How do you build your first AI automation in n8n or Zapier tonight, and which one should you pick?

Your first AI automation should be three steps, built in one evening: a form submission or an email arrives, an AI step labels it lead, support or spam and writes a two-line summary, and the result lands in a Slack channel or a Google Sheet. Pick n8n if you want to see what every step did and are willing to paste in an OpenAI key, and Zapier if you want the AI step with no key and accept paying per run. I based the prices and step names below on the n8n, Zapier and Make pricing pages and the n8n and Zapier docs for their AI steps, all opened on 26 October 2026, plus the automations of this shape that run my own consultancy, where the AI steps do the sourcing, research and first drafts and I check them and send.

The hub guide gives the shape and the five rules; this page is the build.

What is the smallest useful AI automation, and why is it three steps?

The smallest useful AI automation moves one message from where it arrives to where someone reads it, with one AI step in between that turns free text into a label and a short summary. It is three steps because a trigger and a destination are the minimum for anything to happen, and one AI step is the only part a plain filter could not do.

A trigger is the event that starts a run: a new form response, a new email, a new row in a sheet. An AI step is a node that sends text to a language model with your instructions and gets text back. A destination is the app that receives the result. In n8n these are nodes on a canvas; in Zapier they are steps in a Zap. The no-code automation glossary entry has the general shape.

The worked build here is lead triage. A website form or a shared inbox receives a mix of prospects, existing customers with problems, and people selling things. The automation does the deciding and the summarising; a person still replies. The sourcing and research flows that feed my own pipeline started at this size and grew only after a week of runs matched what I would have done by hand.

Should you start in n8n, Zapier or Make?

The n8n trial costs nothing and the Zapier build needs a paid plan from day one, so n8n is the cheaper place to find out which you prefer. Make's pricing page did not print the paid prices when I opened it, so Make's price is missing from the table, which holds only what each vendor's page said on 26 October 2026.

	n8n Cloud	Zapier	Make
Entry paid plan	Starter, "20€/mo, billed annually", 2.5K executions a month, unlimited active workflows	Professional, "Starting from $19.99 /month", with a "Pay yearly (Save 33%)" toggle; the tasks included at the entry price are not printed	Core, Pro and Teams are named, no prices printed on the page I opened
Free tier	Free trial, no credit card, 1,000 executions, length not stated. A self-hosted version is on GitHub	Free, "$0 /month", 100 tasks a month, two-step Zaps only, 15-minute polling	"1000 credits per month" on the Free plan
How an AI step is added	OpenAI node with your own OpenAI credential; "Agents, MCP and AI nodes" are listed on Starter	AI by Zapier step, no OpenAI account needed, on Professional, Team and Enterprise; on Free it can only be previewed in an unpublished Zap on the Standard tier	AI modules exist; the page says AI Provider features "may use more" credits
What one run costs you	One execution, "a single run of your entire workflow", however many steps	For the AI step: 1 task on Standard, 3 on Advanced, 5 on Premium, 1 with your own key	Credits per operation, more for AI
Where it runs	n8n's cloud, or your own server	Zapier's cloud	Make's cloud
Who it suits	People who want to read what each step did and run many flows for one price	People who want the fastest first run with no API key	People who want a large free allowance and will check the price themselves

Zapier's Free plan can't run it. The flow has three steps, Free allows two, and AI by Zapier is marked as not available on Free for published Zaps. In Zapier this build is a Professional plan from day one.

n8n's price does not grow with the number of steps. An execution is one run of the whole workflow, so a logging node or a second destination later costs nothing extra. Zapier counts tasks, and the AI step on the default Premium tier is 5 tasks per run.

n8n needs your own OpenAI key. The OpenAI node docs say to "Create or select an existing OpenAI credential", and from then on the model usage is on your OpenAI bill. Zapier's blog on the ChatGPT integration says AI by Zapier runs "without using your own OpenAI account" and recommends it over connecting your own ChatGPT account.

If you are undecided, build it in n8n during the trial. The 1,000 trial executions cover a week of parallel running for any inbox under about 140 messages a day (1,000 over 7 days).

How do you set up the trigger (form or inbox) in about ten minutes?

The trigger is the step you set up first and test on its own, before any AI is involved. Ten minutes is enough if you pick one source and resist adding a second.

