Workflow
A workflow is a fixed chain of steps that runs on its own, with an AI model doing the thinking inside some of them.
01 · Introduction
What a workflow is
A workflow is a series of steps that run in a set order, started by a trigger.
Example: when a new row lands in a spreadsheet → send its text to an AI model to classify it → if it is a complaint, post it to a Slack channel and create a ticket.
Each box is a step (also called a node). The arrows are the order. The first box is the trigger: a schedule, a new email, a form submission, a webhook.
Workflow vs. agent: a workflow follows the same path every time, which makes it predictable and cheap. An agent decides its own path, which makes it flexible but harder to trust. Many real automations are a workflow with one or two "AI steps" inside.
02 · When to use it
When a workflow is the right choice
Pick a workflow when:
- The steps are always the same and you can draw them on paper.
- It must run unattended: nightly, on every new email, on every order.
- A mistake is costly, so you want each step visible and testable.
- Several tools must be connected: forms, sheets, mail, chat, CRM.
Typical first workflows for a working professional:
- Summarize new support emails into a daily digest.
- Tag incoming leads and write the first reply draft.
- Turn meeting notes into tasks in the project board.
- Translate every new help article into a second language.
03 · Setup
Building your first one
Use a visual tool before writing code. n8n (self-hosted or cloud), Make and Zapier all work the same way: drag steps onto a canvas and connect them.
- Start from the trigger. Pick "new row in Google Sheet" or "every day at 8:00".
- Add an AI step. Paste a clear prompt, for example: "Classify this message as complaint, question or praise. Answer with one word."
- Add a branch ("if complaint").
- Add the actions: send a message, create a record, write back to the sheet.
- Run it once by hand with a real example and look at the output of every step.
- Switch it on.
Self-hosting n8n is one Docker command, which is why Docker is in this handbook set.
04 · How to use it
Day-to-day use
- Keep each AI step small: one question, one short answer. Chain steps rather than writing one giant prompt.
- Ask the model for a fixed format (one word, or JSON with named fields) so the next step can read it.
- Log every run. When something looks wrong a week later, you need to see what the model was given.
- Put a human check before anything irreversible: sending to a customer, paying, deleting.
- Set a budget alert on the model's API key. A loop in a workflow can spend money fast.
- Name workflows by what they do, not by the tool: "Daily support digest", not "Sheet to Slack v3".
05 · Pitfalls
Pitfalls
- Silent failures. A workflow that stops erroring may simply have stopped running. Check that it ran today.
- Free-text answers. If the model answers "I think this is a complaint", the branch step cannot read it. Insist on a fixed format and handle anything else as "unknown".
- Loops. A workflow that writes to the same sheet it watches can trigger itself forever.
- Stale credentials. Connected accounts expire. Reconnect them before the Monday run.
- Over-automation. Automate the boring 80 percent and leave judgment calls to a person.
06 · Resources
Where to go next
A workflow connects tools; the next step is giving a model tools directly (MCP) and writing down your procedures so an agent can follow them (skills).