9
min read
The wild west of generative UI
Interfaces are starting to generate themselves. The real design work is moving upstream, into the systems and principles that guide what gets generated.
Dinesh Davé
Co-founder + Creative Director
You can start a Starbucks order in ChatGPT now. There's an official app for it: type @starbucks, browse the menu inside the chat, pick your drink, and at the moment of paying it hands you back to the Starbucks app to finish the job. The whole loop takes longer than the three taps you'd have made in the app you already have, and clips of the chatbot losing that race made the rounds as proof that all this talk of AI replacing our interfaces is overblown.
The joke lands, but I think it points the wrong way. Interfaces are starting to generate themselves, and from what we're seeing as we build with these tools at our firm, the change arrives in three levels, roughly the way autonomous driving did. The chat box that every software company is currently bolting onto its product is a transitional step, and to see what it's transitioning toward, it helps to look at how we got here.
Two ways of talking to technology
Until a few years ago there were two. The first is the screen. You tap, swipe, and pinch, and decades of product design have made this so efficient that you can call a cab or order a coffee almost without looking, because the buttons live where you know they live, and a screen with a handful of options beats typing a sentence every time. The limitations show up as products grow: the options that don't fit get buried under menus, sub-menus, and settings panels, which is a big part of why professional software can feel intimidating by the time the company has been in business for a decade.
The second is voice. Siri arrived in 2011 and Alexa in 2014, both pitched as the future of computing, and a decade later the NPR and Edison Research surveys of what people actually do with them read like a light-switch manual: music, weather, timers. The reason, in part, is cognitive load. With a screen the options sit in front of you, but with voice you compose the command from scratch every time. I, for one, love our smart home. My wife? Not as much. I’ve become the dedicated smart home speaker of the household as she tries to command Siri to “turn off the blinds” in her slumberous state every night.
And when the industry pushed voice past those jobs, the people pushed back. A six-year-old in Texas famously ordered a $160 dollhouse through the family Echo. The Humane Pin (RIP) raised $230 million to replace your phone with a wearable voice assistant and was sold for parts to HP within a year of shipping. The Rabbit R1 sold 100,000 units, and its founder later shared that only 5,000 people were still using theirs at any given time. The one device in the category that's working, Meta's Ray-Ban glasses, got there by adding a camera and keeping your phone in the loop instead of trying to replace the screen.
When a new way of interacting with technology shows up, we throw it at everything, and then over a few years we converge on the narrow set of jobs it's actually good for. Voice converged on controlling the home and setting timers. Everything else went back to the screen.
Consumer tech goes first
ChatGPT launched at the end of 2022 and reached an estimated hundred million users in about two months, the fastest consumer technology adoption in history. Those two months trained an enormous number of people on a third way of talking to technology, after the tap and the voice command: describe what you want in natural written language, get work back. And because consumer products train people at a scale business software can’t, the pattern is now flowing into B2B the way it always does. Salesforce shipped Agentforce. Framer launched agents inside its canvas this June. Figma announced its design agent that can make its own tools at Config two weeks ago. Nearly every software company is adding a conversation to its product, and if the history above is a guide, this is the throw-it-at-everything phase, with the convergence still ahead.
What the conversation did fix is fragmentation. Planning a trip used to mean using a dozen apps: to book flights, book a hotel, rent a car, check the weather, make a packing list, download maps, etc., each pulling from its own databases while you did the assembling in your head. Today an LLM with enough context does the assembling for you and hands back one consolidated answer. The limit is the format, because a wall of text is a poor interface for a hotel booking, a trip itinerary, or, in my recent case, a kettlebell workout plan.
The three levels of generative UI
Generative UI is the interface the model assembles on the spot, for you, in the moment you ask. Google has been shipping this in Gemini since late 2025 and said at I/O this year that it's bringing it to everyone in Search this summer. ChatGPT's apps render interactive interfaces in the conversation. Claude began building interactive visuals inside its chats in March 2026, sliders you can drag and diagrams you can play with instead of a wall of text. And from what I’m seeing, it arrives in stages.
I’ve been thinking about those stages by borrowing the auto industry’s classification of self-driving: the SAE defines three tiers of driving automation — supervised, conditional, and unsupervised — and software seems to be climbing a similar ladder.
