Reference
Automation, workflow, or agent?
The front of your take-home card, with practice. Pick the smallest thing that does the job.
Automation
You know the steps, and none needs judgment.
No AI required.
Workflow
You know the steps, but some need judgment.
Put AI in those steps, and nowhere else.
Agent
You know the goal, but not the steps.
Let AI choose the steps, with limits and a human approving.
The agent test: who picks the next step? A person: it's a workflow. The AI: it's an agent.
The most expensive mistake: an agent doing a workflow's job.
Practice
Six real jobs. Choose, then read why.
When an invoice is paid in QuickBooks, mark the job "paid" in the job tracker.
Automation. Fixed steps and no judgment: a paid invoice is a paid invoice. Adding AI here adds cost and a way to be wrong.
Read each contact-form message, decide whether it is sales, support, or billing, and draft a reply for review.
Workflow. The steps never change, but sorting and drafting need judgment. That is Playbook 1.
"Look through this week's service calls and figure out why repeat visits went up."
Agent. You know the goal, not the steps: what to look up next depends on what it finds. Give it read-only tools and read its log.
Every morning, read overnight email, page someone for emergencies, and draft replies for approval.
Workflow. It sounds ambitious, but you know every step. That is Playbook 3. Making this an agent would be the most expensive mistake.
Send a reminder text the day before every scheduled appointment.
Automation. Same message, same timing, every time. Your scheduling software may already do it.
Answer customer questions using only the company handbook, and flag anything it does not cover.
Workflow. Search, answer, flag: fixed steps, with judgment in the answer. That is Playbook 2.
Four rules for anything you automate
| Rule | What it means | Where you saw it |
|---|---|---|
| One narrow task | Each AI step does one job you can describe in a sentence. | The Sort step in Playbook 1; each specialist in Playbook 6. |
| Audit every run | Every run leaves a log you can read, step by step. | "Read the log" in every playbook. |
| A human approves | Nothing reaches a customer or moves money without a person's click. | The approval step in Playbook 3; Marcus in Playbook 6. |
| Measure the KPI | Pick one number before you start, and check it at 30 days. | Bring it home. |
Where the tools sit
You do not need to write these down. Most businesses start in the first two rows.
No-code, in the cloud
Zapier, Make, Microsoft Power Automate, and Microsoft Copilot Studio. Fastest start, least control, priced per task.
Low-code, host it yourself
n8n, used in every playbook here. Visual like the first row, and you can run it on your own server so your data stays with you.
Personal agents
Assistants that act for one person, such as OpenAI's dots and OpenClaw. Useful for individuals; harder to govern across a company.
Code frameworks
LangGraph and CrewAI. For developers building custom agent systems. Most control, most work.