AI as material, not spectacle.

A good architect does not decorate a building with concrete. Concrete is what the building is made of. It is structural, quiet, doing its work in the background. Nobody thinks about it, which is exactly the point.

We keep coming back to that image when we talk about AI. Most of the AI in software today is decorative. It is announced with a badge, a sparkle icon, a modal that appears when you did not ask for one. The product wants credit for using it. That is almost always a design failure.

The best AI features are the ones you do not notice, until you try to imagine the product without them.

Three quiet places AI belongs

The first is at the seams. The places where a product used to require the user to translate between two systems, write this in the right format, tag this correctly, remember what you called that thing last week. These are the boring layers. AI absorbs them well, and taking them away is genuinely felt.

The second is in the readerly work. Grouping, summarising, comparing, noticing. Small acts of attention that a person could do, but that quietly erode focus when repeated a hundred times a day. Done well, this kind of AI does not replace anyone's judgement, it just clears the desk so the judgement can happen.

The third is in the long tail of things a product would like to do but cannot afford to hand-craft. Personal onboarding. A helpful error message that actually understands the user's context. An empty state that is written for this specific person. These are the small courtesies of software, and AI finally makes them economical.

What we try not to do

We try not to make the AI the interface. Chat is a fantastic tool and a mediocre product surface. We reach for it last, not first.

We try not to make the AI a personality. The product does not need a name, a voice or a face to be useful. Adding one is almost always a way of hiding the fact that the underlying feature is not strong enough on its own.

And we try, quietly, to keep the surface honest. When the model is uncertain, the interface should say so. When it makes a mistake, undoing that mistake should be the most obvious thing on the page. These are old-fashioned interface concerns, and they matter more, not less, in an AI-native product.

Most of the AI-native software worth using in five years will not feel AI-native. It will just feel like software that finally knows what it is doing. That is the standard we are trying to hold ourselves to.