What to Automate With AI vs Doing It by Hand: A Fair Test

Handing a task to AI is not free. You spend time explaining it, then more time checking what comes back. So the honest question about what to automate with AI is not "can it do this", but "does the whole loop beat doing it by hand?" For some jobs it clearly does. For others the manual way wins, and it is better to know which before you spend an afternoon setting up.

This post gives you a way to compare the two on your own tasks. It does not claim numbers from a test, because the answer depends on your work. If you want a guided version, Lesson 1.5 covers it: Find the jobs to hand to AI.

AI wins when a task repeats often, follows a pattern you can describe and is quick to check. The manual way wins when the task is rare, delicate or hard to explain.

The real cost of each way

Doing a task by hand has one cost: your time each run. Handing it to AI has three:

• Setup: writing the brief, gathering an example, testing it once or twice.
• Run time: usually short, but not zero, because you still supply the inputs.
• Checking: reading the output and fixing it. AI tools can make mistakes, and the vendors say so on their own pages, so this step cannot be skipped.

Setup is paid once, run time and checking are paid every time. That is why frequency matters so much: a task you do daily spreads the setup cost across many runs, while a task you do twice a year never pays it back.

A worked comparison (illustrative)

Say you run a small cleaning company and write a weekly schedule message to your team. The numbers below are invented to show how to compare.

Under these made up numbers, you save about 20 minutes a week after a 45 minute setup, so it pays back in the third week. If the message went out only once a month, the same setup would take much longer to repay. Change the numbers to match your own task. If you cannot guess them, time yourself on the manual version for two weeks first.

When the manual way wins

Keep a task by hand when any of these is true:

1. It is rare. Anything you do a few times a year rarely repays the setup.
2. It depends on judgement about a person. Refunds, complaints, hiring decisions and sensitive messages need you.
3. You cannot explain it. If your instructions would be a list of exceptions, the AI has to guess.
4. Mistakes are expensive and hard to spot. Legal wording, tax figures and medical information need a qualified human and official sources.
5. The task is the point. Writing a personal note to a long-term client may matter precisely because you wrote it.

When AI is worth the setup

AI tends to pay back when most of these are true:

1. The task repeats at least weekly.
2. It follows the same rough steps each time.
3. You have a good past example to paste in.
4. The output is checkable in a minute or two.
5. A mistake is cheap and visible.

Good fits for small businesses and solo creators include first drafts of routine emails, summaries of your own meeting notes, reformatting information into a standard layout, and turning rough notes into a post draft that you then edit.

The middle ground: split the task

Many tasks are neither fully manual nor fully hand-off. A client proposal, for example, has a human part (the offer and the price) and a repeatable part (the standard sections, the formatting, the first summary). You can hand the repeatable part to AI and keep the decision. Splitting like this often gives the best return, because the part you keep is the part that needed you.

A simple decision test

Ask these four questions about a task and count the yeses.

1. Does it repeat weekly or more?If not, stay manual for now.
2. Can I explain it on one page?If not, stay manual or split it.
3. Can I check the result in a couple of minutes?If not, the checking cost may erase the saving.
4. Is a wrong result cheap to catch?If not, keep a human in charge of the final step.

Four yeses is a strong candidate. Two or fewer is better kept by hand for now. Three is worth a small trial with a low risk version.

Do not forget the quality side

Time is easy to count, quality is harder. A hand-written message may have warmth that a draft lacks, while an AI draft may be more consistent than something you write when tired. Compare both on the same real task, not on what you imagine. Run the job both ways once, put the results side by side and ask which one you would send. Anthropic's prompting guidance says examples are among the most reliable ways to steer tone and structure, so if the AI version feels flat, add a better example to the brief before you decide it cannot match you.

If you use a tool that works with files on your own computer, such as Claude Code, which its documentation describes as able to edit files and run commands, the same comparison applies. Count your setup, run and checking time.

A one week trial

You do not have to guess. Try this:

1. Pick one task that scored well on the decision test.
2. For the first run, do it by hand and time it.
3. Write a brief and run it with AI on the same material. Time the setup, the run and your checking separately.
4. Compare the results side by side.
5. Repeat for two or three runs, because the first run carries the setup cost.
6. Decide using your own numbers.

For help with the brief, see How to Brief AI Like a Senior Hire.

Questions people ask

Is AI always faster than doing it myself?

No. Once you include setup and checking, rare or delicate tasks are often faster by hand.

How long should I trial a task before deciding?

Two or three runs. The first includes the setup cost, so it can make AI look worse than it is.

What if the AI result is good but I still check everything?

That checking time is part of the cost. If it still saves you time, keep it. If not, try a better example in the brief or leave the task manual.

Can I switch back to manual later?

Yes. Nothing about a hand-off is permanent, and keeping the brief means you can pick it up again.

Where to go next

If you want a guided path to pick the jobs that pass this test, Lesson 1.5 walks you through it: Find the jobs to hand to AI. For a repeatable weekly routine, see the Find framework.