Back to the studio
Studio Calanque FR

AI agents10 min

n8n AI agent vs Make vs Zapier: what each one costs per month, and which I install

n8n AI agent, Make or Zapier? What each tool bills for, one real month priced four ways, where your data sits, and the one I install by default.

Most of the time I install n8n. An n8n AI agent is billed per complete run, however many steps it takes, and you decide which server it lives on. Make, if your scenarios already live there; at low volume nothing costs less. Zapier for three or four simple zaps your office manager can fix herself on a Friday afternoon. For an agent that loops over the 2 or 3 attachments in every email, almost never.

"Three weeks ago this said 750. Is it going to stop?"

That was the whole email, under a screenshot of a Zapier invoice: Professional plan, 5,000 tasks, $133.50 a month. It came in April from the owner of a yacht crew agency in La Ciotat, and the answer was no, not on Zapier. It stopped when we rebuilt the workflow as an n8n AI agent on a small server in France.

Nothing was wrong with the agent. It reads about 320 applications a month in season, and when the STCW certificate or the ENG1 medical is missing it emails the candidate. When the PDF comes back, the workflow runs again from step 1: about 8 steps, nearly all of them billed as Zapier tasks.

Then a captain asked where the passport scans of his 9 crew members had been sitting since March, and who could open them. The owner had no idea. More on that below.

What you are comparing

These three tools are not the AI. They're orchestrators. When an email arrives, the tool sends it to a language model (GPT, Claude or one of Mistral's), gets back what the model decided, and then does it in your other software, HubSpot or Google Calendar for most of my clients. The model bills you separately for the text it reads and writes, and that cost is the same whichever orchestrator you use. If the difference between an agent and a classic automation is still unclear, I ran one real customer message through both in AI agents, explained for business owners.

Zapier, easy until the agent loops

"It took me twenty minutes," the owner of a dive center in Cassis told me last spring. That was for hooking the booking form to a Google Sheet with Zapier, no developer, no call to me. They're still on the free plan, 100 tasks a month, and they've never gone over.

The crew agency went over in a week. Zapier uses one task each time a step succeeds, and triggers or built-in tools like Formatter and Filter don't count. Reading the email, extracting the attachments, calling the model, looking up the candidate and writing the reply came to 7 or 8 tasks per application, and about 4 applications in 10 came back a second time. On monthly billing, Professional costs $29.99 for 750 tasks, $73.50 for 2,000 and $133.50 for 5,000. Billed yearly, the same tiers cost $19.99, $49 and $89 a month.

Zapier Agents is billed separately, in "activities". The Pro add-on includes 1,500 a month, and a run can use 40 at most before the agent stops and asks for confirmation. That's enough to test an idea, not to process a season of applications.

Then there's storage. According to Zapier's documentation, customer data is kept in AWS US-East-1, in the United States, and "region-specific data residency is not available unless otherwise agreed." I don't know of any small business that has negotiated that.

Make, when you want to see the wiring

Make shows each scenario as a diagram of bubbles connected by lines. That's its main advantage for me: someone who has never written code can look at it and see what happens.

An e-bike rental in Cassis had 14 scenarios in Make, built by the owner's nephew one winter. In July they get about 25 emails a day about the climb up the Route des Crêtes and how long the batteries last, and last year they asked me for an agent to answer them. I didn't migrate anything. The AI step went into scenario number 9 in an afternoon, and the nephew still maintains the other 13.

Make bills in credits. Most module actions use one credit. Be careful with iterators: if one goes through 50 rows, the next module runs 50 times. The Core plan costs $10.59 a month for 10,000 credits on monthly billing, or $9 a month on an annual plan. At low volume that's much cheaper than the other two, and I tell clients so even when I'd prefer to put them on n8n.

Make has two European data zones, EU1 and EU2. The rental's owner had chosen one and thought that settled the privacy question. But Make's help page on AI providers says that choosing a zone "does not automatically mean every request gets processed and stays in the EU." OpenAI models used through Make's AI provider are processed in the US, whatever the zone. Plug in your own OpenAI or Anthropic key and the requests go to that provider's default endpoint, with no promise of EU processing.

Why an n8n AI agent is my default

n8n bills per execution, and one execution is one complete run of a workflow, whether it has 3 steps or 30. For an agent this is the main point, since the number of steps changes with each message. A candidate who sends three phone photos of certificates instead of one PDF can add four or five steps to a run. On n8n Cloud, Starter costs €24 a month for 2,500 executions (€20 on annual billing), and Pro costs €60 for 10,000 executions (€50). The pricing page says data on hosted plans is stored in the EU, in Frankfurt.

We didn't use the cloud, though. The Community edition is free under n8n's Sustainable Use License when it's your own people building the workflows, and the license FAQ lets a contractor like me install it on a client's server, build the workflows and maintain them, as long as I'm not also the host. So it lives on a VPS at OVHcloud in France, rented in the agency's name, the $8.50 size rather than the $4.54 one, and even that turned out to be a bit small.

