AI Automation with n8n: A Founder's Practical Guide
· 3 min de leitura · Andrés Morales
Most founders I work with in Bogotá think "AI automation" means hiring a senior engineer to build a custom backend. It doesn't. A small business can replace 15–25 hours of weekly manual work in under three weeks using n8n, a few well-chosen AI agents, and a webhook or two.
This is the system I ship for clients who need to move fast — without an engineering team.
Where the time actually leaks
In every discovery call I run, the answer is one of three things:
- Lead routing — a salesperson copy-pasting contact forms into HubSpot, then chasing the wrong lead.
- Reporting — a manager pulling numbers from 4 dashboards every Monday morning.
- Customer support triage — the same five questions, answered manually, 50 times a day.
All three are solvable without code. Let's walk through one.
The 3-week plan
Week 1 — Pick one workflow, define it carefully
Don't automate 5 things. Pick the one that hurts most. For most of my clients, it's lead routing.
Write the workflow down, step by step, in plain English:
When a new contact form arrives, enrich it with Clearbit, score it against my ICP, and either (a) push to HubSpot with the right owner assigned, or (b) drop it into a Slack channel for human review.
That's your spec. If you can't write it in plain English, you can't automate it.
Week 2 — Build it in n8n
n8n is the workflow tool I recommend over Zapier or Make for three reasons:
- Self-hostable — runs on a $6 VPS, no per-task pricing.
- Native AI nodes — OpenAI, Anthropic, and Ollama all work out of the box.
- Code nodes — when the visual builder isn't enough, drop down to JavaScript.
The first build is the slowest. Expect 2–3 days from "blank canvas" to "first successful test run."
Week 3 — Watch it break
Every workflow I ship has a 2-week observation window. Things that always break:
- APIs that rate-limit without warning
- Forms that send unexpected payloads (Spanish accented characters, old browser quirks)
- AI agents that hallucinate fields on edge cases
Build observability in from day one: log every run, alert on failures, review the misses weekly.
What this actually costs
| Item | Cost |
|---|---|
| n8n (self-hosted on Hetzner) | €3.79/month |
| OpenAI API (gpt-4o-mini for triage) | ~$20/month at 200 runs |
| Clearbit enrichment | Free tier covers most SMBs |
| My hourly rate for setup | ~$70/hour, 15 hours typical |
The breakeven vs. a virtual assistant is around month 4.
When AI is overkill
A workflow that runs 10 times a day with structured data does not need an LLM. Use deterministic logic (if/else, regex, lookups) for everything except:
- Open-ended text classification
- Sentiment-based routing
- Free-form summarization
If your "AI" piece is actually a glorified if-statement, you're paying $0.005 per run for nothing.
What to build next
Once the first workflow is stable, the second is 50% faster, and the third is 70% faster, because the patterns repeat. Within 90 days, most of my clients have 4–6 stable automations running. That's when the compounding effect kicks in: ops costs don't scale with revenue anymore.
Want an audit of your current ops to see which workflows are worth automating first? The first 30 minutes are free — book a strategy call.