Zapier, Make, n8n. Where each one stops.
We build on all three, which means we have no reason to talk you into any of them. The choice comes down to volume, how much logic you need, and whether you want to own uptime.
Why this question is hard to answer honestly.
Almost every comparison of these three is written by someone selling one of them, which is why the honest version is hard to find and why people end up asking on Reddit. The tools are not really competing on features any more — the app catalogues have converged enough that most teams could build most workflows in any of them.
The difference that dominates everything else is what each one charges for. Zapier bills per step, Make bills per module, and n8n bills per workflow run. That sounds like an accounting detail and it is not: a twenty-step workflow costs twenty times as much as a one-step workflow on Zapier, and exactly the same on n8n. Every other cost argument is downstream of that one sentence.
After that it comes down to how gracefully each handles logic that is not a straight line, who is responsible when something breaks at 2am, and — new since 2025 — how each has absorbed AI agents. None of those appear on a feature comparison table.
Zapier.
The largest app catalogue — around 8,000 integrations — and the lowest barrier to a first working automation. Billed per step, where a step is one action on one record.
What it does well
- Widest integration coverage by a distance, including long-tail SaaS nothing else supports
- A non-technical person can genuinely build and maintain simple workflows
- Managed entirely — no infrastructure, no upgrades, no uptime to own
- Agents ship as a first-party product, so AI steps sit inside the same billing and support relationship
Where it stops
- Per-step pricing compounds with workflow length as well as volume: a twenty-step workflow costs twenty times a one-step workflow for the same trigger
- Multi-step logic, branching and loops are workable but awkward, and complex scenarios become hard to reason about
- Limited control over retries and error handling, which matters once a workflow is load-bearing
Who it suits
Lower volume, wide variety of apps, and a team without engineering capacity to spare.
Make.
A visual scenario builder with a genuine flow-control model and roughly 3,000 apps. Billed per module — finer-grained, and generally cheaper per unit of work than Zapier.
What it does well
- Branching, iteration, aggregation and error handlers are first-class rather than bolted on
- Materially cheaper than Zapier at mid volume for equivalent work
- The visual canvas makes a complex scenario legible in a way a linear step list does not
- Maia, its conversational builder, plus an agent builder in beta — the AI story is real rather than a checkbox
Where it stops
- Operations accumulate faster than people expect — an iterator over 100 records is 100 operations, not one
- Steeper learning curve; the flexibility that makes it powerful also makes it easy to build something fragile
- App coverage is good but thinner than Zapier in niche categories
Who it suits
Real branching logic, mid-to-high volume, and someone willing to learn the tool properly.
n8n.
Source-available and self-hostable, with code steps that run real JavaScript or Python. Billed per workflow run on cloud — not per step — or infrastructure cost only if you host it.
What it does well
- Dramatically cheaper at high volume when self-hosted — you pay for a server, not per operation
- Arbitrary code where the visual model runs out, which removes the ceiling entirely
- Data stays in your infrastructure, which resolves conversations that stall on compliance grounds
- Version 2.0 (January 2026) added native LangChain support and around 70 AI nodes, so agent work does not need a separate platform
Where it stops
- Self-hosting means you own uptime, upgrades, backups and security patching — a real operational commitment
- Fewer prebuilt integrations, so more workflows involve talking to an API directly
- The licence is fair-code rather than open source; worth reading if you intend to embed it in something you sell
Who it suits
High volume, engineering capability in-house, or data-residency requirements that rule out managed tools.
How we actually decide.
The question we ask first is not which tool, but how many operations a month the workflow will actually generate at steady state. That single number eliminates at least one of the three immediately, and it is usually the number nobody has estimated.
The second question is whether the logic is linear. If it genuinely branches, loops, or needs to aggregate before acting, Zapier will do it and you will regret how it reads in six months. If it is a straight line, Zapier's simplicity is a real advantage and paying more for it can be entirely rational.
The third is whether anyone will own the infrastructure. Self-hosted n8n is the cheapest option on paper and the most expensive if nobody is responsible for it. We have inherited more than one instance that stopped running weeks before anybody noticed.
Since 2025 there is a fourth question: whether the workflow needs an AI agent making decisions rather than a rule following them. All three now offer this — Zapier Agents, Make with Maia, n8n 2.0 with LangChain and roughly seventy AI nodes — and the honest position is that none is clearly ahead yet. If that is the requirement, prototype in whichever you already run rather than switching platform for it.
In practice a lot of businesses end up with two: something managed for the long tail of simple connections, and something cheaper or self-hosted for the high-volume path. That is a legitimate outcome rather than a failure to decide.
Questions people actually ask.
Self-hosted, at volume, by a wide margin — you are paying for a server rather than per task. At low volume the saving is negligible and the operational burden is not, so the honest answer is that it depends entirely on how many operations you run and whether someone will maintain it.
Almost always because a task is counted per action per record, not per workflow run. A single automation touching five systems for every order is five tasks per order. Multiply by monthly volume and the number stops being small quickly.
Logically, generally yes, and often more elegantly. The gap is app coverage: Zapier supports more niche tools. If your stack is mainstream, that gap rarely matters. If you depend on one obscure SaaS product, check it is supported before committing.
Cloud unless you have a specific reason not to. The reasons that genuinely justify self-hosting are data residency, very high volume, or an existing platform team who will treat it as a production service. Cost alone is usually not sufficient once you price the maintenance.
n8n, usually by a wide margin, because it bills per workflow run rather than per step. A twenty-step automation costs the same as a one-step automation there, and twenty times as much on Zapier. Make sits between the two, billing per module. If your workflows are long rather than merely frequent, that single difference will dominate your bill.
There is no honest winner yet. Zapier ships Agents as a first-party product, Make has Maia plus an agent builder in beta, and n8n 2.0 added native LangChain with around seventy AI nodes. All three are moving monthly. Our advice is to prototype on whichever platform you already run, because switching platform for the agent feature alone rarely survives contact with the rest of your workflows.
We build on all three and pick per workload rather than per client. Give us the volume estimate, the shape of the logic, and whether anyone owns infrastructure, and the answer is usually obvious within an hour. Anyone who recommends the same tool every time is telling you about their business, not yours.