You don't need to learn Python. You don't need a developer on retainer. You need a free afternoon, a Claude API key, and the willingness to drag boxes around on a canvas. That's it.
I'm writing this for the business owner who keeps hearing "AI agents" and feels behind. You're not behind. You're 47 minutes away from owning your first one.
The agent we're building today: a lead-intake agent. Someone fills out a form on your site. The agent reads the submission, looks up the person on LinkedIn (or wherever), decides if they're qualified, drafts a personalized reply, and books a call on your calendar if they pass. All while you're at lunch.
Why n8n and not some "no-code AI platform"
There are a hundred tools promising magical AI agents in 5 minutes. Most of them are wrappers around OpenAI with a pretty UI and a $97/month subscription. n8n is different. It's an open-source workflow tool with a real AI agent node, and you can run it on a $5/month server or use their cloud version if you want zero infrastructure headaches.
I like n8n because when something breaks at 11pm on a Tuesday, I can see exactly which step failed and fix it. The drag-drop tools that hide everything? Good luck debugging those.
Two paths to start: - n8n Cloud. Sign up at n8n.io, get a hosted instance in about 90 seconds. There's a free trial and paid tiers above that. Check their pricing page for current numbers, it shifts. - Self-hosted. Docker container on a $6/month Hetzner box, or one click on Railway. If you have a technical friend, ask them. If not, start with cloud.
You can always migrate later. Don't let infrastructure paralysis stop you from building the actual thing.
The 6 pieces of your lead-intake agent
Here's the workflow we're building. Every box on the canvas is a node. You connect them with lines. That's the whole "coding."
- Webhook node … receives the form submission
- Set node … cleans up the data (name, email, phone, message)
- AI Agent node … this is where Claude does the thinking
- Cal.com or Calendly node … books the meeting
- Gmail node … sends a personalized reply
- Notion or Airtable node … logs everything for later
That's the spine. You can add fancier stuff later (Slack notifications, CRM updates, lead scoring). Start with the spine.
Step 1: The webhook
In n8n, click the orange "+" to add a node. Search "Webhook." Drop it on the canvas. n8n gives you a unique URL.
Take that URL and paste it into the form tool you use. Tally, Typeform, Jotform, even a basic HTML form on your site… all of them can POST to a webhook. Now when someone submits, n8n catches it instantly.
(Quick aside, the first time I saw a webhook fire in real time, I felt like a wizard. It's the closest non-developers get to feeling like they wrote actual software.)
Step 2: Clean the data with a Set node
The webhook gives you raw form data. Usually messy. Names with weird capitalization, phone numbers in 8 different formats, emails with trailing spaces.
The Set node lets you map fields to clean names. So `fields.email` becomes `email`. `fields.full_name` becomes `name`. You don't write code. You click dropdowns and pick the field you want.
Takes about 4 minutes.
Step 3: The AI Agent node, where the magic happens
This is the node that's going to do the actual thinking. Drag in the "AI Agent" node (n8n's Advanced AI section).
You'll plug three things into it: - A model. Pick Claude (Anthropic) or OpenAI. I prefer Claude Sonnet for this kind of task because the responses sound more like a person and less like a chatbot. You'll need an API key from console.anthropic.com or platform.openai.com. - A system prompt. This is where you tell the agent who it is and what to do. - Tools (optional). The agent can call other nodes mid-thought. Cool, but skip this for v1.
My system prompt looks roughly like this:
You are the lead-intake assistant for [Your Business Name]. You receive form submissions from people interested in working with us. Your job is to: 1. Read the submission and decide if they look like a real prospect (ignore obvious spam, students doing research, agencies trying to sell us SEO). 2. If real, write a 4-sentence personalized reply that references something specific from their submission. 3. Output a JSON object with: qualified (boolean), reply_text (string), suggested_call_length (15 or 30). Tone: warm, direct, no corporate fluff. Sound like a real person who's read what they wrote.
That's it. Save it. Test it with a fake submission. You'll see the agent's response in the right panel.
Step 4: Calendar booking
Add a Cal.com node. (Calendly works too, the n8n Calendly node has limited write actions last I checked, so Cal.com tends to be smoother for booking.) Configure it with your API key. Tell it to create a hold or send a booking link based on the `suggested_call_length` the agent returned.
If your agent returns `qualified: false`, skip this step. Use an IF node to branch.
Step 5: Send the reply
Gmail node. Connect your account via OAuth (one click). Tell it to send `reply_text` from the AI agent to the email from the form. Set the from address, the subject, done.
Pro tip: cc yourself on the first 20 emails the agent sends. Watch what it writes. Tweak the system prompt until you'd hit send yourself without editing.
Step 6: Log it
Airtable or Notion node. Append a row with the submission, the agent's decision, the timestamp. This is your audit trail. When the agent screws up (and it will, occasionally), you'll want this.
What it actually costs to run
Honest math, monthly: - n8n Cloud starter tier or a $6 self-host server - Anthropic API: pennies per submission on Claude Haiku, a few cents on Sonnet. For a business getting under 500 leads/month, you'll spend under $20. - Cal.com or Calendly: whatever you already pay - Gmail: free if you have Google Workspace
So under $50/month for an agent that handles intake while you sleep. Compare that to a VA at $1,200/month who only works business hours and needs training.
Three things that will go wrong (and they're fine)
- The first 5 emails will be slightly off. Tweak the prompt. Each tweak takes 60 seconds.
- You'll have a junk submission slip through as "qualified." Add a rule in the prompt: "If the message is shorter than 12 words OR doesn't mention what they need help with, mark qualified=false."
- The agent will book a call for someone who clearly didn't want one yet. Add a step: agent has to explicitly include `wants_call: true` before Cal.com runs.
You iterate. You're building a system, not casting a spell.
You don't need permission to start
The agents you'll build next year start with the one you build this weekend. Stop saving the n8n tutorial for later. Open n8n.io, hit start free, and build the spine I just described.
If you want the full version of this system, including the prompt templates I use across 6 agents in my own business, plus the qualification scoring rubric, that's all inside Level Up. Krista walks you through it on real screens, not theory.