Your client just asked: "Can you build us an AI system that handles our onboarding emails, schedules follow-ups, and summarizes support tickets?"
You said yes. Now you're staring at n8n's canvas, wondering where to start.
—
n8n is powerful. Claude is smart. But connecting them for a real business process — not a demo — requires architecture, not just drag-and-drop.
Here's what breaks in production:
After building 50+ workflows for agencies, here's the pattern that holds:
Don't build A → B → C. Build states: intake → validate → process → review → deliver → confirm. Each state has entry conditions, exit criteria, and error handling. If Claude fails at "process," the workflow pauses at "review" for human intervention — it doesn't crash.
Don't pass raw data to Claude. Pass structured context:
Role: You are a client onboarding specialist.
Context: This is a new lead from [source]. They booked [service] for [date].
Task: Draft a personalized welcome email that confirms the booking and asks one clarifying question.
Constraints: Max 150 words. Friendly but professional. Mention their specific service by name.
Every Claude call gets 3 retries with exponential backoff. If all 3 fail, the workflow logs the error, notifies a human, and continues with other branches. Nothing stops completely.
Store prompts in a database, not in the workflow. Tag versions. A/B test. When Claude's behavior shifts (it happens), you update one row — not 47 workflows.
A real client workflow we built:
Build time: 4 days. Client result: 23% reply rate (up from 4% with templates).
The agency owner positions this as their own "AI onboarding system." Their client never sees n8n or Claude. They see a branded dashboard and professional results.
That's the model: we build the engine, you own the client relationship.