A year ago, the dominant question in AI product design was: where does the chatbox go? The answer, increasingly, is: nowhere visible. The most successful AI products of the last twelve months have moved the model deeper into the interface and surfaced the result, not the conversation.
From conversation to outcome
The first wave of AI tools assumed the user wanted to talk to a model. Most of them were right, sometimes. But the second wave has realised that the user wants the thing the model can produce — and they'd rather not have to prompt for it in prose.
- Inline suggestions over dedicated chat surfaces.
- Undo and branching as first-class features.
- Outputs that match the existing design system, not a separate playground.
Latency as a material
We now design around the model the way we design around a slow network. The first response is sketched; the second is detailed; the third is committed. This is closer to how illustrators and writers already work — and it's a better mental model than 'wait for the answer.'
“The best AI products don't feel like AI. They feel like good software that happens to be smarter.”
That's the bar. Not 'it can do anything,' but 'it does the thing you need, in the place you need it, with the latency you can live with.' Most of the products we're proud of shipping in 2026 meet that bar quietly.