Library
00/07 · ~30 min
GUIDEDECK · wiring apps and LLMs together without (much) code

AI Workflow
Automation
from no-code to real code.

A 30-minute tour of the visual builders that glue your tools together — Zapier, Make and n8n for classic automation, and Flowise, Langflow and Dify for AI-native flows. We'll build one end to end, and be honest about the day you should put the canvas down and write code.

~30 MINBEGINNER → INTERMEDIATENO-CODE → CODE
SCROLL
01 · What workflow automation is 4 min

A trigger fires, nodes run,
an action finishes the job.

Almost every automation tool — old or AI-flavored — is the same three ideas on a canvas. Something happens, you do a few steps to the data, and you make something happen somewhere else. Learn this vocabulary once and every tool in the deck reads the same.

Workflow automation a sequence of steps that runs on its own when an event occurs, moving and transforming data between apps without a human clicking through each step. You assemble it visually as a graph of nodes, instead of writing and deploying a script.
TRIGGER new email Filter is urgent? Transform map fields ACTION Slack one run = trigger → nodes → action

Every builder in this deck is a variation on this line: a trigger, a few nodes, an action.

Trigger

What starts the run

The event the workflow waits for. Three common shapes: webhook (an app pushes an HTTP request to you), schedule (cron — every hour, every morning), and polling (the tool checks an app on a timer for new rows or emails).

Nodes

The steps in between

Each box does one thing: call an API, filter, branch on a condition, loop over a list, reshape JSON, or — in AI tools — call a model. Data flows down the wire from one node to the next.

Action

The effect on the world

The node that changes something outside the workflow: send a message, write a row, create a ticket, charge a card. A workflow can end in several actions, or fan out to many.

Like a factory conveyor belt — a part drops on (trigger), stations shape it (nodes), and it ships at the end (action). You design the belt once; it runs itself.

02 · The classic automators — Zapier vs Make vs n8n 5 min

Same job, three very
different shapes.

These three connect SaaS apps to each other. They differ on one axis that decides almost everything else: how much logic the canvas lets you express, and whether you can run it on your own servers. Pick the simplest one that fits — power you don't use is just complexity you pay for.

iPaaS Integration Platform as a Service: a hosted product whose whole job is connecting other apps with prebuilt connectors and a visual editor. Zapier and Make are pure iPaaS (SaaS only). n8n is the outlier: you can self-host it or use its cloud.
ZAPIER · linear Zap trigger step action MAKE / n8n · graph trigger router path A path B loop

Zapier favors one tidy line per Zap; Make and n8n give you a branching canvas with routers and loops.

Read the difference

  • Zapier— the largest connector catalog (thousands of apps) and the gentlest on-ramp. Optimized for a non-technical person automating "when X, do Y."
  • Make — a true visual canvas with rich routers, iterators and data mapping. More power per dollar at volume, a steeper first hour.
  • n8n — node-based like Make, but source-available and self-hostable, with a code node when the boxes run out. The technical-team favorite.
  • All three now ship AI nodes(call a model inside a step) — so "classic" vs "AI-native" is a spectrum, not a wall.
Zapier

Pro— biggest app library, fastest for simple "if this, then that" glue.

Con — priced per task; branching and real logic feel bolted on. SaaS only.

Choose when non-engineers own the automations and breadth of connectors matters most.

Make

Pro — powerful visual logic, cheaper operations, great data mapping.

Con — steeper learning curve; complex scenarios get visually busy. SaaS only.

Choose when you need branching, loops and transforms beyond what Zapier does cleanly.

n8n

Pro — self-hostable, code node, fair-code license, strong AI/LangChain nodes.

Con — you run and maintain it (self-host); fewer prebuilt connectors than Zapier.

Choose when a technical team wants control, data residency, and an escape hatch into code.

03 · n8n in focus — node-based and self-hostable 4 min

The bridge between
no-code and code.

n8n earns its own section because it sits exactly where most engineering teams land: a visual builder you can run yourself, with a real code node for the 10% the boxes can't express. It's where the classic automators and the AI-native builders overlap.

Self-hosting running the software on infrastructure you control instead of a vendor's cloud. With n8n that means your data and credentials never leave your network, cost is flat (your server, not per-task), and you can reach internal systems directly. The trade: you patch, back up, and scale it. n8n is fair-code — source-available under its Sustainable Use License, free to self-host, not OSI open-source.
// n8n Code node — when no prebuilt box fits for (const item of $input.all()) { const t = item.json.text.trim() item.json.words = t.split(/\s+/).length item.json.tag = t.length > 280 ? "long" : "short" } return $input.all() // hand it to the next node
Webhook trigger HTTP prebuilt Code JS / Python SAME WORKFLOW RUNS ON: self-host (Docker) n8n Cloud

Mix prebuilt nodes with a Code node; run the very same workflow self-hosted or on n8n Cloud.

