OpenAI's ChatGPT now builds you a calculator.

The way people reach your product is changing fast. They used to click a link, land on a page, and move through your screens. Now an assistant can answer the whole question before anyone sees your name. The interface itself is turning into the product. Let me catch you up on what moved this month and what it means for your team.

The deep cut

• The interface now completes the task. OpenAI's Intelligent UI turns an answer into a working calculator or chart, no page visit needed.
• Access without judgment fails the user. Siri AI found the right contacts but opened Google Maps from Durham, not Emmett's address.
• A citation is not a reader. Pew found 1% clicked links inside Google's AI summaries.

The answer that builds its own screen

OpenAI started rolling out Intelligent UI with GPT-6 on October 7. Ask a question, and ChatGPT can reply with charts, forms, tappable buttons, even a retirement calculator or a bill splitter you use inside the chat. OpenAI says it trained the model to decide when a visual beats plain text, and when text is enough.

Product manager Aarush Selvan put the goal plainly: for everyday tasks, the most helpful answers aren't just text. To build this, OpenAI made a library of components the model pulls from, then trained it to make good design decisions about layout and interactivity. The chatbot is now assembling interfaces on the fly. That is design work, done by the model, without your team in the room.

Who actually captures the visit

Here is the part that should worry a product leader. When the answer is a working tool, the source visit becomes optional. Karo, writing as Product with Attitude, names three ways you lose the user: they ask the bot instead of searching, an AI summary answers before they click, or they get a usable tool and leave.

The numbers were already bad. Pew's March 2025 data found users clicked a Google result in 8% of visits with an AI summary, versus 15% without. Only 1% clicked a link inside the summary. A citation under an interactive chart credits your work. It does not tell you if anyone noticed your name, opened your page, or signed up. If your business depends on the visit, that gap is now your problem.

Context beats raw smarts

The other shift is where the assistant gets its information. Apple's new Siri AI, tested by NN/g's Caleb Sponheim, found three breakfast spots a brother had texted, then checked whether the user had gone to any of them. ChatGPT, with no access to those messages, searched connected work accounts and found nothing. Same question, different context.

Access alone is not enough. Siri AI confirmed it should route from Emmett's address, then opened Google Maps starting from the user's actual location in Durham, NC. A detail confirmed in one step did not survive into the app that acted on it. The lesson for your team: judge the whole task, not whether the assistant found a contact or launched the right app.

The part you can still own

So where does design still win? In the handoff between automation and the person. The Muzli piece on UX in the AI era frames the new flow as intent, then interpretation, then recommendation, then user review. That review step is yours to design. Automation should cut effort without hiding what the system understood. Show confidence levels. Make editing, rejecting, and undoing easy.

There is also a storage problem few people see coming. NN/g's study of Claude power users found that even experts rarely edit their context files and often cannot find them. One kept 2,700 files. Another found a rule governing what Claude could do that he did not know existed. If the context that drives your product's behavior is a mess nobody maintains, your output drifts and nobody notices until it is bad.

Where it stays on the device

One more move worth watching. Google released Edge Foresight, a Mac note-taker that transcribes meetings fully offline using an on-device model. Your files, audio, and notes never leave the computer. That is a direct answer to the trust question these assistants raise. For teams in regulated or privacy-sensitive markets, local-first processing is becoming a feature you can sell, not a limit to apologize for.

Three questions for your team

• If an assistant can answer the user's question with a tool built from our content, what reason do we give them to come to us at all?
• When our product acts for the user, where is the review step, and does the confirmed detail actually survive into the final action?
• Who owns and maintains the context that shapes our AI's behavior, and could they find and edit those files today?