n8n vs Zapier: what running close to a hundred workflows taught us
We run close to a hundred active automations on n8n. Here is where n8n wins, where Zapier is the better choice, and the questions that decide it for a small company.
- Zapier is the faster start: thousands of ready-made app connectors and almost nothing to maintain. You pay per task, which means per action step.
- n8n gives you the whole workflow: code steps, webhooks, AI agent nodes and workflows saved as JSON that you can keep in git. You pay per run, or you host it yourself.
- Choose Zapier for a handful of simple, low-volume connections between mainstream apps. Choose n8n when the automation is part of how the business runs and has to be versioned, tested and reviewed.
Flowsmith runs on n8n. As of October 2026 our n8n Cloud instance has 97 active workflows: inbox triage, drafting replies for review, bank feeds posted into the accounting system, backup watchdogs, CRM imports, image generation for this blog, and a few dozen small helpers. We did not pick n8n because it is fashionable. We picked it because of what happened once automations stopped being side projects and became part of daily operations.
This is not a feature-by-feature table copied from two pricing pages. It covers what actually decides the choice, and it is honest about where Zapier is the better tool.
The real difference: tasks versus runs
The two tools count work differently, and the counting changes how you build.
- Zapier bills by task. Roughly speaking, every action step that runs successfully is one task. A five-step Zap that fires 1,000 times a month uses about 5,000 tasks.
- n8n Cloud bills by execution. One run of a workflow counts once, however many nodes it passes through. Self-hosted n8n (the Community Edition, under n8n's Sustainable Use Licence) has no execution limit at all. You pay for the server and your time instead.
In practice, per-task billing quietly pushes you towards short, shallow automations. Per-run billing lets you add the steps that make an automation safe: validation, deduplication, logging, a check before anything is sent. Our sales-reply flow runs through dozens of nodes because each step does one checkable thing. On a per-step meter we would have been tempted to cut exactly the steps that protect us.
Prices and plan names change often, so check both vendors' current pricing pages before you decide. The billing model has stayed the same for years, and the model is what shapes your workflows.
Where Zapier is the better choice
We would still recommend Zapier in these cases:
- Breadth of connectors. Zapier connects to thousands of apps out of the box, many of them niche SaaS tools that n8n reaches only through a generic HTTP request. If your stack is mostly mainstream tools and you never want to read API documentation, that matters.
- Nobody to maintain it. If the person building automations is not technical and there is no one to look after a workflow when an API changes, Zapier's polish and support are worth paying for.
- A handful of simple connections. "New form entry → add a row → post to Slack" at a few hundred runs a month is cheap and quick in Zapier. Moving that to n8n saves little.
Where n8n wins, from our own operations
1. Workflows are files you can version
Every n8n workflow exports as JSON. Ours live in a git repository next to the website and the forms gateway, with a changelog. When a workflow misbehaves we can see what changed and when, and roll back. This sounds like an engineering nicety until the day an automation sends something it should not have. After that, a history is the first thing you want.
2. Code where you need it, nodes where you don't
Most steps are ordinary nodes: Gmail, Google Drive, HTTP requests, the database. When the logic gets specific, such as deduplicating customer cases by thread and sender or refusing to act on a draft that did not come through our own flow, a short Code node says it exactly. Zapier has code steps too, but n8n treats code as an ordinary node, which matters once your rules get detailed.
3. Webhooks that agents can call
Many of our workflows begin with a webhook, so other systems and AI agents can call them like a small API. The illustration at the top of this article was made that way: an agent called an n8n webhook that generates images through fal.ai, and a second webhook committed the files to a branch for review. The agent never handles an API key. The key stays in an n8n credential. That boundary is easy to build in n8n, and it is the pattern we teach in the Academy lesson Tools, not copy-pasted JSON.
4. Human review is part of the workflow
Our replies to sales enquiries are drafted by a model and approved by a person before they are sent. The review screen is itself built from n8n webhooks that return a simple HTML page. With a per-task meter, iterating a draft three times with a reviewer would be the expensive path. With n8n it is simply the correct path. We describe the rules behind it in Human-in-the-loop automation: three rules.
5. You choose where it runs
We use n8n Cloud. It is managed, and we have no wish to patch servers. The option to self-host is still valuable. It caps cost at high volume, and it answers the data-residency question some customers ask before they will let an automation read their inbox.
What n8n costs you that the pricing page doesn't mention
The trade-off is real, and it is easy to underestimate:
- You own the failure modes. We once had an "email sender" workflow that had been created but contained zero nodes, so approved replies were never actually sent, and nothing reported an error. We now treat "it ran" and "it did the job" as two separate checks. The pattern is common enough that it has its own Academy lesson: When failure looks like success.
- Back up outside the tool. A backup of n8n must not run inside n8n. If the instance is down, so is your backup. Ours runs from a separate scheduled job and is checked by its own watchdog.
- Errors need somewhere to go. We use one shared error workflow that notifies a person, with a 30-minute cooldown so a single broken API does not produce a hundred emails.
How to decide in ten minutes
| Question | Leans Zapier | Leans n8n |
|---|---|---|
| How many automations in a year? | Under ten, simple | Dozens, some complex |
| Steps per automation | Two to four | Many, with checks and branches |
| Who maintains it? | Nobody technical | Someone, or a partner like us |
| Does anything leave the company automatically? | Rarely | Yes, so it needs review and a history |
| Will AI agents call it? | No | Yes, through webhooks or MCP tools |
| Data residency matters? | No | Yes, so self-host or choose the region |
If most answers fall in the right-hand column, start with n8n and plan for its maintenance from day one. If they fall on the left, Zapier will serve you well, and you can move later. A clear Zap is easy to rebuild as an n8n workflow. A messy one is hard to rebuild anywhere.
Starting without regret
Whatever you choose, begin with the shape every automation shares: trigger, brain, action. Build one flow end to end, test it the way the person on the other side experiences it, and only then add branches. The Academy lesson Your first automation walks through exactly that in n8n, and Models, limits & cost helps you keep the AI step affordable.
Want this running in your company?
Flowsmith designs and operates automations like this for small and mid-sized teams — with a person approving anything that leaves the building.
