SmartDesk started as a side project. A question I couldn't stop thinking about: could I build an AI-powered IT helpdesk triage system — one that actually works in production — without a data science team, a six-figure platform budget, or months of integration work?
The answer is yes. And the stack that made it possible is n8n + Claude API.
Why This Stack
n8n is an open-source workflow automation platform that sits comfortably between Zapier's ease and Airflow's power. It handles the orchestration: receiving tickets, calling APIs, routing outputs, sending notifications. Claude handles the intelligence: understanding ticket context, classifying intent, generating responses.
Together, they cover the full ITSM automation loop without requiring you to write a production application from scratch.
The Architecture
SmartDesk runs a five-stage pipeline:
Stage 1: Ticket Intake
An n8n webhook node receives incoming tickets from email (via IMAP), ServiceNow API webhook, or a simple web form. The node normalises the payload into a standard schema: { id, subject, body, submitter, timestamp }.
Stage 2: AI Analysis
The normalised ticket is sent to Claude API with a structured classification prompt. Claude returns a JSON object with category, priority (P1–P4), confidence score, suggested resolution team, and a 50-word agent briefing.
You are an ITIL-aligned service desk triage specialist. Classify the following IT support ticket. Return ONLY valid JSON with: category, priority (P1=Critical/P4=Low), team, confidence (0-1), briefing (max 50 words), auto_resolve (boolean). If confidence < 0.75, set auto_resolve to false. Ticket: {{ticket_body}}
Stage 3: Smart Routing
An n8n switch node reads Claude's JSON output and routes the ticket accordingly:
- High confidence + known resolution → auto-resolve path
- High confidence + needs human → route to specific L1/L2 queue with briefing
- P1/P0 → immediate Slack alert to on-call engineer + escalation path
- Low confidence → route to senior L2 with full AI analysis attached
Stage 4: Auto-Notification
Claude drafts a user-facing response. For auto-resolved tickets, this is a step-by-step resolution guide. For routed tickets, it's an acknowledgement with estimated resolution time based on historical SLA data. n8n sends it via the original channel (email or ServiceNow portal).
Stage 5: Logging & Dashboard
Every ticket, AI decision, routing outcome, and user response is logged to a Google Sheet (prototype) or ServiceNow (production). A Power BI dashboard tracks classification accuracy, auto-resolution rate, and time-to-first-response.
What I Learned
Prompt Engineering is 70% of the Work
The n8n plumbing is relatively straightforward. The intelligence lives in the prompt. I went through 14 iterations of the classification prompt before settling on the current version. Each iteration was tested against 200 historical tickets with known correct classifications.
Confidence Thresholds Matter
The 0.75 confidence threshold for auto-resolution wasn't arbitrary — it came from testing. Below that threshold, AI errors increased enough to erode user trust in the system. Human-in-the-loop for borderline cases isn't a fallback; it's a feature.
The Feedback Loop is Essential
Every agent override of an AI decision is logged with the agent's corrected classification. These are batched weekly and used to evaluate prompt performance. When accuracy drops in a specific category, the prompt for that category gets refined.
Deployment Considerations
SmartDesk is designed as a parallel layer, not a replacement for your existing ITSM platform. It sits in front of ServiceNow, reading tickets and enriching them before they reach your agents. This means zero disruption to existing workflows during rollout — and easy rollback if needed.
The full source code and n8n workflow JSON are available on request. If you're interested in deploying this in your environment, reach out via the contact section of my profile.