There were problems. In the second week a bosun sent his seaman's book as one 140-page scan of 60 MB and the server ran out of memory.

The second problem was the retention setting. n8n's default is to keep successful executions, with their data, until pruning at 14 days (336 hours). For the agency that was two weeks of passport scans on a server for no reason. We set EXECUTIONS_DATA_SAVE_ON_SUCCESS=none, as n8n's scaling docs recommend, so only failed runs are kept. There's also a 15 MB limit on attachments now, with an automatic reply asking for the relevant pages, and updates on the first Tuesday of each month.

n8n is harder than Make for beginners. Some logic needs about ten lines of JavaScript in a Code node, and at the agency I'm the only person who edits them.

One real month, priced four ways

The numbers below are from May at the agency, the first full month on the new setup, when 320 applications came in. Each took about 8 steps, and 4 in 10 came back for a second pass, so the month had about 3,600 steps and 450 complete runs. Make would have used more credits than steps because of the attachment loop.

Tool What it counts Plan that covers it List price per month
Zapier each successful step Professional, 5,000 tasks $133.50
Make each module action Core, 10,000 credits $10.59
n8n Cloud each complete run Starter, 2,500 executions €24
n8n self-hosted nothing OVHcloud VPS-2 from $8.50

None of that includes the model. Ours is a Mistral model, which cost about €30 in May going by the invoice, and it would have cost the same behind any of the four tools. Setup isn't in there either; at my studio in the south of France it starts at €900 before tax and takes one to three weeks.

The agency's owner asked the obvious question. Why not Make, at $10.59 a month and no server to look after? Fair. If she'd already been running Make and hadn't cared about an American model reading crew passports, I would have.

Where the data goes

Two weeks after the migration, the captain of a 50-meter motor yacht wrote to the agency. He wanted to know where the copies of his crew's passports were stored and who could see them. The owner forwarded his email to me without comment.

Is it illegal to send crew data to the US? No, and I get told it is about once a month. The European Commission adopted its adequacy decision for the EU-US Data Privacy Framework on July 10, 2023, and for an American company on the Commerce Department's list (the CNIL's page explains it well) you don't need standard contractual clauses or anything like them.

But the owner's question was not about the law. She told me: "I want to answer him in one sentence."

My view, which I know isn't the usual one: people fuss over where the orchestrator is hosted and then send whole emails to a model nobody has read the terms of. It's the model that reads. At the agency the 2 PDFs per candidate stay on the OVHcloud server, a Code node gets the expiry dates out of the certificates, and the model is only given the email text, the names of the files and one line about which certificates are valid. Her one-sentence answer comes from that, more than from n8n.

What I pick

Most of the businesses that ask me this question are small. In the Census Bureau's Business Trends and Outlook Survey, 19.8% of US firms said they were using AI as of May 3, 2026, against 37% of firms with 250 or more employees. A company of 4 or 5 people usually has no one whose job is to maintain a server, and that matters more than any difference in features.

If the agent handles personal data and the client is fine with me maintaining a server, I install n8n on their own VPS; that was 3 of my last 6 projects. No server wanted? n8n Cloud, hosted in Frankfurt. If the client already uses Make and it works, they stay on Make, because a migration would mean paying twice for the same result. Zapier is for simple automations, such as a plumber's quote requests going into a Google Sheet, and not for agents.

And sometimes it's none of the three. When the workflow becomes the main part of the business, with its own screens, client accounts and rules that change every month, I build a web application and there is no orchestrator. At that point it's no longer an automation job but a web application, the third of the services the studio offers.

The tool is the easy part. Which job to automate first is covered in where to start with AI automation, and the scoping, the test on old messages and the first week of drafts in how to create an AI agent for your business.

If your Zapier bill has gone up twice this month, a 15-minute call is usually enough to decide what to do.

FAQ

Is n8n really free?

The self-hosted Community edition is free under the Sustainable Use License, as long as the workflows are built and edited by your own people, or by a contractor working on your server. You still pay a few dollars a month for the server, plus someone's time for updates and backups, about an hour a month. The hosted version starts at €24 a month, or €20 if you pay for a year.

Is Make cheaper than Zapier?

In almost every case I've priced, yes. Make's Core plan gives you 10,000 credits for $10.59 a month, which is less than Zapier charges for 750 tasks on its entry Professional plan ($29.99). Zapier is worth the extra money when the person who maintains the automations won't learn a new tool.

Can you build an AI agent in Zapier?

Yes, with Zapier Agents. It's billed in activities, 1,500 a month on the Pro add-on and up to 40 per run, and it's a quick way to test an idea. For an agent that processes hundreds of files a month I still use n8n, because of the per-step billing and the US storage.

Is n8n GDPR compliant?

Not by itself, and no tool is. n8n Cloud stores data in Frankfurt, and a self-hosted install keeps it on the server you choose, in France for the crew agency. The rest is up to you: limit what the model sees, keep executions for as short a time as you can (the default is 14 days) and sign a data processing agreement with each provider.

Sources

Related