Designing for AI-native interfaces

The shift

For 40 years, software design has been about surfaces — buttons, menus, forms, layouts. You design the thing the user sees, and they figure out how to use it.

AI-native products flip this. The primary interaction is conversational. The user says what they want in natural language, and the product does it. The "surface" is whatever the AI decides to render in response.

This breaks most of our design instincts. Here's what we've learned building Ion.

1. The input matters more than the output

In a traditional app, you spend 90% of your design time on the output — the dashboard, the list, the form. The input is just buttons and clicks.

In an AI-native product, the input is the product. If users can't articulate what they want, nothing else matters. We spend more time on the command palette, the prompt suggestions, the example queries, and the error states than on any single output surface.

2. Citations are non-negotiable

An AI that answers without citations is a magic 8-ball. Users don't trust it, and they shouldn't. Every answer in Ion links back to the source doc, ticket, or thread it came from. This is a UX decision, not a technical one — we could ship answers without citations and save 200ms of latency. We don't, because trust is more important than speed.

3. Show your work

When the AI takes an action — drafts a doc, closes a ticket, posts a message — show the user what it did, immediately, with a way to undo. This is the equivalent of a confirmation dialog in a traditional app, but more transparent. Users need to see the AI's reasoning, not just its output.

4. Defaults still matter

Even in a conversational product, most users will take the default action 90% of the time. Spend disproportionate effort on what the default is — which model, which tone, which format, which length. The conversational interface is the escape hatch; the defaults are the product.

5. Failure states are the product

The AI will fail. It'll misunderstand, hallucinate, time out, or refuse. These moments are where users decide whether they trust the product. Design failure states with the same care you design the happy path — explain what happened, offer a fix, and let the user retry without retyping everything.

What stays the same

The fundamentals don't change: clear hierarchy, good typography, fast feedback, accessibility. AI-native doesn't mean throwing out 40 years of design craft. It means applying it to a new interaction model.