My business runs on an app AI built me

I sell a claim that's easy to say and hard to prove: AI should be invisible. Built into the work you already do, not bolted on as a chatbot you have to remember to visit.

Rather than keep arguing it, I built the thing. The studio runs on a Mac app I made for it. Not a demo kept alive for screenshots — the software I open every morning to see what the business needs from me today.

Here's what's actually in it.

Dictation that beat the thing I was paying for

Hold a key, talk, let go. The words appear at the cursor in whatever app I'm in — cleaned up, punctuated, with my proper nouns spelled the way I spell them. It works everywhere on the machine, including places that historically fight back.

It runs on models on my own computer, which matters twice: it's under two seconds end to end, and nothing I say leaves the Mac. Client names, pricing, half-formed ideas — none of it goes to a server.

There's a well-known tool that does roughly this for about twelve dollars a month. Mine took a day.

That day-versus-subscription maths is the actual story. It's not that I'm fast. It's that the cost of building a small, exactly-right tool has collapsed, and most people haven't updated their instincts yet. The default is still "find software that nearly fits and pay forever." That default is now often wrong.

A production line that is also the demo

Client sites get built from inside the app. Templates, tasks, deploys — the whole path from "yes" to "live."

I like this part because it collapses two things that are usually separate. The system I use to do the work and the system I use to show the work are the same object. Every improvement to how I operate is automatically a thing I can put on a screen for someone.

The app talks to the AI that improves it

This is the part I'd have found far-fetched a year ago.

When something in the interface bothers me — a screen that's cramped, a button in the wrong place, a number that's wrong — I pin a note directly on it. That note goes into a shared mailbox that my AI coding sessions read. A session picks it up, fixes it, and reports back. Often overnight.

In the morning I look at what changed. Then I approve it, or I don't.

The decision that makes all of it safe

Every one of those loops could have been fully automatic. None of them are.

The overnight fix waits. The site build stops before it ships. Dictation puts words on my screen and never sends them anywhere. It's the same pattern everywhere: the machine does the work, I keep the throttle.

That isn't nervousness. It's the thing that makes this shippable into someone else's business.

Small business owners will happily let software do the typing, the drafting, the chasing, and the reconciling. What they won't do — correctly — is let it act in their name without looking first. Every real objection I hear is some version of "what if it says something stupid to my client." An approval step answers that permanently, and it costs nothing to build in from the start.

Most AI products get this backwards. They chase full autonomy, hit the trust wall, and then bolt on a review step that feels like a nag. If you design the human in from the beginning, the same step feels like control.

What I'd say to another owner

I'm not telling you to build a Mac app. I'm telling you what it means that I could.

Look at where your week actually goes. Not the work you're paid for — the admin around it. The re-typing. The rebuilding of the same quote from a blank page. The follow-up you meant to send on Tuesday. The reconciling of four tools that each know a quarter of the truth.

Those are small, dull, specific jobs. They are exactly what this technology is good at, and exactly what generic software refuses to shape itself around. A tool built for your business — that does only those jobs, quietly, with you holding the approve button — is now a days-to-weeks project rather than a years-and-a-funding-round project.

That's the whole thesis. The studio is just me proving I'll eat it first.

If your week has one of those jobs in it, tell me what it is. I'll tell you plainly whether it's worth building — and when it isn't, I'll say so.