You have been told to “automate” things. You looked, and the first thing you hit was three names: Zapier, Make, and n8n. Every comparison you found was written by someone selling one of them, or by a developer who assumes you already know what a webhook is. You do not have a developer. You have a business and maybe an hour a week to think about this.
This is the n8n vs Make vs Zapier comparison we would give a friend who runs a company: how each one charges you, where each one breaks, where your data goes, how each handles an AI step, and which kind of business should pick which. The short version: Zapier for a handful of simple workflows, Make for volume on a budget, n8n when AI or sensitive data is central or you want to own the system outright.
What these tools actually are
All three are automation platforms: software that watches for something to happen in one system (a form is submitted, an invoice is paid, an email arrives) and then does a series of things in other systems (adds a row to a spreadsheet, sends a text, creates a task). Each of those things is a step. A chain of steps is a workflow. Zapier calls it a Zap, Make calls it a scenario, n8n calls it a workflow.
The ordinary-business analogy is an office assistant with a binder of standing instructions: “When a new job comes in from the website, put it on the schedule, text the customer, and tell the estimator.” The platform is the assistant. The binder is the workflow you build. The differences between the three are how the binder gets written, what each instruction costs, and what happens when the assistant hits something the binder does not cover.
1. Learn how each one charges you, because it decides everything else
Zapier charges per task: any step that does something counts, so a trigger plus four actions is about four tasks per run. Make charges per operation, and nearly every module that executes counts as one. n8n’s cloud plans charge per workflow execution, meaning one run of the whole workflow counts once no matter how many steps it has. Self-hosted n8n has no per-run charge at all.
Work it through with a number. A twelve-step workflow that runs 1,000 times a month is roughly 12,000 tasks on Zapier, roughly 12,000 operations on Make, and 1,000 executions on n8n cloud. Entry plans cost about what a cheap subscription costs (check each current pricing page), but at volume the gap opens wide. Count your steps and your monthly runs before you pick.
2. Zapier has the widest catalogue and the highest price per step
Zapier’s strength is breadth and ease. It lists thousands of integrations, the interface reads like a form, and a non-technical owner can build a working two-step Zap in ten minutes. Its reliability record is good and its documentation is the best of the three. If you need to connect a niche piece of software, Zapier is the one most likely to already have it.
Where it breaks is cost and logic. Branching (Zapier calls it Paths) is clunky and every branch adds tasks. Looping over a list is awkward. A workflow with real decisions in it becomes expensive and hard to read. Zapier is the right call when you will have fewer than ten workflows, each with fewer than five steps, and nobody will ever want to look under the hood. That describes a lot of businesses and there is nothing wrong with it.
3. Make is the visual canvas, the cheapest start, and the operations trap
Make (formerly Integromat) draws your workflow as circles connected by lines. For people who think visually this is a real advantage: you can see the branching, see where data flows, and see which module failed. It has proper routers (branching), iterators (do this for each item), aggregators (collect results back together), and error handlers. It is far cheaper than Zapier per unit of work, and many businesses run everything they need on a plan that costs less than a lunch.
Where it breaks is the operations count and the spaghetti. An iterator that loops over 200 line items runs the following modules 200 times, each an operation, so a scenario that looked cheap in testing can burn its monthly allowance in a week. Forty modules with routers inside routers is not something the next person can maintain. Error handling exists but has to be built deliberately, and most people never do. The habits that fix this are in Make.com Automation Guide for Business Owners Who Want It to Last.
4. n8n is the most capable and assumes someone technical is within reach
n8n is the developer-friendly one. It has a code step where a few lines of JavaScript do what would take a dozen modules elsewhere. It has the deepest AI support of the three: purpose-built nodes for talking to language models, giving them tools and memory, and forcing them to answer in a fixed format. And it can be self-hosted, meaning it runs on a server you rent rather than in the vendor’s cloud, with no per-run fee.
Where it breaks is the learning curve and the maintenance. The interface holds your hand less, its catalogue of polished integrations is smaller (though its generic HTTP step can reach almost any product), and if you self-host, someone has to keep the server updated and backed up. n8n is the right call when AI is central, when data is sensitive enough to want on your own server, when you have dozens of workflows, or when you want to own the system. See Self-Host n8n for Your Business Without Being Held Hostage for what ownership takes.
5. Know where the workflow runs and where your data pauses
Every run passes your customer’s name, email, invoice amount, or appointment through the vendor’s servers, and usually leaves it in a log there for a while. Data residency is the question of which country’s servers those are. Zapier runs in the United States. Make offers a choice of regions including the EU. n8n cloud is hosted in the EU (Germany) by default, and self-hosted n8n runs wherever you put the server.
For most US businesses this is not a blocker. For a clinic, a law firm, or anyone whose clients care about confidentiality, it is. Whatever the tool, ask the vendor for a data processing agreement (the contract that says what they do with your data), set log retention so old runs full of customer data are deleted, and if you handle health data, ask in writing whether they will sign a business associate agreement. Do not assume.
