There’s a peculiar form of collective delusion running through the AI industry right now. Every week brings a new model benchmark, a new reasoning capability, a new context window expansion. And every week, the products built on top of these models remain almost comically unchanged: a text box, a send button, a streaming response.

We have arrived at the most powerful general-purpose technology since electricity, and the dominant interaction paradigm is still a chatbox.

This isn’t a failure of engineering. It’s a failure of product imagination.

The chatbot interface emerged as the default for AI products for one reason: it was the fastest thing to ship. When GPT-3 arrived, the obvious move was to let users type a prompt and see what happened. That was fine for a demo. It became a problem when the demo became the product.

Consider what a text box actually demands of the user. It assumes you know what to ask. It assumes you can articulate your need with precision. It assumes the right interaction model is request → response, a fundamentally transactional pattern borrowed from search engines and command lines.

This is the interaction design equivalent of giving someone a Ferrari engine and asking them to push-start it every time. The power is there. The experience is broken.

The few AI products that have found genuine product-market fit share a common trait: they didn’t start from the model. They started from the workflow.

Cursor doesn’t ask you to describe code in natural language. It watches where your cursor is, reads the surrounding context, and offers completions that respect the architecture of your codebase. The AI is powerful, but the product insight is about placement — putting intelligence exactly where the work happens.

Linear integrated AI not as a feature you invoke, but as infrastructure that silently triages, categorizes, and routes work without requiring anyone to write a prompt. The interface is unchanged. The work just flows better.

This is the difference between model-centric design and context-centric design. Most AI products today are model-centric: they showcase the model’s capabilities and ask users to figure out how to apply them. The products that work are context-centric: they understand the user’s environment, task, and intent, then deploy AI silently in service of that context.

If we move past the chatbox, what replaces it? After studying dozens of AI-native products, I’ve identified four interaction patterns that actually work:

The AI operates continuously in the background, surfacing insights or taking action without being asked. Think of how Gmail’s smart compose works — not as a feature you activate, but as a presence that’s always contextually aware. The best ambient interfaces feel like competent colleagues who anticipate needs.

Instead of a separate AI panel or chat window, the AI’s output appears directly within the user’s existing workflow surface. Figma’s AI features, for instance, generate variants and alternatives inside the canvas, not in a sidebar. The spatial continuity is crucial — it means the AI speaks the language of the tool, not the language of prompts.

The next wave of AI products will be defined not by what the model can do, but by how the interface translates capability into utility. This is fundamentally a design problem, not an engineering problem.

The companies that understand this — that invest in interaction design with the same intensity they invest in model training — will build the enduring products of the AI era. Everyone else will be building chatbots on top of commodity APIs, competing on price in a race to zero.

The interface isn’t a wrapper around intelligence. It’s the mechanism through which intelligence becomes useful. And right now, that mechanism is broken for almost everyone.

The paradox is that we’ve never had more powerful AI, and we’ve never been worse at making it usable. That gap is the opportunity.