ai-automation
Every failed AI automation we've reviewed died the same way: the model was fine, the system around it wasn't. Retries, idempotency, human review queues — the boring parts are the product.
Three rules
1. Make every step resumable. If step 4 of 6 fails at 2am, the run should pick up there, not start over.
2. Put a human gate where mistakes are expensive. Drafts can flow freely; anything customer-visible goes through a review queue.
3. Log inputs and outputs, not just errors. When output quality drifts, you need last month's inputs to prove why.
Tool shapes
Zapier and Make win for speed to first automation. n8n wins when you need self-hosting, versioned workflows, and code steps. Agents win when the task genuinely branches — and lose everywhere else.