---
id: con-automation
schema: concept/1
layer: garden
status: established
name: Automation
aka: ["workflows", "AI workflows", "Zaps", "flows", "no-code automation", "agentic workflows"]
summary: >
  Setting AI up to handle a repeatable job on its own, triggered by a time or an event like a new
  form entry, so the work happens without you starting it each time.
first_seen: unknown
relations:
  builds_on: [con-tools-and-connectors]
  evolved_from: []
  related: [con-evals]
  parallels:
    - idea: "an assembly line"
      field: "manufacturing"
      note: "Break a job into fixed steps, and put inspection points where errors would be costly."
    - idea: "mail merge"
      field: "office work"
implementations:
  - { provider: zapier, name: "Zapier AI steps and agents" }
  - { provider: make, name: "Make AI scenarios" }
  - { provider: n8n, name: "n8n AI workflows" }
  - { provider: microsoft, name: "Power Automate with Copilot" }
competencies: [cmp-make-it-repeatable, cmp-check-the-result]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "Where 'make it repeatable' ends up; links the AI map to the long history of workflow automation."
---

Automation means a task runs without you pressing go. Something triggers it (a new form submission, an email arriving, a date) and a chain of steps follows, some done by AI. A new volunteer signs up, AI drafts a welcome email in your voice, and it lands in your drafts folder for a quick look.

Workflow automation has been around for decades. AI adds steps that need judgment: summarising, sorting by meaning, drafting replies, pulling details out of messy text. Those steps can also get things wrong, so plan for mistakes. Put a person at the points where a mistake would be costly, and send the AI's output somewhere you will see it before it goes out.

A good rule is to automate something only after you've done it by hand with AI several times and know what good looks like.

## Example

A nonprofit's intake form feeds into a workflow: AI reads each request, tags it as housing, food or legal, and drafts a reply. A caseworker approves the replies each afternoon. Misfiled requests drop from common to rare after one prompt fix.

## How it connects

- Needs **tools and connectors** to reach the apps involved.
- **Scheduled tasks** are the simplest version; **agents** handle the messier ones.
- **Human in the loop** checkpoints and **evals** keep it trustworthy.

## Try it

Write down a task you've done with AI at least three times, then mark the step where a mistake would hurt most. When you automate the task, put your approval step there.