Hand the boring work to an agent.

To automate tasks with AI, start with a task you repeat often. Write down its steps and its finish line, then hand it to the simplest tool that can do it reliably. Fixed steps suit a small script or a no-code automation. Messy inputs and judgment calls, like reading receipts or sorting email, suit an AI agent.

Keep a human check on anything that sends, pays or deletes, and measure the time you actually get back. This guide shows how to find the right tasks, pick the right tool and brief an agent, with five examples you can copy.

01Which tasks are worth automating?

Every repeat task has a hidden yearly cost: how often you do it, times how long it takes. A five-minute chore you do ten times a week costs you more than 43 hours a year.

Spend a week noting your repeat tasks, then do the math. A small shop’s list might look like this:

Good candidates share four traits. They repeat, their inputs are digital, you can tell when they are done right, and a mistake is cheap to catch. Tasks that happen twice a year, or that hinge on a relationship or your judgment, are poor candidates.

Setup time counts too. The xkcd chart “Is It Worth the Time?” shows how long you can spend improving a routine task before you lose more time than you save over five years. Shave five minutes off a daily task and you save about 150 hours in that span.

If you run a small business, our list of jobs you can hand off to AI this week is a good place to find more.

02Script, no-code tool or AI agent: which should you use?

The right tool depends on the shape of the task, not on what is newest.

The categories are blurring. No-code tools are adding AI, and Zapier, for example, now offers its own agents. If the difference between an agent and a fixed automation is still fuzzy, our guide to what an AI agent is explains the loop that sets agents apart.

03Automate tasks with AI: five examples with prompts

Each example includes a prompt to adapt and the human check that keeps it safe. The pattern stays the same: goal, inputs, finish line, limits. Run each one on a small sample first. For more on writing briefs, see our prompt engineering guide. For ten more ready-to-run examples, see our list of agentic workflows.

1. Clean up a messy contact list

Exports from different tools rarely agree. An agent can merge them, fix formatting and flag likely duplicates without deleting anyone.

Merge contacts-shop.csv and contacts-newsletter.csv into one file called contacts-clean.csv.
Standardize names (capitalized), emails (lowercase) and phone numbers (international format with country code).
Do not delete anyone. If two rows look like the same person, keep both and mark them "possible duplicate" in a new column.
When done, tell me how many rows went in, how many came out, and how many you marked.

Human check: review the rows marked as possible duplicates before you import the file anywhere.

2. Turn the weekly export into a report, then script it

The first run is a conversation. After that, it should be a script, or see our guide to AI for Excel and Google Sheets if a formula would do the job instead.

Every Monday I export sales.csv from our shop. Build a one-page summary: total revenue, number of orders, top five products, and the change from last week (last week's file is sales-prev.csv).
Then write a script that makes the same summary from any two files, so I can run it myself next week. Tell me in one line how to run it.

Human check: compare the script’s first report with the agent’s, and spot-check two numbers against the raw export.

3. Triage your inbox, with drafts only

Email is where agents save a lot of time, and where a mistake is most visible. Let the agent read and draft, but never send. For the drafts themselves, our guide to writing emails with AI covers phrasing that still sounds like you.

Go through today's unread emails. Sort each one into: needs my reply, for my information, or can wait.
For each email that needs my reply, write a short draft in my usual tone and save it as a draft. Do not send, forward or delete anything.
Finish with a list: sender, subject, category, and one line on what they want.

Human check: read every draft before it goes out, and be suspicious of any email that asks the agent to do something.

4. Move data between apps that do not connect

Some jobs live in apps with no integration: an old supplier portal, a government form, a desktop program. A computer use agent can read one screen and type into another, the way you would. Our guide to computer use agents explains how they work and where they stumble.

Open orders-this-week.pdf. For each order, go to the supplier portal in the browser (I am already signed in), click New order, and enter the item code, quantity and delivery date.
Stop before you click Submit on each order, and show me the filled form. I will approve each one.

Human check: approve each submission. Once the first few are right, you can review in batches.

5. Watch a few web pages for changes

Checking competitor prices, job boards or grant deadlines is a classic weekly chore.

Visit these five pages: (paste the links). For each one, record the current price or deadline and compare it with last week's table in tracker.md.
Update the table, and put a short summary of anything that changed at the top.
Only read these pages. Do not sign up, sign in or fill in any forms.

Human check: skim the summary of changes. If a number looks wrong, open that page yourself.

04Keep a human check where it matters

Automate the work, not the decision. An agent can do nearly all of a task, as long as you own the few steps that cannot be undone. Put a checkpoint in front of each of these:

• Sending anything to a person. A wrong email cannot be unsent.

• Moving money. Refunds are slow and never guaranteed.

• Deleting or overwriting. Work on copies and keep backups.

• Publishing. Public mistakes travel fast.

• Unusual input. Odd data is where automations break, and where attacks hide.

05How to measure the time you saved

Time the task by hand two or three times before you automate it. Afterwards, time everything the automation still needs from you: briefing, checking and fixing.

Hours saved per year = (minutes before minus minutes after) x runs per week x 52 / 60
Weeks to pay back setup = setup minutes / minutes saved per week

Take the weekly report above. By hand it took 45 minutes. With the script the agent wrote, running and checking it takes 10. That saves 35 minutes a week, about 30 hours a year, and three hours of setup pays for itself in about five weeks.

Keep a simple log for the first month: the date, minutes spent and anything you had to fix. If fixing takes longer than doing the task by hand, the automation is not working yet. Tighten the brief, or move the task to a script.

06FAQ

Read next: is it safe to let an AI agent control your computer?, or AI agent vs chatbot if you are still choosing.