Pick One Workflow to Automate First
Score five candidate tasks against four tests, then pick a platform by how it bills: Zapier counts actions, n8n counts runs.
Assumes you've done: Map the Work Before You Automate It. Advisory, not a gate — this page stands on its own.
You have a ranked list of tasks. Now you pick one. Exactly one, because the failure pattern in small businesses is not picking badly, it is picking six things at once and finishing none of them while the trial expires.
Two things decide the pick. Whether the task survives four tests, and whether the platform charges you in units that match how the task actually runs.
Score your candidate on four things you can check today
Take the top five rows from your inventory and put each through this. A candidate has to pass all four. Three out of four is a task you automate later, once you have one working system and some confidence.
| Test | Passes when | Fails when |
|---|---|---|
| Machine-detectable trigger | An email arrives, a form is submitted, a clock hits 9am, a row is added to a sheet | The trigger is you noticing something, or a customer phoning |
| Same steps every time | You could write the steps down and a temp could follow them | Each instance needs a judgement call about what to do next |
| Cheap to get wrong | A bad result creates a duplicate row or an internal note | A bad result sends a price, a refund or a legal commitment to a customer |
| You can tell it worked | There is a place you can look each week and see the output | Success is invisible until something goes wrong months later |
The fourth test is the one people leave out and regret. An automation you cannot inspect is an automation that has been quietly broken since April.
Worked example. "Reply to stock enquiries" passes the first, second and fourth tests and fails the third, because the reply goes to a customer. Downgrade it: the automation drafts the reply and files it, and you press send. Now it passes all four. That downgrade move works on most tasks and it is the single most useful habit in this whole series.
Zapier bills per successful action, n8n bills per run
This is where the money goes, and the two dominant no-code platforms count completely different things. Get this wrong and a workflow that looked like it fit the free tier eats a month's allowance in four days.
Zapier's own help documentation is blunt about the unit: "A task is any successful action that runs in Zapier. Only successful actions count toward your task usage." The list of what does not count is longer than people expect, and it is what makes the free tier usable. Per that page, all trigger steps are free, so is any Filter or Paths step, so are action steps that error or halt, so are steps skipped because a filter stopped them, and so are Formatter, Delay, Looping, Digest, Storage, Zapier Manager, and Tables and Forms steps.
Some steps cost more than one. The same page lists Lead Router at 5 tasks per lead, Zapier MCP at 2 tasks per tool call, Code by Zapier at 1 or more, and AI by Zapier as varying by model tier. If your first automation is a three-step Zap where only one step is a real action, you are spending one task per run, not three.
n8n counts differently. Its documentation defines the unit in one sentence: "An execution is a single run of a workflow." Not per action, per run. A workflow with fifteen nodes and a workflow with two nodes both burn one execution. The same page adds the detail that saves you during development: "Only production executions count towards this quota. These are executions started automatically by triggers, schedules, or polling." Runs you fire manually from the editor while building do not count.
| Zapier | n8n | |
|---|---|---|
| Billing unit | Successful action step | One workflow run |
| Triggers cost | Nothing | Included in the run |
| Failed steps cost | Nothing | Still a run |
| Test runs while building | Count if the action succeeds | Manual editor runs excluded |
| Cheap shape | Few action steps, many filters | Fewer, fatter workflows |
The practical consequence: on Zapier you make workflows narrow, filtering hard before you reach an action. On n8n you can pile logic into one workflow without the bill moving, and you are instead watching how often it fires.
What the free tier actually holds
Zapier's pricing page lists the Free plan at $0 per month with 100 tasks, unlimited Zap workflows, Tables and Forms, two-step Zap workflows, Copilot on a daily limit, and basic access to its AI products. Professional starts at $19.99 per month billed annually at the 750-task tier, or $29.99 billed monthly; Team starts at $69.00 per month billed annually at the 2,000-task tier. Those headline figures are tier-specific, because the page prices each plan against a task slider, so "$19.99" is the entry point rather than a flat plan price.
What happens when you run out is worth knowing before it happens. Zapier's help page states: "When your plan reaches its task limit Zapier will hold any new Zap runs until your billing cycle resets." Held, not queued and delivered late. If you switch on pay-per-task billing instead, the same page describes email notifications at 80% and 100% of the pay-per-task limit, which it puts at three times your plan's base limit, with tasks continuing to run and being charged separately.
100 tasks a month sounds tiny until you apply the counting rules. A daily automation with one real action step is roughly 30 tasks a month. Three of those fit. A per-email automation on an inbox getting 60 enquiries a month does not.
One gap to name: n8n's documentation defines executions and confirms paid plans carry an execution quota, but the quota numbers themselves are not on the documentation page. They live on the commercial pricing page, which changes independently of the docs. Check the number the day you sign up rather than trusting any figure you read in an article, including this one.
Build the smallest version that can fail safely
Whichever platform you pick, the first build is deliberately underpowered.
- One trigger, one action, no branches. If your chosen task needs a condition, hardcode it for now. Branches are where a beginner workflow becomes unreadable.
- Send the output to yourself. Not to the customer, not to the shared inbox. Your own address, or a spreadsheet row. For the first two weeks the automation's only job is to prove it fires at the right moment with the right data.
- Run it manually before you schedule it. On n8n this is free by documented policy, since manual editor runs are excluded from the quota. On Zapier a successful test action does consume a task, so budget a handful of them.
- Write one line in your inventory sheet. Date switched on, what it does, where the output lands. In six months this is the only record of why the thing exists.
- Look at it every Friday for a month. If you have not looked at it in four weeks and it is still sending things, you have an unmonitored process, which is worse than the manual job you replaced.
Only after a full month of correct output do you point it at a real destination. That month is not caution for its own sake. It is the cheapest way to find out that your trigger also fires on the automated receipt from your own accounting system, which is the kind of thing nobody predicts and everybody discovers.
The candidate you should not pick first
Two tasks look ideal and are traps.
The first is anything that talks to a customer without you in the loop. It scores highest on time saved and it is where every horror story comes from. Downgrade it to draft-and-review, as above, and it becomes a fine second project.
The second is the task you find most annoying but do rarely. Annoyance is a terrible proxy for value. A quarterly job you hate is four instances a year. A daily two-minute job you barely notice is around 500 minutes. The inventory already told you this. The trap is that the ranked list says one thing and your gut says another, and the gut usually wins unless you have written the numbers down.
If none of your top five passes all four tests, that is a real result, not a failure of the exercise. It means the work in your business is triggered by people rather than by systems, and the fix is upstream: a booking form instead of phone calls, a shared inbox instead of your personal one. Software cannot hook onto a trigger that only exists in someone's head.
Changelog (1)
- July 31, 2026 — First published.