n8n, Make or Zapier: choosing by the constraint that will actually bind

Every comparison of these tools published by the vendors counts integrations. Nobody has ever chosen wrong because a connector was missing — you can call an HTTP endpoint from all three. Teams choose wrong because they pick on the demo and get caught by one of three constraints six months later.

The pricing shape decides more than the price

Zapier bills per task, where a task is roughly one action executed. A workflow with five action steps that runs a thousand times a month is five thousand tasks. Filters that stop early are free, which quietly makes filter placement a cost decision.

Make bills per operation, and an operation is finer-grained than a Zapier task — an iterator producing forty items costs forty operations downstream. Make is usually cheaper per unit and easier to make expensive by accident, because a single scenario that fans out over a list can consume a month's quota in one bad run.

n8n self-hosted bills nothing per execution. You pay for the VM and your own time. n8n Cloud bills per execution — the whole workflow, however many nodes — which is the most forgiving shape of the three for multi-step work.

The practical consequence: high-volume, many-step, low-complexity work is dramatically cheaper on n8n. Low-volume work spread across many teams is cheaper on Zapier, because you are not paying anyone to run it.

Where the data sits, and whether that is a conversation

If your automation moves customer records, health data or anything covered by a DPA someone had to negotiate, the question is not "is the vendor compliant" — all three are, on paper, with the right plan. The question is whether payloads transit a third-party processor at all, because that is the sentence you have to write in the register of processing activities and defend in an audit.

Self-hosted n8n inside your own VPC answers it by removing the processor. That is the single most common reason we see European companies move off Make, and it is usually driven by legal, not engineering. For a team already running on a cloud account with a landing zone, adding one more internal service is cheap; for a five-person company it is not worth it.

What happens when the workflow needs real code

Every automation of any value eventually needs a piece of logic that does not fit in a node: a retry with backoff, a signature to compute, a payload to reshape in a way the mapper cannot express.

Zapier gives you a sandboxed Code step with a short timeout and no package installs. Make gives you less, and pushes you toward chaining more modules, which costs more operations. n8n gives you a Code node with real Node.js and, self-hosted, the ability to install npm modules on the instance.

That gradient decides the ceiling. Teams with engineers hit Zapier's ceiling fast and resent it. Teams without engineers hit n8n's floor — the instance needs an owner — and resent that instead.

The honest recommendation

Zapier when the users are non-technical, the volume is modest, and the value is spreading automation across departments without IT. Its reliability and its connector depth are genuinely the best of the three, and you are paying for not having an operator.

Make when the work is visual, branching and data-shaping heavy, the team likes the canvas, and volumes are predictable. Watch the operation counter on anything that iterates.

n8n, self-hosted when volume is high, engineers exist, data residency is a real constraint, or you are building LLM steps into workflows and want to control the model, the prompt and the cost per run. It is also the only one of the three where the workflow is a file you can review in a pull request — see self-hosting n8n in production for what owning it actually involves.

The migration between them is not free but it is not terrible either: the connectors differ, the logic does not. Budget a day per non-trivial workflow.

What to do this week

Export your current task or operation usage for the last three months and group it by workflow. In almost every account we have looked at, two or three workflows account for more than seventy percent of the consumption. Those are the only ones whose pricing shape matters; the rest of the catalogue can stay wherever it is.