Most guides on how to automate work with AI tools start by picking a platform. That is the wrong end. The platform matters far less than two things nobody puts in step one: which tasks are actually worth automating, and how the platform you choose counts what you use — because that counting method, not the sticker price, is what your bill is made of.
What follows is the order of operations that works, with the arithmetic you need to avoid a surprise invoice and the rule for where AI belongs inside an automation. Pricing checked 20 August 2026.
Step 1: find the tasks worth automating
Keep a log for one week. Every time you do something repetitive, write down the task, how many minutes it took, and what would happen if it were done wrong. At the end of the week you will have a list, and three columns tell you what to do with it.
Frequency × minutes × cost of an error. High frequency and low error cost is where automation pays. High error cost belongs to a human no matter how often it recurs. Something you do twice a month for four minutes is not worth two hours of building, however satisfying the build would be.
Be honest about the second number. People consistently overestimate how long routine tasks take, and a workflow that saves ninety seconds twice a day saves you about thirteen hours a year — real, but not the transformation the marketing implies.
Step 2: check what you already own
Before subscribing to anything, look at the apps you already pay for. Email clients have rules and templates. Spreadsheets have scheduled scripts. CRMs have built-in workflow builders. Meeting platforms have summaries. Accounting tools have recurring invoices and reminders.
A large share of what people build in an automation platform duplicates something included in a subscription they already hold. That costs nothing extra, breaks less often, and keeps the data inside a vendor you have already vetted. The same audit logic applies here as in AI tools that replace expensive software — the cheapest tool is usually the one already on the invoice.
Step 3: understand how each platform counts
This is the section that saves money. The three main platforms bill for completely different units, and the difference compounds with workflow complexity.
| Platform | Billing unit | Entry pricing (Aug 2026) |
|---|---|---|
| Zapier | One task per action step — the trigger is free, every action costs | Free 100 tasks/month, single-step only. Professional from about $19.99/month annually (roughly $29.99 monthly) for 750 tasks |
| Make | One operation per module run — similar per-step counting, much cheaper per unit | Free 1,000 operations/month with multi-step scenarios included; paid tiers start around $9–11/month |
| n8n | One execution per whole workflow run — a twenty-node workflow costs the same as a one-node workflow | Cloud from roughly $20/month; self-hosted has no execution limit, at the cost of running a server and maintaining it |
Work through one example. A workflow fires on a new payment, looks up the customer record, posts to a team channel, and appends a row to a spreadsheet. Run it a thousand times a month and Zapier counts roughly three thousand tasks, Make counts a similar number of operations at a much lower unit price, and n8n counts one thousand executions regardless of how many steps sit inside.
The practical rules that follow:
- Few steps, low volume, non-technical team: Zapier. The app library and the interface are worth the premium under a few thousand tasks a month.
- Many steps, or volume above a few thousand runs: Make. Per-step counting still applies but the unit cost is far lower.
- High volume, long workflows, or data that cannot leave your infrastructure: n8n, especially self-hosted. Per-execution billing means complexity stops costing extra.
- Everyone: buy annually if you commit — monthly billing typically runs around a third higher — and set a usage alert at 80% of your limit. Unused tasks do not roll over.
Step 4: put the AI in one place only
Here is the mistake that ruins most first attempts to automate work with AI: using a language model for the plumbing.
An automation has two kinds of step. Deterministic steps — move this, copy that, if the amount is over 500 then route here — should be handled by filters, formatters and conditions. They are free, instant, and they behave identically every time. Judgement steps — is this email a complaint, what is the invoice number in this PDF, summarise this thread — are where a model earns its place.
| Step | Use |
|---|---|
| Route by amount, sender, date or keyword | A filter. Never a model |
| Reformat a date, split a name, build a URL | A formatter step |
| Classify free text into your own categories | A model, with a fixed list of allowed outputs |
| Pull structured fields out of an unstructured document | A model, with the output validated before use |
| Draft a reply or a summary | A model, routed to a human for approval |
| Decide whether to send, pay, delete or publish | A person |
Two disciplines make model steps behave. Constrain the output — ask for one value from a fixed list, or a specific JSON shape, never free prose that a later step has to interpret. And validate before acting: if the model returns something outside the allowed set, route it to a human rather than letting the workflow continue on a guess.
