What to automate with AI, and what not to

A cheap mistake and an expensive one aren't the same task

Before you hand a task to AI, ask one question: if this goes wrong, who finds out first? If it's you, two minutes later, and the fix is deleting a paragraph and rewriting it, automate it. If it's the customer, three days later, standing in your driveway with a screenshot, don't.

It's the same technology either way. What changes is where the person is standing: in front of the task, reading it before it goes out, or behind it, cleaning up after it already went. A bad draft costs ten seconds. A bad message that went out on its own at eleven at night costs you a lost job, or worse, a job you now have to do at the price the machine made up.

Notice the test says nothing about difficulty. There are brutally hard tasks where the mistake is obvious on sight, and dead simple ones where the mistake travels for a week before anyone looks at it. The model handles the difficulty. You pay for the damage.

Safe to hand over: drafts, sorting, and data someone was going to retype anyway

The estimate written up from your notes after the walkthrough. The 48-hour follow-up email. A service description for the site. The Spanish version of a page you were going to read anyway. In all of them, someone was going to write it regardless: what gets automated is the blank page, not the decision. The draft lands in twenty seconds, gets fixed in two minutes, and if it's bad you delete it.

Sorting is an even better candidate, because a wrong answer sits there in plain sight. Triaging incoming leads by type and urgency, separating price shoppers from people ready to book, tagging job photos by trade, pulling line items off a PDF invoice into a spreadsheet. Mislabel something and you catch it the moment you open the list: the cost of the error is one click.

On the contractor site I built, more than a hundred photos ended up as 54 after two rounds of culling. A machine can run the first round — group by job type, drop the duplicates, the blurry ones, the ones shot at dusk — and save you the tedious part. Which photo goes on the homepage is still picked by someone who knows which work they want more of.

Not safe: anything that goes out the door unread

The classic case is the bot answering the chat window or the business line. It's fine ninety times, and on the ninety-first it quotes a price range you never set, confirms a slot you don't have, or says yes to work you don't do. That mistake doesn't show up the day it happens. It shows up when the customer arrives certain of something you never said, and arguing about it costs more than the job was worth.

The working rule is short: anything with a number in it that will later hold you to it doesn't go out alone. Price, timeline, warranty, service area, license number. In regulated trades, add any sentence that sounds like promising an outcome, which in several states isn't just bad marketing; I wrote about that separately in building a legal services site without promising results.

Same goes for anything published under your name. Answering a one-star review with generated text fails in an unusually expensive way, because it stays up and public forever. And with money, the same split: automate the invoice getting built, keep a person on the send button. An extra zero is the kind of mistake you catch immediately — one second after it has already gone out.

You start with drafts because the review already exists

The reason to start with drafts and sorting isn't that they're easy. It's that reviewing them is already free: the person was doing that task anyway, so you're not adding a checkpoint, you're removing the slowest part of one that already existed. Payback from day one, risk close to zero.

It also hands you the number you need for the next decision. Keep count for two or three weeks: out of every twenty drafts, how many went out untouched, how many needed small edits, how many got thrown away whole. If five in twenty get thrown away, that task isn't ready to run on its own, no matter how good the other fifteen looked. If it's one in twenty and the edits are cosmetic, now you can talk about automatic.

And the move to automatic is a small one: a single category of message, the most repeated and the most harmless — confirming the request came in and saying how many hours until someone calls — with no prices or dates inside it, and a copy to your inbox for the first month. If you still need to work out which task goes first, the minutes-per-week test is in where to start automating in a small team.