In n8n, add the first step and choose the form trigger, which gives you a hosted form URL with the fields you define, or the Gmail trigger, which polls a mailbox. For a form, add three fields: name, email, message. For an inbox, set a label or a search filter so the trigger only sees the messages you want triaged, never the whole mailbox. Test the trigger by submitting the form once or sending one email. The trigger node now shows the exact fields it received; the AI step will map to them.

In Zapier, create a Zap, choose the trigger app (your form tool, or Gmail with a search filter), connect the account, and run the test. Zapier pulls a recent real record and shows its fields. The polling interval is 15 minutes on Free, 2 minutes on Professional and 1 minute on Team, per the pricing page, so a test email can take a quarter of an hour to appear on Free.

The failure case at this step is the test record. Both tools test on one sample, and if yours is a "testing 123" submission, the AI step you build next will be tuned to clean input. Submit a realistic message: a pricing question with a company signature, or a real spam email forwarded to the inbox. In my builds the first thing that broke was a message field with line breaks the sample never had.

How do you write the brief for the AI step so it returns a one-word label?

Write the brief as a job description with a fixed output shape, and switch on structured output in the tool so the model cannot answer in prose. The glossary entry's common mistake applies here: the next step expects a label, gets a paragraph, and the mapping breaks.

In n8n, add an OpenAI node after the trigger. The docs call the operation "Generate a Chat Completion" (renamed from "Message a Model" in n8n 1.117.0). It takes a Model, a Messages list with roles User, Assistant and System, and a switch called "Output Content as JSON", which the docs say to "Turn on to attempt to return the response in JSON format". The docs suggest gpt-4o-mini for speed and low cost. Put the brief in a System message and drag the form fields into a User message.

In Zapier, add an AI by Zapier step. The help page lists templates for Summarize, Write, Classify and Extract in the Prompt field, and under Settings an "Output Fields" section where you add a "Field name", a "Field type", a "Field description" and whether the field is required. Define two fields: label and summary. Set the model to Standard; the help page says it is "Suitable for simple prompts without tools" and costs 1 task per run against 5 on Premium, the default on paid plans.

The brief is the same in both tools. This is the one I use, edited to a generic inbox.

You triage incoming messages for a small consultancy.

Return exactly two fields.
label: one of lead, support, spam. No other value.
summary: two sentences, no more than 40 words. Sentence one
says who wrote and what they want. Sentence two says what
the reader should do next.

Definitions.
A lead is a message from a company email address asking about
pricing, timelines, availability or a call.
Support is an existing customer, or anyone referring to an
existing order, invoice, login or delivery, with a problem.
Spam is anything selling us a product or service, and
messages with no readable request.

Edge cases.
A customer asking to buy more is a lead.
A recruiter is spam. A journalist asking for a comment is a lead.
If you cannot tell, use support, so a person looks at it.

Example.
Message: "Hi, we are a 40-person agency, what do you charge
for a 2-week sprint and when could you start?"
label: lead
summary: An agency asks for sprint pricing and a start date.
Reply with the pricing page and two slots this week.

The allowed values give the model a closed set. The edge cases are the ones a person would argue about, written down once. The fallback is support, because a mislabelled lead in a support queue gets a reply, while one marked spam gets nothing.

What goes wrong here is the open-ended word. "Decide if this is important" returns a paragraph about importance. "Classify as lead, support or spam" with definitions returns a label. If a label comes back outside the set, such as "Lead" with a capital letter or "lead (probably)", the structured output switch is the fix: "Output Content as JSON" in n8n, Output Fields in Zapier.

How do you send the result somewhere a person reads it, and test it on real data?

Send the result to a Slack channel or a Google Sheet that you already look at every day, and run the automation on real messages for a week while the manual process keeps going. Those are the hub's rules three and five.

For Slack, add a Slack node (n8n) or a Slack step (Zapier) after the AI step, pick a channel you created for this purpose, and compose the message from fields: label, summary, sender's email, original message. Keep it out of channels other people rely on until the week is over.

For a sheet, add a Google Sheets append-row step with one column per field: timestamp, sender, label, summary, original message. The sheet is also your log, the hub's rule four. n8n saves every run as an execution showing what each node received and returned, and Zapier has Zap history; the sheet is the version you can sort and count.

