AJ Learning Hub

Beginner · Lesson 6 of 6 · 18 min

Your first AI workflow

Build a no-code flow from a trigger to an AI step to an action, with a human review step so nothing reaches a customer unchecked.

The shape of every AI workflow

Almost every useful AI automation follows the same shape: a trigger starts it, an AI step interprets or drafts, a review step lets a person check, and an action does something in the real world.

A concrete example: a customer submits a support form (trigger). A model labels the message and drafts a reply (AI step). The draft lands in a review queue where a team member approves or edits it (review). Only then is the reply sent and the CRM record updated (action).

Build the review step in from the start. It is far easier to remove a review step once you trust the output than to explain to a client why the system emailed a customer something wrong.

Pick a platform you can finish on

n8n offers a Cloud free trial with no credit card and a free self-hosted community edition. Its AI Agent node needs at least one tool connected, so for a simple single-prompt step use the Basic LLM Chain node instead. Make has a free plan billed in credits, and its built-in AI features can use more than one credit.

Zapier's free plan is limited to two-step Zaps, so a trigger, AI step, review and action flow needs a paid plan or trial. Its AI by Zapier step can require approval before running, set per tool. If your leads already live in GoHighLevel, its workflows are the natural home for anything that touches contacts, pipelines or messages.

For the AI step you need model access. Use your platform's built-in AI step if it has one, or create a key on a provider's free tier, such as the Gemini API's free tier in Google AI Studio.

Design the AI step

Send only the fields the model needs, usually the message text and maybe the customer's first name. Do not send full CRM records or payment details.

Use a fixed list of categories with an explicit 'other' option, and ask for output in labelled lines or JSON so the next step can map each field. The prompt from the prompting lesson works here almost unchanged.

text
Prompt in the AI step:
Classify this support message as billing, scheduling, complaint or other.
Then draft a short, polite reply. Do not promise refunds, dates or prices.
Reply in exactly this format:
category: <category>
draft: <reply>

Message: {{form.message}}

Mapped into the review sheet:
timestamp | customer_email | message | category | draft | status=pending

Add the human review step

The simplest review queue is a spreadsheet tab: the workflow appends a row with the draft and a status of pending. A person reads it, edits the draft if needed and changes the status to approved. A second workflow, or a later step, only sends rows marked approved.

Other good queues are a Slack message with approve and reject buttons, or a task in your CRM assigned to an owner. Whatever you choose, someone must own the queue and check it on a schedule, or drafts pile up and customers wait.

What breaks if you skip it: a model that misreads sarcasm, invents a policy or promises a refund will do so directly to a customer, at the speed of automation.

Test with ugly inputs and watch the runs

Test with at least five realistic messages, including an empty message, a very long one, an angry one and one in another language. Each platform keeps a run history (Executions in n8n, run history in Make, Zap History in Zapier). Open each run and confirm what went in and what came out.

Decide what happens on failure before you go live. n8n lets you assign an error workflow that starts with an Error Trigger node and can send you an alert. A failed AI call should alert a person, not silently drop a customer's message.

Key points

  • ◆Trigger, AI step, human review, action is the shape of most useful AI automations.
  • ◆Choose a platform whose free tier supports enough steps to finish the flow.
  • ◆Send the model only the fields it needs and ask for a fixed, mappable format.
  • ◆A review queue needs an owner and a check schedule, or drafts pile up.
  • ◆Test with messy inputs, read the run history and set up failure alerts before going live.

Terms in this lesson

TriggerA trigger is the event that starts an automation, like a new form submission or a set time of day.ActionAn action is a task the automation does after it starts, such as sending an email or creating a contact.Workflow AutomationWorkflow automation means letting software do a routine process for you, step by step, every time a certain event happens.Human-in-the-LoopA person checks and approves what the AI wants to do before it actually happens, especially for risky steps like sending money or messaging customers.n8nn8n is a workflow automation tool you can use in the cloud or run on your own server, with strong support for code and AI agents.MakeMake is a visual automation tool where you connect apps by dragging steps onto a canvas and linking them into a flow.ZapierZapier is an online tool that connects thousands of apps so one app's events can automatically trigger work in another, with no coding.No-Code / Low-CodeNo-code means building apps or automations by clicking and dragging instead of writing code. Low-code is the same idea, with the option to add a little code when needed.

Try it · exercise

Build a support-triage workflow on a free tier that drafts replies into a review sheet and sends nothing on its own.

  1. Create a Google Sheet with two tabs: Inbox (timestamp, email, message) and Review (timestamp, email, message, category, draft, status).
  2. In an n8n Cloud trial or a Make free account, create a workflow triggered by a new row in the Inbox tab.
  3. Add an AI step using the platform's built-in AI or a free-tier model key, with the prompt from this lesson.
  4. Add a step that appends a row to the Review tab with the category, the draft and status set to pending.
  5. Add three test rows to Inbox, including one angry message, and let the workflow run.

You're done when: The Review tab shows three new rows, each with a category from the allowed list, a draft reply and the status pending, and no email or message was sent to anyone.

Quiz me on this lesson