Level: Zero
Level zero is where business software sat in the recent past, and it's still most of what we use day to day: a tool with alerts and customizable dashboards. Those dashboards already admit that every business is different, but the software handles it by handing you the work: you drag the widgets, save the views, build the reports, define the custom fields, and decide which alerts matter. And when your CRM pings you that a form came in from an unregistered domain, you go fix it yourself (I’m looking at you, HubSpot). The customization is real but soft, and a person performs most of it.
Level: One
Level one is autonomy one prompt at a time, and it's what many traditional, pre-AI software tools are launching today. The chat boxes that Framer and Figma just shipped are this level: an agent inside the product that you hand a task in plain language, sometimes with a toggle for which model you'd like to do the work. You ask it for a thing, it goes and does that thing autonomously, you check the result, and you prompt again, and on the more complex work it iterates with you until it's right, though the simple asks land in one shot. In the driving analogy this is driver assistance. You supervise every action, because the agent still gets it wrong often enough that you have to.
Level: Two
Level two is the product that does the job without being asked, and the AI-native products are there now. Clarify, an AI-native CRM, transcribes all of your sales conversations, knows your prospects, and your deals, and processes all of it autonomously: notes generate themselves after a meeting, action items appear, deal records update based on what was said on the call, and when a new deal gets discussed, the CRM opens it for you, no prompting involved. In driving terms this is conditional automation: within the domain it knows, the system acts on its own, and your job shifts from asking for the work to reviewing it.
Level: Three
Level three is the product that generates itself around you, and we think it's a few things arriving together. The product keeps doing the job on its own, the way level two already does. On top of that, the product becomes hyper-personalized, generated for the company, the team, and the person using it. We're a two-person firm and we run on HubSpot, and we use about half of it, while the other half lingers in our sidebar untouched. The level-three version hides what we don't use, and then goes further: it takes everything it knows about our company, the meetings, the clients, the marketing data, and generates the interface a two-person marketing firm actually needs, which is a different product from what a hundred-person startup needs, and different again from a thousand-person company with a thirty-person marketing team.
The small apps ChatGPT and Claude generate on demand are the small-scale preview of this. Before every gym session I ask ChatGPT for a kettlebell workout, and today what comes back is a list. I then paste it into my notes app, dip into YouTube to learn the moves, and start a timer in my clock app. The generative version is a mini app built for that hour, timer already included, a figure demonstrating the swing, made just for me, built once, evolved for every work out. At level three the whole product behaves that way.
The last piece is that you can extend the product yourself. Figma shipped a version of this two weeks ago: you can now prompt the agent to build you a plugin that lives in your file as a small reusable tool your whole team can run. Point that at any business software. We duplicate the same newsletter template in HubSpot every week and paste in new HTML by hand, and the level-three version is a button we built ourselves that pulls the draft from Notion and rebuilds the email, on top of an already hyper-personalized product experience custom-built just for me.
The industry is laying the plumbing for this already. Salesforce, after 25 years of perfecting screens, shipped Headless 360 this spring, every capability exposed as an API or an agent tool, with co-founder Parker Harris asking out loud why you should ever log into Salesforce again. In driving terms level three is unsupervised: the Waymo you step into has no steering wheel, and soon enough it will know you when you get in, seats adjusted, dashboard arranged your way. Like unsupervised driving, it needs years of training data before it's trustworthy. Tesla built full self-driving by processing camera footage from millions of customer cars, and every clumsy generative widget we use today is the software equivalent of that footage. The wild-west messiness of this moment, where the same question produces three different interfaces on three different days, is what the data-gathering years look like.
That's also why I don't buy the version of this future where one super-app swallows all software. A carmaker had to master one domain completely before it could remove the driver, and I'd expect software to work the same way: the products that reach level three will be the ones that know a single job inside and out and have the data infrastructure to keep learning it, which should be reassuring if you're building one of them.
The work moves upstream
Every level moves the same work in the same direction. At level one a person still checks every output, so the principles can live in your team's heads. At level two your team has to write down the logic that decides what the product does on its own, because nobody is prompting it. And by level three, principles are most of what a product team makes: the components, the patterns, and the if-this-then-that logic that decide what gets generated for each company, each team, and each person, because the screens themselves get assembled on demand and their quality is set by the quality of the principles underneath.
How a company codifies its brand and design for this future is its own subject, and we'll write about it next. If you're thinking through what generative UI means for your product in the meantime, I'd be happy to compare notes.
