Make (Integromat) and n8n let non-developers wire APIs, spreadsheets, and LLM steps into reliable automations. This lesson teaches triggers, structured outputs, error handling, and one complete scenario you can adapt.
What you will learn
- Map trigger → AI step → action architectures
- Force JSON-shaped LLM outputs for downstream nodes
- Add retries and failure notifications
- Estimate cost per run (tokens + platform ops)
Prerequisites
- Accounts on Make or n8n (self-host n8n if you need data residency)
- API keys stored as platform secrets — never in plain text modules
- Lesson 1 prompt skills
Step 1: Choose Make vs n8n
| Factor | Make | n8n |
|---|---|---|
| Learning curve | Gentle UI | Steeper, more flexible |
| Self-host | Cloud-first | Strong self-host story |
| Pricing | Operations-based | Fair for high volume self-host |
Start with whichever your team already pays for.
Step 2: Reference scenario — form → summary → Slack
- Trigger: Google Form or Typeform submission
- AI module: Summarize answers; extract
{ urgency, topic, next_step }as JSON - Router: If
urgency=high, ping#support-urgentSlack channel - CRM: Create HubSpot/Pipedrive note with summary
- Log: Append row to Google Sheet for audit
Step 3: Structured output prompt
Read the form JSON below. Return ONLY valid JSON:
{"urgency":"low|medium|high","topic":"string","next_step":"string","customer_quote":"string"}
No markdown fences.
---
{{form_fields}}
Parse JSON in next node; on parse error, route to human review queue.
Step 4: Reliability patterns
- Idempotency: store processed submission IDs
- Rate limits: queue when OpenAI throttles
- Dead letter: email admin when scenario fails twice
- Versioning: export scenario JSON to Git monthly
Templates: AI templates & workflows.
Step 5: Security
- Least-privilege OAuth scopes
- Rotate keys quarterly
- Do not pass PCI/PHI through consumer automations without review
Common mistakes
- Chaining 5 LLM calls when one structured call suffices
- No failure alerts (silent data loss)
- Editing production scenario without duplicate test scenario
Practice exercise
Build the form → Slack scenario on free tier with fake data. Document:
- Trigger name and filter conditions
- Exact JSON schema your prompt enforces
- What happens when the LLM returns invalid JSON
Run 10 test submissions; aim for zero silent failures.
FAQ
vs Zapier?
Same category — pick one platform; concepts transfer.
Key takeaway
Automation value is in structured handoffs and monitoring — the LLM step is only as good as the JSON contract you enforce.
Related on AIFree.vn
AIFree.vn — practical AI & IT education. Updated June 2026.