: 87.4% [███████████████░░░] 156/179 TOP MISCATEGORIZATIONS: 1. "API rate limiting" → Product Question (should be Technical/Integration) — 23 cases 2. "Cancel subscription" → Account Question (should be Billing/Refund) — 18 cases 3. "Two-factor setup" → Technical Issue (should be Security/Access) — 14 cases ACTION ITEMS: [ ] Add API-related keywords to Technical Issue training set [ ] Adjust confidence threshold for Product Question (lowering from 80% → 75%) [ ] Retrain model with 50 new labeled examples for miscategorized types ``` ## Integration Points - **Ticketing systems**: Zendesk, Freshdesk, Intercom, Jira Service Desk, Help Scout, Front - **ML platforms**: AWS SageMaker, Google Vertex AI, Azure ML, Hugging Face - **NLP frameworks**: spaCy, NLTK, transformers (BERT, RoBERTa) - **Analytics**: Tableau, Power BI, Looker (categorization dashboards) - **Monitoring**: Datadog, New Relic (model drift, inference latency) - **CRM**: Salesforce, HubSpot (customer tier, history enrichment) - **Logging**: ELK Stack, Splunk (classification audit trail) ## Edge Cases | Scenario | Handling | |----------|----------| | Ticket spans multiple categories | Multi-label classification; route to primary, tag with secondary | | New issue type not in taxonomy | Route to "unclassified" queue; flag taxonomy owner for review | | Low confidence on all categories | Default to manual review queue; log for model improvement | | Multilingual tickets | Language detection → translate → classify → route to appropriate language queue | | Spam/phishing tickets | Separate spam filter before classification; auto-archive with zero routing | | Bulk ticket submission (100+) | Batch processing mode with rate limiting; stagger inference requests | | Attachment-only tickets (no text) | OCR on screenshots; file-type based routing; flag for manual review | ## Output ### Live Categorization Result ``` TICKET #48291 — CATEGORIZATION RESULT ===================================== Subject: "Can't export report to PDF, getting 500 error" Customer: Enterprise tier | Channel: Web form | Language: English CLASSIFICATION: Primary category: Technical Issue — Bug/Error [confidence: 94.7%] Sub-category: Performance / Export Function Custom tags: [requires-investigation, export-module, enterprise-customer] Severity indicator: Medium (functional impairment, data intact) ROUTING: Assigned queue: Engineering — Backend Team SLA: Response within 4 hours (Enterprise tier) Escalation trigger: If unresolved > 24 hours → Engineering Manager PROCESSING TIME: 0.8 seconds ```