Task Automation: The Three Levels and Which One You Need

Most task automation advice fails because it starts with a tool. The tool decides what you can build, and you end up bending a real task into a shape the tool likes.

Start from the task instead. Every automatic task lives at one of three levels, and picking the wrong level is why so many setups get abandoned:

• Device level — it happens on one machine. File renaming, backups, scheduled scripts, phone shortcuts. Free, fast, no accounts.
• App level — it moves information between services. New form response creates a row creates a notification. This is what people usually mean by automation.
• AI level — the task needs reading, judgment, or writing before anything can move. Sort these emails by what they're actually asking for.

Below: how to place a task on that ladder, what to use at each rung, and the four tasks you should refuse to automate.

Quick comparison

Cost climbs as you go down. So does the chance of a wrong answer. Always solve a task at the lowest level that can hold it.

Level one: automatic tasks on your own machine

The most underused tier, because it isn't sold to anyone.

• Apple Shortcuts — the same automations run on iPhone and Mac, and they can be triggered by time, location, arriving messages, or opening an app. Genuinely powerful and genuinely under-explored. We've collected practical ones in iPhone automations.
• Windows Task Scheduler — runs anything on a schedule or a system event. Ugly, ancient, completely reliable.
• cron — the Unix scheduler, still the correct answer for "run this script every night" on macOS and Linux.
• Folder actions and watchers — do something the moment a file appears in a directory. The quiet workhorse of document processing.

If your task is "every Monday, take these files, rename them, and put them somewhere," you do not need a cloud platform, an account, or a subscription. You need fifteen minutes and one of the above.

The single limitation: it runs on one machine, and only when that machine is awake. That's the entire reason level two exists.

Level two: moving things between apps

This is the classic shape. Something happens in one service; something should happen in another.

• Zapier — widest app coverage, lowest learning curve. If both your apps are supported, this is a fifteen-minute job.
• Make — a visual canvas that handles branching and iteration properly, which is where linear tools get awkward.
• n8n — self-hostable with code fallback. More setup, more control, the right call when data can't leave your infrastructure.
• Power Automate — the default inside a Microsoft 365 estate.

The mistake people make here is building the ambitious version first. A twelve-step flow with three branches, built before anything ran in production, will fail in a way you can't diagnose. Build two steps. Run it for a week. Add the third.

The other thing worth knowing before you commit: these platforms charge on different units — per task, per operation, per execution, per seat. A flow that polls every five minutes is nearly free on one model and expensive on another. Our comparison in workflow AI breaks down which is which.

Level three: when the task needs judgment

Move up to AI only when a rule genuinely can't express the task. The test: can you write down the condition? If yes, it's a rule, and a rule is cheaper, faster, and won't be wrong.

Tasks that legitimately need this level:

• Reading unstructured input. Invoices in twelve formats, PDFs, forwarded email threads.
• Classifying by meaning. "Is this ticket a bug report, a feature request, or a billing question?"
• Extracting into structure. Turning a document into fields. The highest-value AI automation for most people, and the least exciting.
• Drafting from real context. A reply that references the actual thread, not a template.

Tasks that don't: anything with a threshold, a keyword, a date, or a status field. Those are rules wearing a costume, and running them through a model makes them slower, costlier, and occasionally wrong.

The reliable pattern in 2026 is a hybrid: rules handle the deterministic parts, one AI step handles the messy part, and rules take over again for the actions. Keep the model as far from the irreversible operations as possible.

The four tasks not to automate

Automation advice never covers this and it's where most wasted effort goes.

1. Things you do rarely. Twice a year is not a candidate, whatever the time saving looks like on paper. You'll spend longer building it than doing it, and it'll be broken by the time you need it again.
2. Things you haven't done manually enough to understand. You'll automate the wrong version. Do it by hand ten times first — the exceptions only show up in the messy runs.
3. Things where being wrong is expensive and silent. Payments, permissions, anything sending to customers. Automate the preparation and keep a human on the trigger.
4. Things that are actually a broken process. If a task exists because two teams don't share a system, automating it makes the broken thing permanent and harder to see.

Point four is the one that costs the most, because it feels like progress.

Getting started this week

Pick one task with all four traits: you do it weekly or more, it takes under fifteen minutes, it needs no judgment, and being wrong is cheap. Automate exactly that. Not the process around it — that one task.

Then leave it alone for two weeks and see what breaks. What you learn from one real automation failing in production is worth more than any amount of planning, and it's the thing that tells you whether your next task belongs at level one, two, or three.

If where you keep stalling is running someone else's setup — you've found the script, the repo, or the workflow, and can't get it going on your machine — that's a different problem from choosing a level. Taku mirrors a working AI setup into your own desktop workspace and runs it there, so you can use a power user's configuration without reproducing their environment first. The free app library is a fast way to check whether your task already has a shape someone published. Taku is in Beta, and the Mac app is available now.

FAQ

What is task automation?

Setting up software to perform a task without you doing it each time. It runs at three levels: on your own device, between cloud apps, or through an AI model when the task needs reading or judgment.

What's the easiest way to create an automatic task?

On a phone or Mac, Apple Shortcuts. On Windows, Task Scheduler. On Linux or macOS for scripts, cron. All three are free and handle scheduled or triggered local tasks with no account.

When should I use AI for task automation?

Only when you can't write the rule down. Unstructured input, fuzzy classification, extraction, and drafting need a model. Thresholds, keywords, and status changes don't.

How do I stop automations from breaking?

Start with two steps, run them in production for a week, then add more. Most breakage comes from complex flows built before anything simple was proven, and from source systems changing their APIs.

What shouldn't I automate?

Rare tasks, tasks you haven't done by hand enough to understand, tasks where an error is expensive and invisible, and tasks that only exist because a process is broken upstream.

Key points

• Place the task first, pick the tool second — the lowest level that holds it is the right one.
• Device-level automation is free, fast, and badly underused.
• App-level platforms differ by billing unit more than by feature; a polling flow exposes the difference.
• Use AI only where you can't write the rule down; keep it away from irreversible actions.
• Refuse to automate rare tasks, misunderstood tasks, silent-failure tasks, and broken processes.