Open any mature SaaS product and count the navigation items. Forty is typical: dashboards, lists, detail views, settings pages nested three deep, each designed for a persona that visits weekly. This structure is not a law of software. It is a filing cabinet inherited from paper offices, preserved because every feature ever shipped needed somewhere to live, and the sidebar was the cheapest available real estate.

The filing cabinet made sense when software was scarce and attention was cheap. A trained operator, sitting down for hours, could learn forty tabs the way a clerk learned forty drawers. But the ratio has inverted: software is now abundant and attention is the scarce input. Nobody has hours. Everybody has a task — refund this order, brief this client, reconcile this invoice — and the forty-tab menu is forty wrong answers standing between the human and the one right action.

What replaces it is already visible in pieces. The command palette (Cmd-K) was the first crack: users stopped navigating and started naming what they wanted. AI chat was the second: users stopped naming menu items and started describing outcomes. The third step, now arriving, dissolves the menu entirely — the interface compiles around the stated intent, performs it, shows the receipt, and vanishes. No sidebar to learn, because there is nothing to navigate to.

The economics finally work because inference got cheap and data got local. Generating a bespoke interface per task was absurd at $20 per thousand tokens with a three-second round trip to Virginia; at sub-cent costs on a local model with millisecond access to your own database, it is cheaper than maintaining the forty tabs. The constraint flipped from compute to trust: users must believe the compiled interface did the right thing, which is why every ephemeral UI ships with an audit receipt — what was touched, what changed, how to undo it.

Critics raise the obvious objection: discoverability. Menus teach; blank prompts intimidate. True, and the answer is the same one command palettes found: type-ahead suggestions drawn from your own recent tasks, plus a shrinking set of pinned rituals for the five things you do daily. Discovery moves from browsing architecture to recalling history. Your software learns your verbs instead of teaching you its nouns.

For existing products, the migration is a strangler pattern, not a rewrite. Instrument which three tasks drive eighty percent of sessions, ship intent-first flows for exactly those, and watch support tickets to find the next three. The old menu stays parked behind a single “browse everything” affordance while the compiled layer absorbs daily volume one verb at a time. Within a year the sidebar becomes what the basement already is in most houses: somewhere things are stored, not somewhere anyone lives.

The admin console does not disappear entirely — it demotes to infrastructure. Somewhere beneath the ephemeral layer, the forty drawers still exist as an API surface for auditors, power users, and edge cases. But the daily interface becomes a conversation with receipts rather than a building you commute to. Square footage collapses; intent density rises.

Builders should take this personally. Every settings page you ship in 2026 is a small admission that your software could not figure out what the user wanted. The winning products of the next decade will be judged not by feature count but by how rarely the interface is visible at all: state the intent, watch it compile, check the receipt, get on with the day. The filing cabinet had a good forty-year run. Time to burn it for warmth.