Where n8n shines
  • Data must stay in-house — self-host keeps credentials and payloads on your network.
  • Volume is high — a flat server cost beats per-task pricing at scale.
  • You need an escape hatch — drop into the Code node instead of contorting boxes.
  • AI flows — its LangChain-based AI Agent and model nodes build RAG and tool-using agents on the canvas.
Where it bites
  • Self-hosting is real ops: upgrades, backups, queue mode for scale, and uptime are now yours.
  • Fewer turnkey connectors than Zapier — niche SaaS may need an HTTP node and manual auth.
  • A sprawling workflow is still hard to test and review — the no-code ceiling (Part 6) applies here too.
04 · AI-native builders — Flowise, Langflow, Dify 5 min

Visual canvases built for
RAG and agents.

Classic automators connect apps and calla model as one step. AI-native builders are the reverse: the model is the center, and the canvas wires up prompts, memory, knowledge bases and tools around it. They're visual front-ends over LLM frameworks like LangChain.

AI-native builder a visual tool whose nodes are LLM building blocks — models, prompts, retrievers, vector stores, memory and tools — for assembling chatbots, RAG pipelines and agents. The knowledge-base path they wrap is RAG & Vector Search; the apps they help you ship are Building LLM Apps.
docs embed vectorstore question retrieve LLM answer

The shape these tools make easy: embed your docs, retrieve the relevant chunks for a question, and let the model answer grounded in them.

What the nodes do here

  • Loader / splitter — pull in PDFs, sites or tables and chop them into chunks.
  • Embedder + vector store — turn chunks into vectors and index them for similarity search.
  • Retriever— fetch the chunks closest to the user's question and inject them as context.
  • Model + memory + tools — the LLM answers; memory keeps the thread; tools let an agent act.

That last row — tools an agent can call — is the same idea taught in AI Agents & Tool Use, here drawn as boxes.

Flowise

Pro — open-source, fast drag-drop chatbots and agents on LangChain/LangGraph.

Con — focused on LLM orchestration, not general app integration.

Choose to stand up a RAG chatbot or agent quickly and self-host it.

Langflow

Pro — open-source, Python-native, great for prototyping flows and exporting to code.

Con — prototyping-first; productionizing and scaling is on you.

Choose when your team is Python-first and wants to graduate flows into a codebase.

Dify

Pro — full LLMOps platform: RAG, agents, prompt IDE, datasets and observability in one.

Con — heavier and more opinionated than a simple chain builder.

Choose when you want to operate AI apps end to end, not just wire a single chain.

05 · A worked example — webhook → LLM → action 5 min

One automation,
end to end.

Let's build a support-ticket triager. A form submits a ticket; the workflow reads it, finds relevant help-docs, asks a model to classify and draft a reply, then routes it. This is the canonical AI automation — and it's the same five ideas from Part 1.

Webhook a URL the workflow exposes so another app can push it an HTTP request the moment something happens. It's the most common AI-automation trigger: instant, no polling, and the request body carries the data you act on.
// support-ticket triage — reads like the canvas on webhook("/ticket") { // 1 · trigger ctx = vectorStore.search(body.text) // 2 · retrieve (RAG) out = llm.chat({ // 3 · classify + draft system: "Triage ticket. Reply as JSON.", user: body.text, context: ctx, }) if (out.urgency === "high") // 4 · branch slack.post("#oncall", out.summary) // 5 · action zendesk.draft(out.reply) // 5 · action (human approves) }
webhook /ticket retrieve docs LLM triage urgency? high → Slack #oncall draft replyhuman approves

Trigger → retrieve → model → branch → actions. Note the human-in-the-loop: the model drafts, a person sends.

Why RAG here

Without your help-docs as context, the model invents answers. Retrieval grounds the reply in real policy — see RAG & Vector Search.

Structured output

Asking for JSON (urgency, summary, reply) lets the next nodes branch and route reliably instead of parsing prose.

Keep a human

The model drafts; a person approves the customer-facing send. High-stakes actions stay behind a review step until trust is earned.

Like a sharp intern: it reads every ticket, looks things up, writes a tidy draft and flags the urgent ones — but you sign off before anything reaches the customer.

06 · Limits & when to graduate to real code 4 min

No-code has a ceiling.
Know when you hit it.

