Visual AI workflows — n8n and Zapier

Motivation

Not every AI workflow needs Python. Marketing ops teams run Slack-to-CRM automations, support engineers route tickets automatically, and finance teams generate weekly summary reports — all without writing a line of code. Visual workflow tools like n8n and Zapier sit at the low-code end of the orchestration spectrum you explored in Lesson 15. Understanding when to reach for them — and when to write Python instead — is a practical AI engineering skill. You will encounter workflows built in these tools in the wild, and you may be asked to extend or debug them.

What visual workflow tools are

A visual workflow tool lets you build automations by connecting nodes on a canvas. Each node does one thing: call an HTTP endpoint, transform data, send a Slack message, query a database, or wait for a webhook trigger. You connect nodes by drawing arrows between them.

The AI-relevant nodes are the key addition of the last two years: every major visual tool now ships direct integrations with OpenAI, Anthropic, Google Gemini, and others. You can build a pipeline like:

New email arrives → Extract action items (GPT-4o) → Create Notion tasks → Reply to sender

without writing Python — and then hand that automation to a non-technical teammate who can read and adjust it in the GUI.

n8n — developer-friendly, self-hosted

n8n (pronounced “n-eight-n”) is open-source workflow automation software you run yourself — on your laptop, a Docker container, or a cloud VM. The self-hosted version is free; they charge for their cloud-hosted version.

Why developers choose n8n

• Self-hosted: your data never leaves your infrastructure. This matters for AI systems handling sensitive documents, customer data, or anything under a compliance policy.
• Code node: when the visual nodes are not enough, drop into a JavaScript or Python function inline. This is what separates n8n from purely no-code tools — you are never blocked by the limits of the GUI.
• 400+ integrations: Slack, HubSpot, Google Sheets, Postgres, HTTP webhooks, plus native OpenAI and Anthropic nodes.

Running n8n with Docker

docker run -it --rm \
  -p 5678:5678 \
  -v ~/.n8n:/home/node/.n8n \
  n8nio/n8n

Open http://localhost:5678. You will see the n8n editor — a blank canvas ready for nodes. The volume mount (~/.n8n) persists your workflows between runs.

Building an AI classifier workflow

This example wires a webhook trigger into an OpenAI call that classifies customer messages:

Step 1 — Webhook node

Add a Webhook node. Set:

• HTTP method: POST
• Path: classify
• Response mode: Last Node (return whatever the last node produces)

This creates an endpoint at http://localhost:5678/webhook/classify.

Step 2 — OpenAI node

Add an OpenAI node. Set:

• Model: gpt-4o-mini
• Operation: Message a Model
• System prompt: Classify this customer message as exactly one of: complaint, question, praise. Reply with only the single word.
• User message: {{ $json.body.message }} (n8n’s expression syntax pulls message from the POST body)

Connect Webhook → OpenAI.

Step 3 — Activate and test

Activate the workflow (toggle in the top-right). Then test it:

curl -X POST http://localhost:5678/webhook/classify \
  -H "Content-Type: application/json" \
  -d '{"message": "Your shipping is always late and support never replies"}'

Response:

complaint

An LLM-powered classifier running as an HTTP endpoint, with no Python. This is the pattern support engineering teams use to route tickets automatically.

The Code node

When you need logic that goes beyond what the visual nodes can express — a loop, a custom transformation, a calculation — the Code node runs JavaScript (or Python in newer n8n versions) inline:

// Code node: filter LLM results below a quality threshold
const items = $input.all();

return items.filter(item => {
  const score = parseInt(item.json.score, 10);
  return score >= 7;
});

The Code node reads from $input (the previous node’s output) and returns an array of items for the next node. This is the escape hatch that makes n8n powerful for developers: you can drop into code whenever the visual layer is not expressive enough, then pop back out into visual nodes.

Zapier — SaaS, no-code

Zapier is a cloud-hosted, no-code workflow tool. It is older, more polished, and designed for non-technical users — marketers, operations people, customer success teams. Zapier’s pricing is per task (each node execution costs a task); your data goes through Zapier’s servers.

When Zapier wins

• Your teammates are non-developers. A Zap is easy to read and edit in the Zapier UI without any programming knowledge.
• You need SaaS breadth quickly. Zapier has 5,000+ pre-built integrations: Salesforce, HubSpot, Mailchimp, Airtable, DocuSign, and hundreds more.
• You do not want to manage infrastructure. Zapier handles hosting, scaling, and uptime.

What a Zap looks like

A “Zap” is Zapier’s name for a workflow. A simple AI-powered Zap might be:

1. Trigger: New row added to a Google Sheet (a list of customer feedback)
2. Action — ChatGPT: Feed the row’s feedback text to GPT-4o with the prompt: Summarise this feedback and identify the main issue in one sentence.
3. Action — Gmail: Create a draft email to the customer success team with the AI summary

Every step is configured in Zapier’s web UI — no code, no deployment. A non-technical person on your team can read this Zap, understand what it does, and update the prompt or the email template themselves.

When to use visual tools vs Python

The decision comes down to one question: who owns this automation after it is built?

Where visual tools fit in the orchestration spectrum

In Lesson 15, you saw LangGraph, LangChain, and CrewAI — Python frameworks that give you fine-grained control over multi-agent workflows. n8n and Zapier sit at the opposite end of the same spectrum:

Most control ←─────────────────────────────────────────────→ Least control
Raw API  ·  LangGraph  ·  PydanticAI  ·  CrewAI  ·  n8n  ·  Zapier

Each step to the right: less Python, more visual, easier for non-developers, harder to customise. There is no single right answer — the right tool depends on who is building, who will maintain it, and how much custom logic the automation needs.

An AI engineer is expected to understand this full spectrum and make the right call for each situation. Knowing n8n and Zapier exist — and when to recommend them over a Python service — is as important as knowing when to use LangGraph.