n8n vs Make vs Zapier vs custom code: how to pick an automation backbone
I run AmazeAI, where we build workflow automation for service businesses, so I have a stake in this. This is a short version of a longer piece with sources, linked at the bottom. The pricing units below were checked agai
I run AmazeAI, where we build workflow automation for service businesses, so I have a stake in this. This is a short version of a longer piece with sources, linked at the bottom. The pricing units below were checked against each vendor's pricing and help pages on November 15, 2026.
The short answer
Zapier fits simple, low-volume workflows owned by someone who is not an engineer. Make fits visual, branching workflows at moderate volume. n8n fits teams that want code steps, self-hosting and pricing per run rather than per step. Custom code earns its cost where a failed run loses real money. Choose on who will maintain it and what a failure costs, not on the monthly fee.
How the four compare
| Zapier | Make | n8n | Custom code | |
|---|---|---|---|---|
| Billed per | Task: each successful action step | Each module run, counted in credits (1 operation is 1 credit by default) | Execution: each full workflow run on Cloud. The self-hosted Community Edition is free to run | Hosting plus engineering time |
| Where it runs | Vendor cloud | Vendor cloud | Vendor cloud or your own servers | Wherever you deploy it |
| Logic | Filters and paths, best for straight lines | Routers, iterators and aggregators, visually | Branches, loops and sub-workflows | Anything you can write |
| Code steps | JavaScript or Python steps | Possible, but built around visual modules | JavaScript and Python nodes as normal steps | It is all code |
| Who maintains it | Ops or marketing | A technical operator | A technical operator or a developer | Engineers |
| Where it hurts | Cost at volume, complex logic | Large scenarios get hard to follow | Self-hosted means you own uptime and upgrades | Build cost, and it needs an owner who codes |
Why the pricing unit matters more than the monthly fee
Take a workflow with a trigger and five actions: a new inquiry arrives, gets validated, enriched, written to the CRM, booked into the calendar and confirmed by text. Run it 3,000 times a month. This is an illustrative example, not a quote, and the figures are billing units, not prices:
- Zapier: about 15,000 tasks, because each successful action step counts.
- Make: about 18,000 operations, because each module run counts.
- n8n Cloud: 3,000 executions, because each full workflow run counts once.
- Custom code: hosting only.
Add a validation step and two of the three bills grow by 3,000 units. That is the behavior to watch. When every extra step costs money, teams cut validation and error handling to save units, and those are the steps that keep a workflow from doing damage.
When Zapier is the right answer
More often than automation builders admit. It fits when the workflow is short and linear and runs a few hundred times a month, the person who will fix it works in operations rather than engineering, a failed run costs a spreadsheet row rather than a booking, and you are still learning what the process should be.
When custom code earns its cost
When a failed run loses money, as with a booking write, a payment or a consent check. When data such as patient records cannot leave your infrastructure. Or when the monthly bill for tasks or credits outgrows what an engineer's time would cost. Most production systems end up as hybrids: the workflow engine orchestrates, and the money-losing steps live in code with tests, version control and someone who reviews changes.
Where the AI steps go
Zapier, Make and n8n can all call a language model as a step. The design question is where its output is checked before it changes anything. A model is good at reading a messy inquiry and poor at deciding whether someone may be called or which slot is free. Those are rules, and rules belong in code or in a step that cannot be talked out of them.
Three mistakes that show up a year later
- Keeping state inside the tool. Lists, flags and counters stored there become a second database no one can query. Keep state in your CRM or a real database, and moving tools later is a rebuild of steps, not a migration.
- Choosing on the monthly fee. The cost that matters is the hours spent fixing it and what a failed run loses.
- Skipping error handling to save units. An automation with no alert fails silently.
A five-minute audit
Write down each automation you run, with its monthly runs and the person who fixes it. Next to each, write what one failed run costs: nothing, an hour, a customer. Mark which ones alert a person when they fail. The unmarked ones with a cost are your first job.
The full article, with sources and the answers to common questions such as whether you can start on Zapier and move later: https://amazeai.co/signal/n8n-vs-make-vs-zapier-vs-custom/
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ full credit and traffic to the original publisher.