The Interface Tax

I don’t want to learn your app. I want to say what I need and land in the part that does it.

The Tell: I Stopped Going to My Tools

I haven’t opened most of my tools in weeks. Not because they broke. Because I stopped needing to go to them.

It started with the analytics dashboard. Instead of opening it, clicking into the right workspace, finding the cohort view, fixing the date range — I just asked. “Churn by cohort, last quarter.” Number came back. Then it happened with the logs, then the billing portal, then the thing I used to call “checking on the deploy.” One by one, apps I’d spent years mastering turned into places I don’t visit.

I know how this sounds. Dorky. The kind of thing a CTO says right before he buys a margarita mixer he has to return in a week. Every couple of years some founder declares a category dead because his own workflow changed, and usually he’s just narrating a phase. So let me be precise, because the precision is the point: I didn’t replace the work. I replaced the commute to it. The dashboard was a building I’d walked so many times I mistook the hallways for the product.

The interface didn’t die. It got skipped.

The App Was Always a Bundle

Strip the branding and an app is a few things welded together: the data, the rules that govern it, and the screens you poke at to reach either one. We say “an app” like it’s one object. It’s a bundle, and the bundle existed for one boring reason — delivery.

For forty years the only way to hand a capability to a person was to wrap it in pixels. You couldn’t ship “score this customer’s churn risk” to a human — you shipped a dashboard that displayed the score, because a human needs something to look at. Presentation got fused to capability out of necessity. Then we spent two decades convincing ourselves the presentation was the capability. That’s how you get a company that genuinely believes its design system is a moat.

An agent doesn’t want the pixels. It reads the data and the rules and never asks for a screen. So for the first time since the GUI showed up, presentation comes unstuck from capability. “Chat is beating the GUI” is the lazy way to say it. The real word is unbundling, and underneath it the unit of software quietly shifts from the app to the capability it was wrapped around.

Most of Using Software Is Looking for the Door

Watch what you actually do inside a piece of software. You’re hardly ever doing the thing. You’re looking for where the thing lives — scanning the top nav, expanding a tree, opening the third tab, digging into a settings pane four levels down, remembering that this tool files exports under “Reports” and that one calls them “Insights.” Every product makes you memorize its own floor plan. The action takes four seconds. Finding it took four minutes.

Nobody itemizes that on the invoice. It’s still the biggest cost of using software, and it’s the first thing the model deletes. You say where you want to go — “open the refund flow for order 4471,” “split this invoice three ways” — and you’re there, dropped into a workflow built for exactly that. Reads like a pitch deck until you notice it isn’t promising anything new. It’s deleting the wayfinding everyone already hates.

Here’s where the dorky version of the prediction falls apart, though. It does not follow that you’ll chat your way through the task once you arrive. Nobody wants to type a sentence to crop an image or nudge a value until it looks right. So draw the line. Navigation — finding things, deciding where to go — goes to chat. Actionability — doing the thing, with precision and state — does not. Two different layers. Jam them together and you get “chat replaces everything,” which is how a true statement ends up sounding insane.

There’s a test for which side you’re on: is naming the action harder than doing it? Opacity at seventy-three percent, crop framed just so, this exact blue — naming that is harder than dragging a slider, so you want the slider. Don’t know how to do it, or it’s a pain to even describe where you’re headed — that’s chat’s job. Navigation fails the test. Actionability passes.

The thing that stitches the two together already has a name: generative UI, or the “liquid interface.” Chat works out where you’re going, then conjures a real, deterministic surface for the actual work and throws it away when you’re done. This isn’t a whiteboard idea — CopilotKit ships a runtime for it, and A2UI and Open-JSON-UI are real component specs. Chat takes the whole navigation layer. Actionability lives on as surfaces the model renders on demand, and the model picks which side of the line you land on, not a sitemap.

The Standard Showed Up Early

“Someday, sure — once there are standards.” That objection died in November 2025, when Anthropic and OpenAI, who otherwise spend their days trying to bury each other, co-authored one. MCP Apps — formally SEP-1865 — shipped as the first official MCP extension on January 26th, 2026. A server can declare UI that renders as sandboxed HTML right inside the conversation: a chart, a form, a control you can actually use. The pristine workflow you land in is now a resource type with a security model attached. It launched with Figma, Slack, Asana, and Canva rendering inside the model’s window on day one. When your two angriest rivals sign the same spec, you stop calling it a trend. It’s plumbing now.

The Interface Was Never the Moat

The market did this math in public in the first weeks of 2026. People are calling it the SaaS Massacre.

Per-seat pricing was a bet that revenue tracks headcount — more humans logging in, more seats, more money. An agent runs a multi-step job across your stack and never logs in. The seat evaporates, and the revenue stapled to it goes with it. For a decade SaaS charged rent on human operators. Take the humans out and there’s nothing left to charge for.

The sentence getting repeated in board meetings is plain: agents don’t need a UI, they need an API. An agent drives an ugly interface as well as a pretty one, or skips it. Your onboarding flow, your decade of design polish — invisible to the only user whose count is still climbing. When an agent talks to Salesforce, HubSpot, and Pipedrive equally well, your CRM is an interchangeable backend and the switching costs that held up the valuation drop to nothing. a16z put the AI-browser-agent market near $12 billion for 2026, up more than 200% year over year. The capital is voting, and it’s voting for the agent layer.

What’s Actually Defensible Now

If the polish is a commodity, what do you defend? The capability, and the quality of the contract you expose it through. Clean data, and semantics that mean what they say. The boring backend everyone treated as an afterthought behind the “real” product turns out to be the only part an agent can’t route around.

A line going around catches it: your UI is just a reference implementation now. If your data and your API semantics are clean, customers will generate their own front end on top of you — in their design system, rendered by their model. You don’t own the last mile anymore. You own whether the last mile is even possible.

Then there’s the part that should scare people: discoverability. Agents don’t scroll a results page or click a blue link. They query, and they synthesize. One analysis found 96% of B2B companies were functionally invisible in AI-driven discovery — only about 4% held a steady presence when an agent asked a category-level question. You can be excellent and absent at the same time, because your entire go-to-market was built for someone who types into a search box.

SEO is turning into GEO — getting yourself legible and callable by a model instead of ranked for a human. Nobody’s really won that ground yet.

Stop Shipping Pixels

If typing code is already a commodity — and it is — then pushing pixels is next, for the same reason. The machine does the rote part well enough that doing it by hand stops being where the value is.

The job moves up, and honestly it’s a better job. You model capabilities instead of decorating screens. You write contracts a machine can reason about. You design the intent and the single surface that serves it, instead of a forty-screen map nobody asked to memorize. The commodity is the typing. The work is figuring out what’s worth building and saying it clearly enough that a human and a model both understand. I’ve been boring people with that line for years. It’s finally just true.

I won’t pretend this is clean. Nobody’s solved monetization — stuffing display ads into a chat window is a joke, and there’s no replacement yet. Security is worse: prompt injection runs wild across agentic browsers, and Amazon’s January 2026 lawsuit over agentic shopping is the opening shot in a long fight about who gets to act through whose interface. It’s early and half-built, and anyone selling you the finished version is selling you the margarita mixer.

So be exact about the claim. Not “no UI.” It’s the general-purpose, permanent app you log in and live inside that’s dying. The throwaway surface a model renders for one action is just getting started. The interface becomes something your software emits on demand.

Which is why I’m sitting here with everything closed. I don’t want to learn your app. I want to say what I need and land in the part that does it.