When a FinTech client came to us with 500+ support emails landing daily — and a 4-person team manually reading, categorising, and drafting replies to each one — we knew automation could solve most of it. What we didn't expect was how completely it would.

After 30 days live, the AI triage system we built in n8n had reduced their support overhead by 98%. Here's exactly how we built it, what it costs to run, and the numbers that matter.

98%
Cost reduction
$0.03
Cost per email
<90s
Avg response time

The problem: 500 emails, 4 people, 12-hour SLAs

The client's support queue was growing faster than their team. Each email required someone to read it, identify what the customer needed, check the knowledge base, draft a reply, and send it. At peak volume, SLAs were slipping to 12+ hours — damaging customer trust and creating a backlog that never fully cleared.

The brief

Read every inbound support email automatically. Classify it. Draft a reply using our knowledge base. Save the draft to Gmail for a human to review and send in 30 seconds — not 30 minutes.

The architecture: a 7-node n8n workflow

The system runs entirely in n8n, triggered by Gmail, and uses GPT-4o for both classification and reply generation. Here's the flow:

  1. Gmail Trigger — polls for new unread emails in the support inbox every minute
  2. Parse & Validate — strips signatures, auto-replies, and spam using a Code node
  3. AI Classify — GPT-4o reads the email and outputs category, urgency, sentiment, and a 1-line summary
  4. DB — Fetch KB Data — pulls the 5 most relevant knowledge base entries by category from PostgreSQL
  5. Assemble Draft Prompt — builds a system + user prompt combining the email context and KB results
  6. AI Agent — Generate Draft — GPT-4o writes the full reply, following brand voice guidelines
  7. Gmail — Save Draft — pushes the draft directly to the correct Gmail thread, ready for human review

What's inside the full article

  • The complete n8n workflow JSON you can import directly
  • The exact GPT-4o system prompt we use for classification and drafting
  • PostgreSQL schema for the knowledge base and email log tables
  • How we handle edge cases: auto-replies, multi-language emails, sensitive content
  • Month 1 vs Month 3 performance comparison after the self-learning loop was added
  • Cost breakdown: what it actually costs per 1,000 emails processed
✍️
Full article coming soon
We're writing up the complete technical breakdown with workflow JSON, prompts, and database schema. Subscribe below to get it the moment it drops.