Every app you have ever shipped is a way to reach some records and some rules about those records. The buttons, tabs, wizards and settings panes existed because typing at a database was not something we could hand to a person. That constraint is what we built an entire industry on, and it no longer holds.
Not all. Most. Take an honest inventory of the enterprise tools and consumer utilities you use in a week and count how many are meaningfully more than create, read, update, delete, wrapped in policy. The interesting question is not whether the rest can be conversational — it is what the exceptions have in common.
Nobody wants to open the app, find the tab, apply the filter and scroll. They want the answer or the effect. Every screen between intent and outcome is overhead we invented, charged to them, and then called a user experience.
A seven-screen workflow becomes one sentence and a tool call. This is not a new skin on the old thing — it inverts the model. The user no longer navigates a structure we designed; they state a goal and the structure is assembled for them, once, and thrown away.
Comparison across many items. Spatial and geographic data. Ambient state you glance at rather than ask for. Direct manipulation of something creative. Anything with real bandwidth, where a sentence is a worse container than a picture. The discipline is knowing which is which — and being willing to say "this one stays a GUI".
What you ship is a set of capabilities an assistant can invoke, with a contract, permissions and a description good enough for a model to choose it correctly. The description is now part of the API surface. Nobody has good taste about this yet — that is precisely why it is worth working on in public.
A stateless chat is a worse app. An assistant that remembers your preferences, your history and your constraints does not need the settings screen, the onboarding flow or the saved-search feature — three surfaces deleted by one capability. Conversion work that ignores state produces a chatbot and disappoints everyone.
When the operating system can route a request to a model, the OS becomes a router and the app store becomes a directory of tools. That is an existential question for anyone whose business is owning the software layer above the hardware — and an opening for everyone who was locked out of it.
Custom GPTs, plugins, skills, MCP servers, App Intents, extensions — these are this year's nouns, and half of them will be renamed or gone. We do not build an identity on a vendor's vocabulary. They are implementations of the transition, not the transition.
Predictions are cheap and everyone has one. What is missing is the boring middle: a repeatable way to take a real product apart, name its intents, tools, state and resources, decide honestly what stays visual, build the thing, and write down what broke. That is the work. It is done one teardown at a time, in public, by people willing to be wrong on the record.