🏢 Business / Industry: E-commerce / Home goods retail

🎯 Main Goal: Automatically classify, prioritize, and route every incoming support ticket within 60 seconds — then enforce SLA compliance through automated escalation and surface operational insights through a conversational analytics agent.

🔹 Context

HouseBox, an online home goods store, handles ~900 support tickets per month across Telegram, email, and phone with a team of three managers. Every ticket lands in one shared inbox — a $10,000 payment failure from a VIP client sits between "what are your opening hours?" and advertising spam. Managers process the inbox top to bottom, meaning the highest-value tickets are discovered by accident, hours after they arrive. The gap between when a ticket arrives and when someone picks it up is where customers are lost.

🔹 Problem

The business was drowning in a queue it couldn't triage. 50% of all tickets were repetitive "where is my order?" and "when is delivery?" questions — answerable with a single tracking fact, yet consuming the same manager attention as a refund dispute. Average response time was 11 hours. Negative reviews accumulated on Google Maps. One manager quit from burnout, citing "the same questions over and over." The owner's proposed fix was hiring two more people — against a $2,500/month support budget that couldn't cover even one full-time hire. The problem was never response quality. It was that no prioritization existed between arrival and pickup.

🔹 Solution

A three-layer n8n system was built to close that gap. Layer 1 classifies every incoming ticket with an LLM using chain-of-thought reasoning across urgency, business impact, and customer context — assigning a category, priority (P1–P4), owner, and SLA deadline, then logging it to Google Sheets and alerting the responsible manager in Telegram. Layer 2 monitors SLA deadlines every 5 minutes and fires a three-tier escalation chain when a deadline is breached. Layer 3 is a conversational analytics agent, published as a chat widget, that the owner queries in plain language ("give me stats for today") — orchestrating two callable sub-workflows to answer with real data instead of guesses.

🔹 Process

Layer 1 — Ticket Intake & Classification

• Gmail Trigger (every 1m): Polls the shared inbox and initiates the pipeline for each new email thread.
• Collect Ticket Context (Set): Captures created_at at the moment of arrival and derives a stable ticket_id from the Gmail thread ID — guaranteeing that follow-up replies update the same ticket rather than spawning duplicates.
• AI Classifier (Chain of Thought): An AI Agent with a Structured Output Parser evaluates each ticket across three dimensions — urgency signals, business impact, and VIP/repeat-contact context — before assigning category, priority, owner, SLA minutes, summary, reasoning, and sentiment. Special rules force P1 on review threats and repeat unanswered contacts, and a P2 floor on VIP customers.
• Flatten Result (Set): Computes sla_deadline by adding the AI-assigned SLA minutes to the captured arrival timestamp — each priority tier gets its own deadline with no hardcoded branching.
• Spam? (Switch): Routes legitimate tickets to the log and alerts; filters out non-customer contact before it reaches a manager.