Chat Is the Wrong Cage: Rethinking AI Interfaces Beyond the Text Box
Reportedly around sixty percent of people who open the chat window never send a single message. They watch the cursor blink, feel the blank box asking them to do the product team's design work, and close the tab.
That number, reportedly clustered across a range of AI features, should have ended the "let's add a chatbot" reflex on its own. Instead it launched a genre.
Count the essays. "Chat is the wrong interface for AI." "Chat Is the Wrong Interface: Why Your Agent Shouldn't Be a Text Box." "Chat Is the Wrong Interface for AI Agents." "Chat Is the Wrong Default for AI Products." Four titles, near-identical in their headline framing, at least two from the last few months.
When a contrarian take becomes a genre, the contrarianism is over. The critique has quietly become consensus, and the interesting question has moved somewhere else.
So this piece is not here to re-litigate whether chat is limiting. Last week we traced how legacy UX collides with the chat box — the friction, the information architecture strain. This week the question is sharper: should chat be the primary paradigm at all, and if not, what decides the replacement?
Start with what everyone now agrees on, because the diagnosis really is shared and really is well-evidenced.
The blank box is a discoverability failure, not a user failure. A search field with no hints tells you nothing about what it can find. A chat box is worse, because the space of things you might say is unbounded and the space of things that will actually work is invisible.
Users face a double bind: they don't know how to phrase the request, and they don't know what the system can even do. Faced with both at once, most people freeze. Hence the 60% — versus far higher engagement when the same capability is wrapped in a designed empty state with examples and one-tap starting points.
The tell is what teams do next. Suggested prompts, example chips, "try asking…" placeholders, slash commands, quick-reply buttons. Each is a small confession that pure conversation wasn't enough. At some point the honest move is to notice you're reinventing a menu one band-aid at a time, and build the menu.
The second half of the diagnosis is about output, and it's more dangerous. A chat transcript is a decent record of a conversation and a bad representation of a process.
When an agent plans a task, calls three tools, edits two files, and errors on the fourth step, all of it flattens into a scroll of prose. There is no plan you can see, no progress you can track, no clean place to intervene before the irreversible step runs.
The sharpest version of this is a single sentence worth memorizing. The agent is 95% right per turn, the transcript looks reasonable, and the one wrong action — the email that got sent, the record that got deleted, the refund that got issued — "scrolls by in the same font and the same gray bubble as everything else."
The one wrong action — the email that got sent, the record that got deleted — scrolls by in the same font and the same gray bubble as everything else.
Chat makes capability illegible and mistakes illegible with the same design choice: everything is text, everything is equal weight, everything is after the fact.
That is the consensus. It is correct. And it has produced a chaos of cures.
Here is where the field fractures. Everyone agrees on the disease; nobody agrees on the medicine, and the proposed medicines aren't even the same kind of thing.
Clicky proposes a voice-plus-point layer that sits over any SaaS — press, ask, listen, point. Instead of paraphrasing a location, it draws a halo on the exact DOM element you need, and the same gesture works across every tool you touch. Its frame is activation to autonomy: collapsing five or six onboarding layers into one on-demand interaction.
Adaline proposes a four-verb taxonomy. The verb surface puts AI behind a button you already press — Cursor's cmd-K, where the selected code does the prompt engineering for you. The generative canvas turns output into an artifact you edit in place. The delegated agent takes the task and reports back. The ambient listener watches and offers.
Adaline's evidence is Apple's own WWDC 2026 keynote, which reportedly shipped two contradictory things in one hour: a dedicated Siri chatbot app, and everything else — a macOS screenshot tool that quietly adds calendar events, Shortcuts built from plain language, a camera that answers about what it sees. Read as one story, Apple looks confused. Read as two, the picture sharpens: Apple added chat to its inventory, but the real product work happened somewhere else.
Then there's the agent-operations camp, which wants a dashboard, not a metaphor at all — trace timelines, permission-review queues sorted by risk tier, memory browsers, context-budget meters that work like a fuel gauge. This is chat replaced by a flight-data recorder.
And Smashing Magazine proposes a matrix. Match modality to context, intent, and cognitive load, via a Task Audit and an Input/Output Alignment Matrix.
Its anchoring scene is the one you should keep in your head. A traveler jogs through a loud terminal after a gate change, roller bag in one hand, coffee in the other. The airline app forces them to stop, balance the coffee, and type a long booking reference into a tiny box.
Then it fails again on output: instead of flashing a large, high-contrast gate number, the AI returns a dense paragraph about the atmospheric weather patterns causing the delay, with the actual gate number buried at the bottom. The input demanded dexterity the traveler didn't have; the output demanded focus they couldn't spare.
Now line these up. Voice halos. Four verbs. Operations dashboards. Alignment matrices. These are genuinely useful vocabularies — and they are not interchangeable.
A verb surface and a delegated agent solve different problems. A halo drawn on a live element and a trace timeline for an overnight run answer different questions. Nobody has published the selection criteria that tell you which shape fits which intent.
That gap — not the death of chat — is the actual unsolved problem. Adaline half-admits it by giving you "a vocabulary and a quick diagnostic you can use tomorrow." Smashing goes further with an explicit matrix. But a diagnostic from one blog and a matrix from another are not a discipline.
Which is why the real maturity signal isn't another manifesto. It's a reference book.
Louise Macfadyen's book, out in translation August 2026, does what the essays gesture at: it codifies non-deterministic UX into named patterns rather than improvising them.
Its spine is a structure, not a slogan: input, computation, output. It names the agentic patterns the essays keep circling — reflection, tool use, planning, multi-agent coordination, ReAct. It gives them design principles: plan disclosure, prioritizing what matters, checkpoints, human-in-the-loop.
This is the difference between a genre and a field. A genre produces four essays with the same title. A field produces a taxonomy where reflection and ReAct are distinct named things you can choose between, and where "which interface" becomes a decidable question instead of a vibe.
There's a coda worth naming, because not everyone is playing. Some practitioners are rationally opting out of the churn entirely — closing the tab, as one put it, on all the AI hype as unnecessary cognitive load, normalizing turning the knob the other way.
That's a legitimate stance toward the noise. But it's not a stance you can take toward the work, if the work is designing the interface layer — which has plainly lagged model capability while everyone argued about proxies and pixels.
The blank box was never the enemy. It was just the easiest thing to ship, mistaken for the right thing to ship. The designers who win the next round won't be the ones with the sharpest obituary for chat.
They'll be the ones who can look at a traveler running through a terminal, or a 95%-right agent about to send an email, and say — with a taxonomy behind them, not a hunch — exactly which interface this intent requires.