AI Can't Automate Your Business (By Itself)

The customer's reply came in at 4:12 on a Friday. Three words and a screenshot: "I'll take it."

The screenshot was a price. $400, for a job that should have run $4,000.

A week earlier, the owner's nephew had watched a few videos about AI agents. Smart kid. He wired one up to read incoming customer emails and fire back quotes — no more waiting on the estimator, no more back-and-forth. It worked in the demo. It worked for a couple of days. Then it read a request for a forty-foot job, decided $400 sounded about right, and hit send.

The AI didn't malfunction. It did exactly what it was told. It was asked to decide a price, so it decided one.

You gave it a decision, not a job

The videos all make the same pitch: hand the AI your tools and let it figure everything out. Point it at your inbox, your calendar, your pricing, and step back.

But "figure everything out" covers a lot of ground. In that one email, the AI had two very different things to do. One was to read a rushed, half-punctuated message and work out what the customer actually wanted. The other was to know what that work costs.

The nephew handed it both. The first job, the AI is good at. The second, it has no business touching.

Every task is really two piles

Take anything you'd want to automate and sort it into two piles.

The first pile is judgment. Reading a messy email and understanding it's a forty-foot job with two add-ons. Noticing a customer sounds ready to walk. Figuring out which of your six services someone is fumbling to describe. That's reasoning with context, and it's the one thing an LLM is genuinely good at.

The second pile is rules and lookups. What does a forty-foot job cost? Where does this get logged? Which crew is open Thursday? None of those are decisions. They're facts. The price lives in a table. You look it up, and it comes out the same way every time, forever.

Give the first pile to AI. Give the second to code. And the second pile is most of the work.

Now run the nephew's disaster again, built right. The email comes in. The AI reads it and works out what's being asked — forty feet, two add-ons, wanted next week. Then it stops. It hands those facts to your code, and the code pulls the real number off the price sheet. The AI never sees a dollar figure until the code hands one back. Then it does the last thing it's good at: writing a clear, friendly reply.

Same tools. Same AI. Opposite wiring. And this version can't quote $400 for a $4,000 job, because it was never allowed near the price in the first place.

Why the broken version demos so well

The "let it figure everything out" approach has one real talent. It demos beautifully.

You watch a video where someone types a sentence and the AI books the meeting, sends the invoice, updates the sheet, all on its own. It looks like magic. Demos always look like magic. Ask anyone who's ever bought software off one.

What the demo doesn't show you is the fourth week, when the AI makes the wrong call on a live customer and nobody notices until the money's gone. The failures don't happen on stage. They happen later, on your account.

The approach that's easiest to demo is the one that hands the AI the most decisions. That's backwards from the one you want running your business.

The test to run before you automate anything

Before you point AI at any part of your business, take the task and split it into the two piles.

What here needs judgment? That goes to AI. What here is just rules and lookups? That goes to code — and it'll be most of the list.

Then find the steps you couldn't sort cleanly, the ones where you weren't sure which pile they belonged in. That's where the $4,000 mistakes live. Those are the steps to slow down on, or to keep a person near.

The businesses getting real value out of AI right now didn't hand it the most to do. They drew the line in the right place: judgment on one side, rules on the other.