Choosing What to Automate (and What Not To)

The biggest automation mistake is not technical. It is picking the wrong task. Beginners either automate something so rare it never pays back the setup time, or something so sensitive that one AI mistake causes real damage.

This lesson gives you a simple filter for choosing good automation candidates, a way to estimate whether a task is worth automating, and clear rules for what to keep manual.

What You'll Learn

• The three tests a task should pass before you automate it
• How to estimate payback with simple math
• Which tasks AI handles well and which it handles badly
• When to keep a human in the loop

The three tests

Run any task through these three questions before building anything.

1. Is it repetitive? Automation pays off through repetition. A task you do daily or weekly is a candidate. A task you do twice a year is not, even if it is annoying. The setup and maintenance will cost more than it saves.

2. Is it describable? Could you write instructions for a smart temp worker who starts tomorrow? If you can describe the inputs, the steps, and what a good result looks like, an automation can probably do it. If your honest description is "I just know it when I see it", the task is not ready. Often the fix is to do the task manually a few more times while writing down what you actually decide, until the hidden rules become visible.

3. Is a mistake cheap? Every automation fails sometimes, and AI steps fail in stranger ways than rule based steps. Ask what the worst realistic mistake costs. A mislabeled internal email costs nothing. A wrong number sent to a customer costs trust. Automate the first kind freely. Gate the second kind behind human review, or do not automate it at all.

A task that passes all three is a green light. A task that fails the first test is not worth it. A task that fails the second needs clearer rules first. A task that fails the third needs a human checkpoint built in.

The payback math

You do not need a spreadsheet for this, just honest numbers.

Weekly time saved = minutes per run × runs per week.

Compare that against setup time, roughly two to four hours for your first version of a simple workflow, less as you get faster. A task that saves 15 minutes a day pays back a three hour setup in about two weeks, and then keeps paying every week after. A task that saves 10 minutes a month pays back in roughly three years, which means never, because the workflow will need maintenance before then.

Two honest additions to the math:

• Count interruption, not just minutes. A two minute task that breaks your focus eight times a day costs far more than sixteen minutes. Tasks that interrupt are worth more than their clock time.
• Count maintenance. Apps change, workflows break occasionally, and prompts need tuning. Assume a little upkeep per month per automation. Ten fragile automations can quietly become a part time job, so prefer a few solid ones.

What AI steps are good at

AI steps shine on judgment over text where "good enough" is genuinely good enough:

• Classifying: is this email a complaint, a lead, or spam?
• Summarizing: turn this meeting transcript into five bullets.
• Extracting: pull the name, company, and amount out of this message.
• Drafting: write a first reply for a human to edit and send.
• Translating and rewriting: same content, different language or tone.

Notice the shape: messy text in, structured or shorter text out, and small errors are tolerable.

What AI steps are bad at

Keep these away from unsupervised automations:

• Precise math and totals. Language models approximate. Let the spreadsheet or a calculator step do arithmetic, and let AI do the reading around it.
• Facts it was not given. An AI step asked about things outside its input will sometimes invent answers. Feed it the data it needs; never rely on what it "knows".
• High stakes final decisions. Refunds, hiring, medical, legal, anything with a compliance edge. AI can prepare the decision. A person should make it.
• Tasks you cannot check. If you have no way to notice when output is wrong, you will not notice. Do not automate what you cannot verify.

The human in the loop pattern

Between "fully automatic" and "fully manual" there is a middle setting that is right for most valuable tasks: the automation does the work, and a person approves the result.

The AI drafts the customer reply, and you tap approve before it sends. The AI flags which invoices look wrong, and you review only the flagged ones. You save most of the time while keeping human judgment exactly where mistakes are expensive.

A practical progression: start new automations with a human checkpoint, watch the outputs for a couple of weeks, and remove the checkpoint only for the cases the AI handles reliably. Trust is earned by track record, not assumed at setup.

Your first candidate

End this lesson with a concrete step. Write down three repetitive tasks from your own week. Run each through the three tests, estimate weekly minutes saved, and circle the winner. Keep it somewhere handy. In the Design Your First AI Automation lesson, you will turn it into a working design.

Key Takeaways

• Good candidates pass three tests: repetitive, describable, and cheap to get wrong.
• Payback is simple math: minutes per run times runs per week, against a few hours of setup plus ongoing upkeep.
• AI steps are strong at classifying, summarizing, extracting, and drafting text, and weak at math, unseen facts, and high stakes calls.
• Human in the loop is the default for anything customer facing: AI does the work, a person approves it.
• Interruption cost counts. Automating focus breaking tasks pays more than their minutes suggest.