The short answer
To automate a task with AI, pick a repetitive job with a clear trigger, describe it to an automation tool's assistant, then test and switch it on. In Zapier, type the idea into the Copilot box, click Continue with Zap Editor, accept each step, click Test and then Publish. Zapier's free plan allows 100 tasks a month and two-step Zaps only. Make's Maia and n8n's Assistant build workflows from a description too, and n8n's Community Edition is free to self-host.
What you need first
- A task worth automating. Good candidates happen often, follow the same steps every time and start with a clear event: a form is submitted, an email arrives, a row is added, a meeting ends. Write it as "When X happens, do Y, then Z".
- Accounts for the apps involved. You will connect them to the automation tool, so you need permission to do that. At work, check with IT before connecting company email or CRM.
- A decision on where AI belongs. Many automations need no AI at all. Add an AI step only where judgement or language is needed: sorting, summarising, drafting, extracting fields.
- A tool. This guide builds in Zapier because its assistant flow is documented end to end. Make and n8n are covered below.
How to build an automation with Zapier Copilot, step by step
- Log in to Zapier. A Copilot prompt box appears on the home page.
- Describe the automation in one or two sentences, naming the apps: "When a new row is added to my Google Sheet of leads, send me a Slack message with the name and company." You can use the microphone icon to speak it.
- Copilot shows its reasoning and may ask which apps you use. Answer its questions.
- When the proposal looks right, click Continue with Zap Editor. Your conversation carries over.
- Let Copilot build the Zap. It pauses when it needs you, for example to connect an account; accept each step in the dialog if it looks correct.
- Refine by chatting in the prompt box at the bottom of the sidebar, such as "only send the message if the company field is not empty".
- When prompted, click Test. Test records are sample data and may differ from what a live run sees.
- Review every step, then click Publish. The Zap only processes new data created after it is published.
Adding an AI step without losing control
The AI by Zapier step can now sit inside any Zap and do agent-style work: read an email, decide what it is, look something up and return structured fields. Zapier merged its standalone Agents into this step. Four settings keep it safe:
- Tools picker. Add only the tools the job needs. Fewer tools means fewer surprises.
- Require approval before running. Switch this on for any tool that sends, deletes or changes something, so a person checks it first. Read-only tools can run on their own.
- Structured output. Define the fields you want back, such as category, urgency and a one-line summary, so later steps can use them reliably.
- Per-run task limit. A run pauses for review if it goes over its limit: 75 tasks by default, adjustable up to 500.
Watch the cost. Everything in Zapier is billed in tasks, AI steps on stronger models cost more tasks, and each MCP tool call uses two.
A worked example: sorting inbound enquiries
Say a shared inbox gets 40 enquiries a week and someone spends an hour a day sorting them. The workflow: when a new email arrives in the enquiries inbox, an AI step reads it and returns three fields (type: sales, support or other; urgency: high or normal; a one-line summary), the email details go into a Google Sheet, and high-urgency sales enquiries post to the sales Slack channel. In Zapier, describe exactly that to Copilot, continue to the editor and connect the three apps. In the AI step, define the three output fields and give two example emails for each type. Leave "Require approval before running" on for the Slack post for the first week. Test with five real emails of each type, publish, then check the sheet every day for a week. When the sorting is right nineteen times out of twenty, turn approval off for the Slack step and keep it on for anything that replies to customers. Note that a workflow with more than two steps needs a paid Zapier plan.
Common mistakes and how to fix them
- Automating a messy process. If people do the task differently every time, the automation will be wrong half the time. Fix: write the steps down and agree them first.
- Letting AI send things unchecked. Fix: keep approval on for anything customer-facing, at least until you trust it.
- Vague AI instructions. "Categorise this email" gives inconsistent labels. Fix: give a fixed list of categories, a definition of each and examples.
- Running out of tasks or credits. Zapier's free plan has 100 tasks a month; Make's free plan has 1,000 credits a month and built-in AI can cost more than 1 credit a step. Fix: test on a small batch and check the usage page after the first day.
- Expecting instant triggers on Free. Zapier's free plan checks for new data every 15 minutes. Fix: accept the delay, or upgrade if timing matters.
- Forgetting the data. Customer emails pass through the automation tool and the AI model. Fix: check each tool's data terms; Zapier, for example, offers a training opt-out form, and Enterprise content is excluded from model training.
Zapier, Make, n8n or your chatbot: which should you use?
| Tool | Pick it if | Free option | The catch |
|---|---|---|---|
| Zapier | You want thousands of ready-made app connections and the easiest start | 100 tasks a month, two-step Zaps only | Costs climb with AI steps and multi-step Zaps |
| Make | You prefer a visual canvas where you can see and branch every step | 1,000 credits a month; 30-day Maia trial | Agents and some AI features are still in beta |
| n8n | You have technical help and want full control or self-hosting | Community Edition is free to self-host; cloud has a free trial | Self-hosting means you own updates, security and backups |
Make's AI co-worker, Maia, sits in the Scenario Builder and builds scenarios from a description, with Build together and Plan first modes. The n8n Assistant plans, builds and fixes workflows from a request and asks before it uses credentials or activates anything.
For a simple recurring job with no other apps, your chatbot may be enough: ChatGPT Plus has scheduled tasks, Claude Pro has scheduled tasks and Gemini's Schedules runs up to 10 scheduled prompts at a time. Compare the automation tools in Zapier vs Make, Make vs n8n and Zapier vs n8n, or read the best AI for automation and Zapier alternatives.
Prompts to copy
Find what to automate
Here is a list of things I do every week at work: [list]. For each, say whether it is worth automating, the trigger event, the steps, which apps are involved and whether it needs an AI step or just rules. Rank them by hours saved per month.
You get: A ranked shortlist of automations worth building
Describe it to Copilot
When [trigger event in app], [step 1 in app], then [step 2 in app]. Only continue if [condition]. Ask me before connecting any account.
You get: A clear instruction an automation assistant can build from
AI sorting step
Read this email and return three fields. type: one of [sales, support, invoice, other], using these definitions: [definitions]. urgency: high if [rule], otherwise normal. summary: one sentence under 20 words. If you are unsure of the type, return other. Email: [email body]
You get: Consistent structured fields that later steps can use
Test cases
Write 10 realistic test inputs for this automation, including three edge cases that could break it, such as an empty field, a reply in another language and a forwarded thread: [describe automation].
You get: Test data that finds problems before real data does
Failure check
My automation failed with this error: [paste error]. The steps are: [list]. Explain in plain English what went wrong, the most likely fix, and how to stop it happening again.
You get: A plain-English diagnosis and fix