The mapping is what breaks at this step. A Slack message that reads {"label":"lead","summary":"..."} means the destination received the whole JSON object instead of the fields, which happens when "Simplify Output" is off in the n8n OpenAI node or the fields were not expanded. A sheet with the summary in the label column means the columns were mapped in the wrong order. Both show in the first test run.

Then the real test. Activate the workflow (n8n) or publish the Zap (Zapier) and leave the manual process running. Each day, open the sheet or the channel next to the inbox and compare: did the label match what you decided, and would you have sent the summary to a colleague? The same compare-to-source habit for any AI answer is the five-minute fact-check routine. The hub's week-3 checkpoint is 20 real inputs compared by hand. The first 20 runs are where the gaps in the brief show up, and each gap becomes a line under "Edge cases". Only after a week with no leads marked as spam do you let the flow write anywhere a customer could see, and even then I keep the reply as a draft.

What does it cost per month to run, and when do you outgrow the free tier?

For an inbox receiving 300 messages a month, the flow runs 300 times, which fits inside n8n Cloud's Starter plan of 2,500 executions at 20 euros a month billed annually. In Zapier the same 300 runs cost at least 300 tasks for the AI step on the Standard tier, or 1,500 on Premium, plus the destination step, on a Professional plan from 19.99 dollars a month.

Cost line	n8n Cloud	Zapier	Make
Platform, entry paid plan	20€/mo, billed annually (Starter, 2.5K executions)	From $19.99 /month (Professional), 33% less on yearly billing	Not printed on the page I opened
What 300 runs a month use	300 of 2,500 executions	AI step: 300 tasks on Standard, 900 on Advanced, 1,500 on Premium, plus the destination step	Credits per operation, number not stated for AI modules
Day you outgrow free, for this flow	When the 1,000 trial executions run out, or the trial ends	Day one, because the flow has three steps and an AI step	When 1,000 credits are used; an AI step may use more than one

The n8n figure leaves out the OpenAI bill for the model, which I did not open a pricing page for, so check it after the first week.

The Zapier figure depends on the model tier, and the help page's formula makes it checkable: tasks used per run equals one times the model rate, plus tool calls times the model rate. With no tools on the Standard tier, that is 1 task for the AI step. New steps default to Premium at 5, so change it before you publish, or the first month's task count will be five times what you planned. The page also sets a per-run task limit, 75 by default and up to 500, after which Zapier pauses the step until someone approves it.

Make is the one I could not price for you. Its page says there are 1,000 free credits a month and that annual payment is cheaper than monthly, but no amounts were printed when I opened it. If you are choosing Make, open the page yourself and read the credit rate for AI modules before you commit.

n8n's trial is outgrown when its 1,000 executions are gone, about three months at 300 messages a month, or earlier if the trial has a time limit the page does not state.

What breaks first, and how do you know?

The trigger breaks first, then the brief, then the destination, in my experience, and you know because the log shows a run with an empty field, a label outside the set, or no run at all. Each has a one-minute check.

The trigger fails silently. A form field gets renamed, a Gmail label gets deleted, or the connected account's password changes, and the trigger stops firing or fires with empty fields. The check is a count: compare the rows in the sheet for the last 24 hours with the messages in the inbox. If the inbox has more, the trigger missed some. In n8n, open Executions and look for runs that stopped at the first node; in Zapier, open Zap history and filter for errors.

The brief fails once a new kind of message arrives that it did not define, and the model guesses. Check it with the 20-input review, repeated monthly, with each wrong label added as an edge case. The n8n glossary entry names the other version: a workflow with no error path "stops silently and can stay broken for weeks". In n8n, add an error workflow that posts the execution link to your own channel when any node fails. In Zapier, look for steps paused at the per-run task limit.

Destinations break when the app changes: a Slack token is revoked, a column is inserted in the middle of the sheet, the channel is archived. Count from the other side: completed runs in the log with no matching row in the sheet.

The rule that catches all three is the hub's: log every run with input, output and timestamp, and look at the log weekly, in a calendar block you set on the day you publish the flow.

The free module Your First Real Automation in n8n/Zapier builds this exact flow in 15 minutes with the error path and the log. If you have never built an automation, the beginner module No-Code Automation: Zapier & Make covers the trigger-and-step shape in 12 minutes first. When you want an assistant to act in your tools from a chat rather than on a schedule, the guide on MCP is the next read.