Step 5: build it in draft mode first
Build the whole workflow but replace the final action with a notification to yourself. For one to two weeks, the automation runs and tells you what it would have done. You count how often you would have had to intervene.
This costs you two weeks and saves you the class of failure where an automation quietly does the wrong thing forty times before anyone notices. It also gives you the number that justifies the subscription, which you will want when deciding whether to keep paying.
Only when the intervention rate is low enough do you swap the notification for the real action. And keep the notification for anything irreversible — the point made in what AI agents can and can’t do holds here too: the tasks worth automating are the ones where a mistake is cheap to spot and cheap to undo.
Step 6: give every workflow an owner and a review date
Automations decay. App APIs change, a field gets renamed, a permission expires, and the workflow either fails loudly or — worse — succeeds while doing nothing useful.
- Keep a simple inventory: what each workflow does, why it exists, what it depends on, who owns it, and when it was last verified.
- Review error logs weekly for the first month after building, then monthly.
- Set a consumption alert at 80% of your plan limit so you can optimise before you hit the ceiling rather than after.
- Delete workflows nobody can explain. An unowned automation is a liability, not an asset.
For small teams this inventory can be a single spreadsheet. It takes ten minutes to set up and it is the difference between an automation stack and a haunted house.
What not to automate
- Anything a person’s judgement is the product of. If a client is paying for your assessment, automating the assessment is selling them something else.
- Processes that change monthly. Maintenance will cost more than the task did.
- Irreversible actions without approval. Sending, paying, deleting, publishing.
- Anything you cannot describe precisely. If you cannot write down the rule, you cannot automate it — you can only approximate it, and you will not notice when the approximation drifts.
- Low-frequency work. The build cost is real and it is usually larger than you estimate.
A realistic first month
If you want to automate work with AI tools and actually keep the result running, do less than you plan to. Pick one task from your log. Check whether an app you already pay for does it. If not, build it on a free tier — 100 tasks on Zapier or 1,000 operations on Make is enough to prove a concept. Run it in draft mode for two weeks. Count interventions. Then either promote it to live, or delete it and pick the next task.
One working automation beats five half-built ones, and the discipline of finishing before starting again is most of what separates people who save real time from people who accumulate subscriptions. For where automation fits in a wider stack, see best AI tools for small business owners and AI tools for solopreneurs; for the cost side, cutting business costs with AI and free vs paid AI tools. If your first candidate is customer messaging, start with building an AI chatbot without code or a no-code support system rather than a general automation platform.
Frequently asked questions
What is the cheapest way to automate work with AI tools?
Usually the features already included in software you pay for — email rules, CRM workflows, spreadsheet scripts. After that, free tiers: Zapier’s 100 tasks a month or Make’s 1,000 operations. Self-hosted n8n costs only server time but requires someone technical to run it.
Why is my Zapier bill higher than expected?
Because Zapier counts every action step as a separate task. A four-step workflow run a thousand times consumes roughly three thousand tasks, not one thousand. Unused tasks also expire each cycle, so plan for peak rather than average usage.
Should I use Zapier, Make, or n8n?
Zapier for non-technical teams with simple, lower-volume workflows. Make when workflows get long or volume climbs, since the per-unit cost is much lower. n8n when you have technical help, high volume, or data that cannot leave your infrastructure — it bills per workflow run rather than per step.
Where should AI actually sit in an automation?
Only at the judgement steps — classifying text, extracting fields from documents, drafting something for review. Routing, filtering and formatting should stay deterministic, because those steps are free, instant and identical every time.
How do I stop an automation from doing something wrong?
Run it in draft mode first: build the full workflow but replace the final action with a notification to yourself for one to two weeks, and count how often you would have intervened. Keep human approval permanently on anything irreversible.
How much time does automation actually save?
Less than the marketing suggests and more than nothing. A task taking ninety seconds twice a day is roughly thirteen hours a year. The wins come from stacking several such tasks, not from one dramatic workflow.
Sources
- Zapier — official pricing and plan limits
- Make — official pricing
- n8n — official pricing
- n8n documentation — self-hosting and execution model
Pricing and plan limits checked 20 August 2026. All three platforms adjust tiers and unit allowances regularly — confirm on the official pages before subscribing.
6 comments