Automating with AI sounds good until you open Make, see 2,000 integrations, and don’t know where to start. This guide is for you to set up your first useful AI automation in an afternoon — something that runs on its own, saves real time, and doesn’t break the next day.
This is not theory. It’s a practical map of the three routes in 2026: Make, n8n, and AI agents. Which one to start with according to your situation, what to avoid, and a real example that can be set up in under an hour.
The 3 Routes to Automate with AI in 2026
Make (formerly Integromat)
The fastest option if you’ve never programmed. Drag-and-drop visual, thousands of native integrations, ready-to-use templates. It works very well for flows up to 10-15 steps.
- Cost: free up to 1,000 operations/month; paid plans from 9 €/month.
- Ideal for: freelancers, marketing teams, operations with few repetitive steps.
- Limitation: from a certain complexity level, the interface becomes uncomfortable and debugging becomes complicated.
n8n
Open source, self-hosted or cloud. More technical but infinitely more flexible. It natively handles AI agents (with LangChain integrated), long workflows, and connections to custom APIs.
- Cost: 100% free self-hosted; cloud from 20 €/month.
- Ideal for: developers, automation agencies, SaaS products that need their own workflows.
- Limitation: a slightly higher learning curve. If you never want to touch JSON or JavaScript, Make is better.
AI Agents (LangChain, AutoGen, Custom)
The highest level. Not linear automation, but an agent that decides the steps on its own based on an objective. In 2026, this already works in production for specific cases (customer support, proposal generation, assisted research).
- Cost: your time + LLM tokens.
- Ideal for: tasks with many intermediate decisions where a linear workflow doesn’t work.
- Limitation: requires constant evaluation, can hallucinate actions, beware of write permissions.
How to Choose Between Make and n8n
Simple decision:
- If you’re going to set up 3-5 simple flows and don’t want to deal with servers → Make.
- If you’re going to live off automation or need complex flows → n8n.
- If you’re in doubt, start with Make. Migrating to n8n later is viable, and you’ll learn the concepts first.
For an in-depth comparison, see our Make vs n8n 2026 analysis.
Your First Useful Automation in 60 Minutes
Case: you receive many emails from customers with similar questions and want AI to classify and draft a response automatically.
Stack
- Gmail (trigger)
- Make or n8n (orchestration)
- ChatGPT API or Claude API (processing)
- Google Drive or Notion (optional storage)
The 5 Steps of the Flow
- Trigger: when an email arrives at a specific label (e.g., “support”).
- AI Processing: passes the email body to Claude or ChatGPT with a classifier prompt: is it a technical question, commercial, complaint, or other? What customer is it? What priority?
- Bifurcation: depending on the classification, different sub-flows. If it’s a technical question → generates a draft response using a knowledge base. If it’s a complaint → escalates to a human with a summary.
- Draft: the draft is created in Gmail as a response in draft state, not sent alone. This is the key: AI proposes, you approve with a click.
- Log: each interaction is recorded in Notion or a Google Sheets sheet to review weekly performance.
Prompt Template for the Classifier
Act as my support assistant. Read this email and return a JSON with:
{
"category": "technical_question | commercial | complaint | other",
"priority": "high | medium | low",
"summary": "1 sentence of 15 words max",
"suggested_response": "draft of 80-120 words, professional and close tone"
}
DO NOT add text outside the JSON. If in doubt about the category, mark "other".
Email:
"""
{{email body}}
"""This flow takes 45-60 minutes to set up from scratch, spends a few cents a day on API, and easily saves 30-90 minutes a day if you receive 10+ similar emails.
Common Mistakes When Automating with AI
- Automating without measuring. Without metrics, you won’t know if your flow is failing until it’s too late. Always log AI decisions and review them weekly.
- Letting AI send alone. For irreversible actions (sending emails, charging, publishing), AI must propose, not execute. A human approval layer prevents 95% of disasters.
- Forgetting the cost of tokens. Flows with thousands of executions a day can skyrocket the LLM bill. Start with a small model (GPT-4o mini or Claude Haiku) and only scale to premium models if the result requires it.
- Not having a plan B when it fails. APIs can fail. Make and n8n can crash. Your flow must assume failures and alert you (Telegram, email, Slack).
When to Switch to Agents
Jump to an agent only when your linear flow has too many branches and maintaining them becomes impossible. A well-defined agent with limited tools (3-5 actions, not infinite) and an automated evaluator is manageable. An agent that “does whatever you want” is unmanageable.
Typical case to switch to an agent: continuous competitive research (reads Reddit, X, Product Hunt, searches for mentions, summarizes each week). There, an agent with web search + memory of what it has already seen works better than a rigid workflow.
Frequently Asked Questions About Automating with AI
Which is better to start with: Make or n8n?
Make if you don’t want to touch anything technical and you’re going to set up a few flows. n8n if you want real flexibility, plan to grow in complexity, or want self-host (free and unlimited).
How much does it cost to automate with AI per month?
For a freelancer with 3-5 flows: 20-50 €/month between automation platform and LLM API. For an agency with active clients: 200-800 €/month depending on volume.
Is it safe to let AI respond to customers automatically?
Only if AI creates drafts that you approve. For direct sending, only in very specific flows (confirmations, verified FAQs). The risk of inadequate response or hallucination is never zero.
Can I automate tasks with AI without programming?
Yes, perfectly with Make or n8n using visual templates. The “code” part in these platforms is reduced to copying prompts and using variables. It’s accessible if you have patience for the first 3-4 flows.
How long does it take for an automation to pay off?
For flows that save 30 minutes/day in a recurring task: 1-2 weeks. For complex agency flows (5-10 clients): the first month you recover the investment.
To Get Started Today
If you’ve never automated anything, go in order:
- Read our n8n guide in Spanish step by step even if you’re going to use Make — the concepts are the same.
- See the Make vs n8n comparison and choose a platform.
- Set up the email classification flow above. If you have time, replicate it with form tickets.