6. All three can call an AI model, but they differ in how much control you get
An AI step sends some text to a language model (software like the current Claude models, GPT-class models, or Gemini, which reads text and writes text back) and uses the answer. Zapier has built-in AI steps and an agents product, both easy. Make has modules for OpenAI, Anthropic, and Google. n8n has a full agent node system where the model can choose tools and the output can be forced into a fixed shape your later steps can rely on.
Our view is that the depth matters less than the discipline. A good AI step has one job, a defined input, and a defined output that the next step checks before acting. A bad AI step is asked to “handle the customer” and wired straight to the send button. Any of the three can do the good version. n8n makes it easiest because it can validate the model’s answer before anything downstream uses it, the idea behind LLM Structured Output and Tool Calling, Explained Plainly.
7. Error handling and visibility is where the cheap choice gets expensive
Every workflow will eventually fail, because the systems it talks to will change, go down, or return something unexpected. What matters is whether you find out. Zapier keeps a run history and auto-replays some failures on higher plans. Make has an error-handler concept and a store of incomplete runs, but only if you set it up. n8n has per-step retries, a full execution log, and a dedicated error workflow that can alert someone on every failure.
The practical test: make a workflow fail on purpose (disconnect an app) and see whether anyone gets told. If not, you have a time bomb with a subscription fee. The full set of rules is in Automation Error Handling for Businesses Tired of Silent Failures.
8. Check whether you can leave before you move in
Lock-in is what happens when leaving costs more than staying. Zapier workflows are not portable in any useful way; if you leave, you rebuild. Make scenarios export as a blueprint file, which helps a developer rebuild elsewhere but imports nowhere. n8n workflows are plain JSON files (a standard text format), you can download them all, and because the software can be self-hosted, the vendor cannot switch you off.
Two years in, a business might have sixty workflows running the back office. At that point the tool is infrastructure, and infrastructure you cannot move is a liability. The reasoning in Build vs Buy AI Tools for Your Business, the Decision Rules applies here word for word.
Picture a business like this one
The business below is a composite of the kind of company that writes to us, not a client. The numbers describe the shape of the problem, not a case study.
Picture a business like this one: a regional HVAC contractor with about 35 people and leads arriving from the website, local services ads, and the phone. Two years ago the office manager taught herself Zapier and built fourteen Zaps: new lead to the CRM, new job to the calendar, invoice paid to the spreadsheet. It worked and it was genuinely valuable.
Now the bill is several hundred dollars a month and climbing. Three Zaps fail a few times a week and nobody notices until a customer asks why they never got a confirmation. Nobody remembers what four of them do. The owner wants an AI step that flags emergency calls in the inbound leads, and the quote for doing that on the current plan made him sit down.
What a business like this would build:
- A one-page inventory of the fourteen Zaps: trigger, steps, who depends on each. Five turn out to be dead or duplicated and get switched off.
- The nine that matter rebuilt in n8n on a small rented server, one at a time, with each old Zap left running until the new workflow has run clean for a week.
- One error workflow that catches every failure and texts the office manager the workflow name and the customer involved.
- One AI step in the lead workflow that returns a fixed-format answer (emergency yes or no, service type, confidence), checked by the next step, with low-confidence results routed to a person.
- A short document, kept with the workflows, saying what each one does and how to turn it off.
What changes: the platform bill drops to the server plus a small retainer, failures get seen the same hour, and the dispatcher calls the right people back first. Nothing here is exotic. It is the difference between fourteen things that happen and nine things that are managed.
What it costs to run
Entry plans on all three sit in the range of a modest software subscription; check the current pricing pages because tiers change yearly. A rough rule from our experience: at tens of thousands of steps a month, Zapier is the most expensive by a clear margin, Make is cheap until iterators or frequent polling push the operations count up, and n8n cloud sits in the middle with predictable per-run pricing.
Self-hosted n8n costs whatever the server costs, typically ten to twenty dollars a month for a small virtual server, plus backups and a few hours of maintenance. AI steps add model usage, charged per token (roughly a word fragment; a thousand tokens is about 750 words). A classification step on a cheap model tier costs fractions of a cent per run, so a thousand runs a month is usually less than a cup of coffee.
The mistakes we see most
Picking on features instead of pricing model. Owners pick the longest integration list and discover a year later that their most useful workflow costs more than the person it replaced.
No failure alert. The workflow breaks quietly on a Tuesday and the business finds out from a customer in three weeks.
Treating the AI step as the workflow. The model is asked to “deal with” something and its answer goes straight to a customer with no check between.
When to bring in help
An owner or office manager can absolutely build and run the first five or ten workflows alone on Zapier or Make. The tools are designed for exactly that and the results are real. If your needs are simple, do it yourself and do not let anyone tell you otherwise.
A developer becomes worth it at three points: when the platform bill is large enough that moving would pay for itself within a year, when a workflow touches money, customers, or compliance and needs proper error handling and logs, and when AI steps enter the picture, because checking the model’s output before it acts is a design problem.
Levelbrook builds this kind of automation for businesses, on a fixed price from a written scope, with everything running in accounts and servers you own so you are never held hostage to us or to a platform. If any of the above sounds like where you are, the form below is how a conversation starts.