There are two ways a product ends up with AI in it, and they produce almost nothing in common.

Bolt-on: a widget in the corner of a finished thing.

The common path is to take a product whose architecture was settled years ago and add an assistant to it. The chat box can see the help documentation and maybe a slice of the account, but it can't reach the data model, can't take a consequential action safely, and carries a disclaimer because the team knows it might be wrong. Pull the assistant out and the product is exactly what it was. The AI is a feature on a shelf, not part of the load the building carries.

AI-native: the model is the substrate.

AI-native means the software was designed around what a model can do, from the first commit. Memory, identity, and the ability to act aren't bolted on afterward; they're the frame. In Friend, the companion isn't a feature inside each app. It's one presence that travels with the person across every surface, remembering what came before, and it's also the identity you sign in with. Remove that and there's no product left, because the product was the companion all along.

If you can delete the AI and the software still works, the AI was never native. It was a widget.

Why the distinction isn't pedantry.

A bolt-on assistant is capped at what the surrounding system will let it touch, which is usually very little, because the system wasn't built to trust it. An AI-native system is built the other way around: the model is a first-class actor with a seat at the data itself, held in check by an explicit consent gate rather than by being fenced away from anything that matters. That's what lets it be genuinely useful without being dangerous, and it's a decision you can only make at the start.

How to tell which one you're looking at.

  • Does the AI reach the data model itself, or only a help index and a thin slice of the account?
  • If you removed the AI, would the product still be the same product, or would there be nothing left?
  • When it does something consequential, does it explain and ask first, or is it fenced away from anything that matters because no one trusts it?

We build the second kind only. It's slower to start and it's the entire reason the result behaves like a colleague instead of a novelty. The posture it comes from is written up in our position on good AI, and what it is like to work through it is on how we work.

Common questions.

Ask three things: does the AI reach the data model itself or only a help index and a thin slice of the account; if you deleted the AI would the product still be the same product; and when it acts, does it explain and ask first. If the AI can be removed with nothing lost, it was a widget, not native.

A bolt-on assistant is capped at the little the surrounding system will let it touch, because that system wasn't built to trust it. An AI-native system makes the model a first-class actor held in check by a consent gate, so it's genuinely useful without being dangerous. That decision can only be made at the start.

Ecodia builds AI-native only, with the model in the room from the first commit. In Friend, its AI companion, the companion is the product and the identity you sign in with, not a feature inside each app. Remove it and there is nothing left, which is the test of whether AI is native or ornamental.