Everyone's explaining AIOS wrong · AI Automation Agency Hub

Everyone's explaining AIOS wrong

We started talking about AIOS in 2025 and put out the first video on it in February this year. Since then over 1,000 agencies in the Accelerator have set one up, and there are now dozens of videos on YouTube explaining what it is.

They all get the name right. AI Operating System. Then most describe a tool stack, which is not what it is.

AIOS is three layers.

Layer 1: the business operating suite. Where the company's data already lives. Google Workspace, Microsoft 365, or locally on someone's machine. Not a decision you make, a fact you go and find.

Layer 2: the AI model. Claude, GPT, Gemini, whatever's winning that month. Everyone argues about this layer. It matters least, because they change every few weeks and they're getting easier to swap.

Layer 3: the connectors. What fills the gaps in Layer 1. Data the operating suite can't see, and platforms it can't write to.

Layer 3 is where almost everyone gets it wrong, so here's the clearest example I can give you.

Brick-and-mortar client. They're on Google, so Layer 1 is Google Workspace. But the thing that actually runs the business is their CRM, and Google can't read it or write to it. Layer 3 is what you build to close that gap, using Layer 2 to do the work.

Every AIOS has these three layers. What separates a good setup from a broken one is the picks, and from watching a thousand-plus get built, it comes down to five questions.

Where does the business already live, and what are they already paying for? Layer 1, and it's the first thing you find out, not the first thing you decide. If the client already pays for Copilot, the data's there and the permissions are set.

Does it need to run when nobody's at a computer? Still Layer 1. If it lives on someone's laptop, it sleeps when the laptop does. Overnight or scheduled work has to live somewhere that stays awake.

Can the client's data leave their building? Constrains Layer 2 and Layer 3 at once. If the answer is no, you're down to models running in their own cloud tenancy or on their hardware, and any connector routing data through a third-party service is out too.

What does it need to read, and what does it need to write? Layer 3, and it's permissions, not models. Read access is easy to grant and hard to regret. Write access can damage live records. It drafts, a human sends, until it's earned more.

How much of what you build is portable? The models swap easily now. What you build around them often doesn't, and that's the lock-in people miss.

Which brings me to Dots. You give one a goal, it gets its own computer, connects to your apps, and keeps working after you close ChatGPT. All three layers in one product.

The catch is question five. Not the model, the wrapper. Moving in is easy, they'll even import your setup from Claude. Moving out is mostly copy and paste. And they can change the rules whenever they want. Custom GPTs were how you built on ChatGPT for three years, and on December 11th they stop working.

So, theirs or yours? Look at Layer 3. List the tools the business cannot run without. If the prebuilt one can read and write to all of them, use it. It's fast and cheap. If it can't, or the data can't leave, or the client needs to own the whole thing, build your own.

The best setup is always your own, because you pick the best option in every layer. That got less risky this year. The same skill files now run in ChatGPT, Claude and Copilot, and your own connectors plug into nearly all of them.

This moves fast. Dots launched at the end of September, Microsoft renamed and rebuilt Copilot around agents, and Claude started working with Google Docs. All of that landed inside a month.

Which is why I'm not writing you a setup guide. It would be out of date before most of you read it.

The three layers don't change.

Nico.

1 comment

Everyone's explaining AIOS wrong