The Great UX Convergence

AI-native UX is still being invented. My bet is that a startup should invent as little of the basic interaction model as possible.

Big Tech + OAI et al. are spending billions teaching people how to interact with intelligence through text and voice. Every product is converging on the same few patterns: an input, a stream of work and a ping when the job is done. I would rather ride that wave than burn startup energy teaching users an entirely new interface.

Do not invent a new interaction model if someone else is already spending billions teaching one to your users.

Start with what people already know¶

Text and voice are becoming the universal interface to software. That does not mean dropping an empty chat box into every product.

Users stare at a blank text box and wonder what they are supposed to say. I take cues from Chinese consumer apps, which surround the input with suggested actions, pills and contextual prompts. The input stays open-ended, and everything around it teaches people what the product can do.

Stream the work¶

AI kills an old assumption about interfaces: that the work has to finish before the interface responds.

ChatGPT made streaming text familiar. Midjourney made partial renders familiar. The same goes for anything an agent builds for you, be it a website, a report or a video. Show it coming together.

Intelligent work takes time, sometimes thirty seconds and sometimes an hour. Do not hide that behind a spinner. Show useful progress, keep a trace of what happened and let the user step in when they need to.

One intelligence, many surfaces¶

The intelligence should not belong to the chat box. The same capabilities should be there through text, voice, mobile, desktop, SMS, a phone call and whatever comes next. I should be able to start a job by voice and review it on my laptop later.

The interface is a surface over the work the system can do. That abstraction is also what lets you route between models, tools and plain old code underneath without the user ever knowing. They do not care which model did the job. They care that the job got done.

What makes all of this reliable (evals, context, memory, tools, routing) sits a layer below the interface. I cover that in AI Native Products.

Less interface¶

The long-term goal is not a better chat box. It is less interface.

For a lot of products, the ideal experience is not opening the product at all. Take software for a small service business. The ideal is a text telling the owner that Thursday’s booking moved, the calendar is updated and the customer already knows. Or that there was an opening tomorrow afternoon, three old leads were contacted and one booked.

The best interface for completed work is often a notification that the work is already done.

Do not anthropomorphize the work¶

Nobody needs “Alice, your AI CMO”. They need more customers.

Giving intelligence a name, a face and a personality makes for a memorable demo, but it puts the technology ahead of the job to be done. I prefer capabilities with clear outcomes and traceability. Show what happened, show why when it matters and otherwise get out of the way.

Start with interfaces people already understand. Build toward a product they barely need to operate.