TL;DR - Plain automation moves data. AI automation adds a judgment step in the middle - classify, summarize, draft, extract - using an AI node. That's what makes a workflow feel intelligent.
Why it matters
Most real value isn't moving data from A to B - it's understanding the data and deciding what happens next. The AI node is where that judgment lives.
Worked example - the upgrade
Trigger: New customer feedback (form/email)
AI node: "Classify sentiment (positive/negative) and topic (billing/bug/feature)."
Branch: negative+bug -> alert engineering; feature -> add to roadmap sheet
Action: post a one-line summary to the right Slack channelNo human reads every message first - the AI routes it.
Steal this - good AI-node jobs
Classify: "Tag this as urgent / normal / spam."
Extract: "Pull name, email, and company as JSON."
Summarize: "One-line summary of this message."
Draft: "Write a first-draft reply for a human to approve."Common mistakes (and the fix)
- Letting AI act unchecked on risky steps. Fix: draft for human approval (next lesson).
- Vague AI instructions. Fix: constrain to allowed values ("one of: urgent/normal/spam").
- No validation of AI output. Fix: check the field exists/makes sense before the next step.
Good to know
Every major tool now has this: n8n AI nodes, Make AI modules, Zapier AI steps/Agents - all can call ChatGPT/Claude/Gemini mid-workflow. Use a cheap, fast model for simple classification; save bigger models for harder reasoning.