AIPROFESSIONAL2024 – Present

AI TICKET TRIAGE

Claude-powered IT incident triage that automatically categorizes support tickets, generates troubleshooting notes for agents, suggests canned responses, and identifies catalog item misroutes — processing thousands of tickets per year without manual intervention.

5K+

Tickets/Year

420+

Hours Saved/Year

3

Category Levels

750+

Hrs/yr at Phase 3

HOW IT WORKS

Every new IT support ticket is automatically intercepted before an agent sees it. The ticket subject and description are extracted and passed to Claude with a structured prompt that instructs the model to perform four tasks simultaneously.

Task 1

3-Level Categorization

Classifies the ticket into a category hierarchy (e.g., Hardware → Laptop → Performance) so agents immediately know the issue domain without reading the full ticket.

Task 2

Private Troubleshooting Notes

Generates private agent-facing notes with step-by-step troubleshooting guidance relevant to the identified issue — speeding up time-to-resolution.

Task 3

Canned Response Suggestion

Suggests an appropriate canned response for the agent to send to the user — reducing drafting time for common issue types.

Task 4

Catalog Redirect Detection

Identifies when users submit incident tickets for items that should go through the service catalog (software requests, new equipment, etc.) and flags them for redirection.

MODEL SELECTION

The system started with GPT-4 via OpenAI's API. After encountering production reliability issues related to model deprecation cycles — where model behavior would shift between versions, causing inconsistent categorization output — the system was migrated to Claude via Vertex AI.

Claude was chosen specifically because Anthropic's approach to model versioning provides better stability for production systems. The ability to pin to a specific model version without unexpected behavioral changes makes Claude significantly more reliable for a zero-tolerance automation environment where consistency directly affects agent productivity.

PHASES

Phase 1 — Initial Intake

COMPLETE

Core triage pipeline: auto-categorization, private troubleshooting notes, canned response suggestions, and catalog misroute detection. Integrated with Freshservice webhooks for real-time processing of every new ticket.

  • • Webhook-triggered on ticket creation
  • • Structured prompt engineering for consistent JSON output
  • • Freshservice API integration for private notes and field updates

Phase 2 — Refinement

COMPLETE

Improved categorization accuracy through prompt iteration and added an agent feedback collection loop. Agents can flag incorrect categorizations, feeding data back for ongoing prompt improvement.

  • • Agent feedback mechanism for flagging miscategorizations
  • • Prompt iteration based on real ticket data
  • • Catalog redirect logic refined based on misroute patterns

Phase 3 — Auto-Resolution

PLANNED

Automatically resolve 20–30% of repetitive tickets without agent involvement — projected at 500–750 additional hours saved per year. Target issues: MFA resets, lost phone workflows, common access request types.

  • • MFA reset automation via Okta API
  • • Lost phone: remote wipe initiation + replacement workflow
  • • Common access issues: self-service resolution paths
  • • Human-in-the-loop escalation for ambiguous cases

TECHNOLOGIES

ClaudeVertex AIWorkatoFreshservicePythonGPT-4Prompt EngineeringREST APIs