Make vs n8n vs code: when workflow tools stop being enough

Stay on Make or n8n until one of three things happens: the billing unit starts growing faster than the work, you need a test suite instead of a green checkmark on the last run, or the license stops you doing what you want with your own system. Until one of those lands, a workflow tool is the correct answer and writing custom code is a waste of your money. After one of them lands, the tool is the expensive option and most people keep paying for it anyway.

That is the whole decision. What follows is the arithmetic behind each threshold, using the vendors’ own published numbers as of August 2026.

How the three tools actually bill you

The comparison that matters is not the feature list. It is what one unit of billing buys, because that is what decides your cost at volume. All three vendors define their unit differently, and the difference is larger than the price difference.

	n8n	Make	Zapier
Billing unit	Execution	Credit	Task
What it counts	One run of the whole workflow, regardless of step count or data volume	One action: read, search, create, update, delete, transform, aggregate, iterate	One successfully completed unit of work
Notable exclusions	Extra steps inside the same run cost nothing	Most actions cost 1 credit, some AI provider features cost more	Failed actions, triggers, polling, Filter and Formatter are not billed
Entry paid plan, as listed August 2026	Starter, €20/mo, 2,500 executions	Core, $9/mo at the 10,000 credit tier	Professional, $19.99/mo annual, 750 tasks
Self-hosting	Offered, under the Sustainable Use License	Cloud only	Cloud only

Sources: the n8n, Make and Zapier pricing pages, all read on 10 August 2026.

Run the same job through those three definitions and you get three different bills. A 40 step scenario that fires 2,500 times a month is 2,500 executions on n8n and roughly 100,000 credits on Make, because Make counts every module call. Your Make plan is set by how complicated your workflow is. Your n8n plan is set by how often it runs. Neither is dishonest, but if you scoped your automation on one model and later moved it to the other, your cost estimate is wrong by an order of magnitude.

Threshold one: the billing unit grows faster than the work

The failure looks like this. A workflow that was 6 steps at build time is 30 steps eighteen months later, because every edge case became another branch. On a per action model your bill went up five times while the number of jobs done stayed flat. You are now paying for your own complexity.

Check it with one number: units billed per unit of business value. Credits per order processed. Tasks per lead handled. If that ratio has climbed since launch and the work has not, the tool is charging you for the workaround, not the outcome. Custom code has the opposite shape. The build costs what it costs once, and the marginal cost of the thousandth run is close to zero.

Zapier’s exclusions matter here too. Triggers, polling and the built in Filter and Formatter tools are not billed as tasks, which makes Zapier cheaper than it looks for jobs that filter out most events. That is a genuine advantage and it is worth checking before you assume you need to leave.

Threshold two: you need a test, not a green checkmark

A workflow canvas tells you the last run succeeded. It does not tell you the run will succeed after you change step 14, and it does not run a test before the change goes live. For a job that emails a Slack channel, that is fine. For a job that touches money, inventory or customer records, it is not.

The 2025 DORA report from Google Cloud, which surveyed roughly 5,000 technology professionals, found that AI adoption has a positive relationship with software delivery throughput but a continuing negative relationship with delivery stability. Their framing is that AI amplifies whatever an organization already has, its strengths and its weaknesses both. The same is true of workflow tools. They make a disciplined team faster and an undisciplined one fragile at speed.

The practical test: can you change this automation on a Tuesday afternoon and know before it goes live whether you broke it? If the answer is no and the automation touches something you would have to explain to a customer, you have crossed the threshold. Our own publishing pipeline is a small example. Nothing merges until a build check and a content scan both pass, and a human approves the merge. That is not possible to express as a canvas.

Threshold three: the license decides what you can do with your own system

This one gets skipped and it is the one that bites hardest. n8n is fair-code, not open source under the OSI definition. It ships under the Sustainable Use License v1.0, which grants you use and modification for your own internal business purposes or non commercial use, and allows redistribution only free of charge and only for non commercial purposes. Files marked .ee are excluded entirely and need a separate n8n Enterprise License.

Read that against your plans. Running n8n to automate your own operations is squarely inside the grant. Hosting it for your clients, or embedding it in a product you sell, is not the same activity and needs a different conversation with the vendor. If your automation is going to become part of what you sell, the license is a design constraint, not a legal footnote.

This is why we build client systems on their own infrastructure with code they own outright. Ownership is either total or it is a rental agreement with better branding. We wrote out the full version of that argument in the handover test.

What AI coding agents changed, and what they did not

The build side of this decision got cheaper, fast. GitHub’s Octoverse 2025 report counted 1.1 million public repositories using an LLM SDK, up 178% year over year. The 2025 Stack Overflow Developer Survey found 84% of its 33,662 respondents using or planning to use AI tools, up from 76% the year before, with 51% of professional developers using them daily. DORA 2025 put AI use at work at 90%, with more than 80% reporting a productivity increase.

So the threshold moved. Work that was clearly not worth coding in 2023 is worth coding now, because the hours went down.

What did not move is the reason people chose the tool in the first place. The same Stack Overflow survey found only 3.1% of developers highly trust the accuracy of AI output, 66% report hitting solutions that are almost right but not quite, and 45.2% say debugging AI generated code takes more time.

Faster code generation with no review discipline produces more code you cannot trust, which is worse than a canvas you can at least see. The tooling changed the price. It did not change the requirement for gates, logging and someone who owns the result.

The rule we use on our own builds

Keep it in the workflow tool when the job is internal, low volume, tolerant of a failed run, and stable enough that nobody has touched it in a quarter. Move it to code when it fails any of the three thresholds above, or when it needs real testing, per record cost logic, or a paper trail you would show an auditor. That second case is what an AI Workforce Sprint is scoped around: one defined job, built on your accounts, documented, handed over.

Our own ad spend reconciliation was never a candidate for a canvas. It has to pull every platform claim, every order, and every cost line including shipping, gateway fees and returns, then compute contribution margin per campaign daily. The platform was reporting 4.87x. The real number was between 0.33x and 0.50x.

That gap only shows up when you can join ad data against the actual ledger, per order, with the cost lines the platform never sees. It is written up in full at the ad spend truth build. The same question in its simpler form, whether a job needs a script or a genuine agent, is in AI agent vs. automation.

Most businesses reading this should not rebuild anything today. Go and check one number first: units billed per unit of work, this month against launch month. If it has doubled, you have your answer, and it is not a feature problem.

If it has, describe the workflow and what it costs you now, and you get a written recommendation back within one business day: stay on the tool, restructure the scenario, or build it properly, with a fixed price and fixed scope if the answer is build. No discovery call, because there are no calls. Start async.