Get AI out of the browser

“What’s the actual first step?” It’s the question I get most from owner-operators, usually after they’ve been using AI for a few months and suspect they’re missing something. They are, but it isn’t a tool or a subscription.

It’s this. Stop watching AI work in a browser tab and put it to work on your own files.

Here’s what I mean. Last week I asked AI to draft a field note about a website I built for a design studio. I didn’t describe the project. I pointed it at the project folder.

It read the code, the content setup, and four months of change history. Then it came back with a detail I wouldn’t have thought to mention: the evening in July the client needed a third image size and got one. That became the spine of the note. It still can’t tell what’s true from what’s plausible, so every claim in that note got checked before it went out. That part doesn’t get handed over.

In a browser tab, the same job goes differently. I type a paragraph explaining the project from memory. The AI writes something tidy from that paragraph, and it sounds like every other post about websites. Not because the model is worse. Because it only knows what I remembered to tell it.

That’s the gap for most people using AI at work. They’re paying for a very capable tool and feeding it through a letterbox. You describe the spreadsheet instead of handing it over. You paste in three emails and forget the fourth, the one that changes the answer. You explain your business again in every new chat. Then you copy the result back out and fix the formatting.

The desktop apps change the shape of that. Whichever AI you already pay for, there’s very likely a Mac or Windows app that can work with folders you give it access to. Which one doesn’t matter much. What matters is that you stop describing and start pointing.

For someone running a business, not writing code, it looks like this:

IN THE BROWSER                    ON THE DESKTOP

"our quotes usually look     →    "read the quotes folder and
 something like…"                  list the ones nobody chased"

paste one supplier price     →    "compare every supplier price
 list                              list from this year"

describe last month's        →    "open the sales export and
 numbers                           tell me what changed"

copy the draft back out      →    "write it into the file"

Every job on the right is one of the boring, repeating tasks I keep telling people to write down. They’re boring precisely because the work is spread across files. And spread across files is exactly where a chat box can’t reach.

A chat window only knows what you remember to tell it.

Two honest cautions, because this is where people get burned.

First, access is a decision. Give it one folder for one job, not your whole drive. Anything with client personal information, payroll, or health details deserves a pause before you point anything at it. If you’d be uncomfortable explaining it to the person the file is about, don’t.

Second, let it read before you let it write. Start with jobs where it only reports back: summarise these, compare those, find what’s missing. Once you trust how it handles your files, let it draft into them. Check its work the way you’d check a new hire’s first week, because that’s roughly where it is.

So that’s the first step, and it’s smaller than people expect. Install the desktop app for whichever AI you already pay for. Make a folder with copies of the files behind one annoying job. Copies, not originals. Then ask it to do that job with the files in front of it, and compare the result to what you’d have got in the browser.

For the owner-operators I work with, this is a bigger jump than any new model has been. The model was never the bottleneck. What it could see was.

There are steps after this one. A file that teaches it how you write, so drafts come back in your voice. Saved instructions for the jobs you repeat every week, so you stop explaining them. But all of it is downstream of pointing it at real work. Do that first.

This is most of what I cover in the workshops: your team’s real files, on their own machines, with the guardrails set up first.