Visual builders are the fastest way to a working prototype and a huge range of real production glue. But complexity that lives in a canvas gets harder to test, review and trust as it grows. The skill is spotting the moment the canvas costs more than it saves.

The no-code ceiling the point where the logic you need is easier to express, version and test in code than to wrestle into nodes. You don't abandon the tools — you keep them for the simple flows and move the gnarly, critical, fast-changing parts into a real codebase.
complexity → the ceiling no-code wins code wins glue RAG bot multi-agent

Simple glue and a single RAG bot live happily in no-code. A custom, multi-step agent system pushes past the ceiling.

Signs you've hit it

  • You can't diff or review it. A canvas resists pull requests, code review and meaningful version history.
  • You can't test it. No unit tests, hard to reproduce runs, debugging is click-and-pray.
  • Logic is sprawling. Dozens of nodes, deep branching, or the same Code node copy-pasted everywhere.
  • Cost or latency hurts. Per-task pricing or per-node overhead beats a tight service at your volume.
  • The agent got real. Custom tools, multi-step reasoning, and evaluation now need a framework.
Graduate to

An LLM app

Real prompt management, streaming, tests and deploys. Start at Building LLM Apps.

Graduate to

A coded agent

Custom tools, control flow and evals beyond drag-drop — see AI Agents & Tool Use.

Graduate to

Owned retrieval

Tuned chunking, hybrid search and re-ranking you control — see RAG & Vector Search.

The healthy pattern is both: prototype on the canvas, keep it for the long tail of simple automations, and move only the hard, high-value core into code — often exporting the flow you already proved.

07 · Tooling, SaaS vs self-host & recap 3 min

Pick the simplest tool
that fits the job.

There is no "best" builder — only the cheapest one that meets your real constraints on logic, data control and team skill. Here's the whole landscape in one place, then the decisions that actually matter.

Tooling landscape

Connect SaaS apps with prebuilt connectors. Pick on logic depth and whether you must self-host.

Zapier

Pro — largest connector catalog, easiest for non-engineers.

Con — per-task pricing; weak at complex logic; SaaS only.

Choose for broad, simple glue owned by non-technical people.

Make

Pro — powerful visual logic, cheap operations, rich data mapping.

Con — steeper curve; busy at scale; SaaS only.

Choose for branching and transforms beyond Zapier.

n8n

Pro — self-hostable, code node, strong AI nodes, fair-code.

Con — you run it; fewer turnkey connectors.

Choose for technical teams wanting control and a code escape hatch.

Build chatbots, RAG and agents visually over an LLM framework. Pick on scope: a chain, a prototype, or a whole platform.

Flowise

Pro — open-source, quick RAG bots and agents on LangChain.

Con — LLM-orchestration focus, not general integration.

Choose to ship a self-hosted RAG chatbot fast.

Langflow

Pro — open-source, Python-native, export flows to code.

Con — prototyping-first; production is on you.

Choose for Python teams prototyping toward a codebase.

Dify

Pro — full LLMOps: RAG, agents, prompt IDE, observability.

Con — heavier, more opinionated platform.

Choose to operate AI apps end to end, not just one chain.

SaaS — let them run it

  • Zero ops: no servers, upgrades or backups to mind.
  • Fastest to start; usually the biggest connector catalogs.
  • But your data and credentials pass through a vendor, and cost scales with usage.

Self-host — run it yourself

  • Data residency and access to internal systems.
  • Flat cost at volume; no per-task meter.
  • But you own uptime, scaling, security patching and backups. n8n, Flowise, Langflow and Dify all self-host.
1It's always trigger → nodes → action. Learn the vocabulary once; every tool reads the same.
2Match the tool to the job — Zapier for breadth, Make for logic, n8n for control, the AI-native trio for RAG and agents.
3Self-host buys control, costs ops. Choose SaaS until data residency or volume makes self-hosting pay.
4Keep a human on high-stakes actions. Let the model draft; let a person send, until trust is earned.
5Graduate at the ceiling. When logic resists testing and review, move the hard core into code — keep no-code for the rest.

A 60-second decision guide

  • Non-engineer, simple "when X do Y"? → Zapier. Ship it today.
  • Need branching, loops, real data mapping? → Make.
  • Technical team, data must stay in-house, want code? → n8n, self-hosted.
  • Building a RAG chatbot or agent? → Flowise or Langflow to prototype; Dify to operate it.
  • Logic you can't test, review or trust? → graduate to code: LLM Apps, Agents, RAG.
Knowledge check

Did it stick?

Five quick questions on triggers, the classic automators, AI-native builders, and the no-code ceiling — instant feedback, no sign-in.

Rate this deck
be the first

Navigate with ← → or scroll · back to library