A one-person business fails at the operations layer, not the product layer. Sales you can do. Support you can do. Doing both while also building is what breaks people. Automation is how you buy back the hours.
Start from the drain, not the tool
List everything you did last week that repeated. Rank by minutes lost. Automate the top item only. Building a beautiful workflow for a task you do twice a year is procrastination with a nicer interface.
The four-layer stack
Layer one — triggers. n8n watches inboxes, forms, payment events and calendars. Everything downstream starts here.
Layer two — reasoning. An LLM step classifies, summarises or drafts. Keep this narrow: one decision per node, structured output every time.
Layer three — knowledge. Dify or a small RAG service holds your docs, pricing and policies so responses stay consistent with what you actually sell.
Layer four — action. Write to the CRM, send the invoice, post the update, open the ticket. Real side effects, guarded.
event -> classify -> retrieve context -> draft -> human approve -> act
Keep the human gate longer than feels necessary
Every automation that touches money or a customer's inbox gets an approval step for its first month. You are not slowing yourself down; you are collecting the failure cases that tell you where the guardrails go.
Cost discipline
- Cap tokens per run and alert on the cap
- Cache anything deterministic
- Use the small model until quality actually fails
- Log every run with its cost so you can kill unprofitable workflows
What not to automate
Anything where being wrong is expensive and being fast is worthless: refunds, legal language, hiring decisions, apologies. Automate the retrieval and drafting around them, then press send yourself.
The compounding part
Each workflow you retire from your own hands is permanent capacity. Six of them is roughly a part-time hire — except it works weekends and never asks for a handover document.