[Beginner] · Triggers and actions — and the specific point at which a platform beats a pile of cron scripts.

Workflow automation is connecting systems that weren’t designed to talk to each other, so that something happening in one causes something to happen in another, without a person in the middle.

A form submission creates a CRM record, posts to Slack, and sends a templated email. An invoice PDF lands in a mailbox, gets its fields extracted, and appears in the accounting system. A support ticket is classified and routed. None of these are hard problems. All of them are tedious, and all of them break in the same handful of ways.

Triggers and actions

Every platform in this category has the same two primitives:

Trigger — what starts a run. A webhook, a schedule, a new row, a new email, a file appearing, a manual click.

Action — what happens next. Call an API, transform data, branch on a condition, write to a database, send a message.

A workflow is a graph of these. An execution is one run of it, with its own data and its own log.

That’s the whole model, and its simplicity is the point — it’s the same shape as an event-driven system → building blocks, with the wiring done in a UI rather than in code.

The honest question: why not just write a script?

For one task, write the script. A cron job calling two APIs is 30 lines and needs no platform.

The case for a platform appears at a specific threshold, and it’s worth naming precisely — because adopting one too early is real overhead:

You haveA script is fineA platform starts winning
1–3 automations✓
Dozens, owned by several people✓
One integration✓
Many, each with its own auth✓
You are the only maintainer✓
Non-engineers need to change it✓
Failures are rare and you notice✓
You need retries, alerts, and an audit trail✓

What you’re actually buying is not the ability to call an API — it’s the surrounding machinery you’d otherwise build repeatedly:

  • Credential management for 200 services, with OAuth flows handled
  • Execution history — what ran, with what data, and what it returned
  • Retries, error branches and alerting → note 05
  • A visual representation a non-engineer can read and change
  • Scheduling and queueing without you running a worker

Fifty small scripts on a box nobody remembers is a genuine operational problem. That’s the failure mode this category exists to prevent, and it’s why the argument is about maintenance, not capability.

Where it sits against its neighbours

Not the same as a CI/CD pipeline → CI/CD. Those are triggered by code changes and produce artefacts. Overlapping tooling, different job.

Not the same as a durable workflow engine — Temporal, AWS Step Functions, Airflow. Those are code-first, built for long-running, strongly-guaranteed, high-volume orchestration, and they’re what you graduate to when the automation is the product → distributed systems.

Not an ESB or an ETL tool, though the boundaries blur. ETL moves bulk data on a schedule; this moves individual events reactively.

The rough positioning:

   script  →  automation platform  →  durable workflow engine
   (n8n, Zapier, Make)                (Temporal, Step Functions)

   one task            tens–hundreds        thousands/sec,
   one owner           several owners       exactly-once,
   fails silently      retries + history    survives anything

Most teams need the middle column and reach for either end.

The landscape

ModelSelf-hostBest at
n8nNode graph, code when neededYes (fair-code)Technical users; the default here
ZapierLinear, huge app catalogueNoNon-technical, breadth of integrations
MakeVisual, more branching powerNoComplex visual flows
TemporalCode-firstYesDurable, high-scale, engineering-owned
Airflow / DagsterCode-first DAGsYesScheduled data pipelines

This folder uses n8n because it’s self-hostable (so the data stays yours), node-based but with real code escape hatches, and its concepts transfer to the others → note 02.

Where the AI part comes in

The recent shift: an LLM is now just another node in the graph. That makes a category of previously-impossible automation trivial —

  • Classification — is this email a bug, a billing query, or spam?
  • Extraction — pull the total, date and supplier from this invoice
  • Generation — draft a reply for a human to approve
  • Routing — decide which branch this should take

What it changes structurally: automation used to require every input to be structured. An LLM node lets a workflow accept messy human input — free text, a PDF, an email — and produce something the rest of the graph can branch on.

What it doesn’t change: the model can be wrong, so anything consequential needs a human in the loop or a validation step → note 04 · verifying what it tells you.

Source: [reference] — written Aug 2026 against the plan this folder’s README had carried since July.

↑ Back to AI Automation