"The Chat Bubble Is Not the Whole Product: Matching AI's Interface to the Moment"

The design community, argues UX researcher Victor Yocco, has fallen into "conversational tunnel vision." Because large language models are trained on dialogue, the industry has collectively decided that the chat bubble is the natural home for every AI capability. In a new Smashing Magazine piece, Yocco makes the case that this is a habit rather than a law — and that great UX means matching the interface to the user's context, intent, and cognitive load, so the tool adapts to the person rather than the other way around.

He opens with a vivid scenario: a traveler jogging through a loud airport after a gate change, roller bag in one hand and coffee in the other, needs to ask the airline's AI assistant where to go. The tool fails the input test immediately — it demands the traveler stop, balance the coffee, and type a booking reference into a tiny box. Then it fails the output test: instead of flashing a large, high-contrast gate number, the AI returns a dense paragraph about weather patterns with the gate buried at the very bottom. The airline built a smart tool, but the interface failed the user.

The first half of the problem is input. A blank chat box gives users no menu of what is possible, which breeds choice paralysis — people have to guess what the AI can do and remember the exact phrasing required to make it happen. Yocco's examples are sharp: a data analyst filtering a spreadsheet should not have to become a writer and describe complex logic in a full sentence, and a manager rescheduling a team should not have to narrate a drag-and-drop calendar move as text. Composing a prompt, he notes, is a creative act that forces users to translate a vague thought into a precise command — a linguistic barrier that a slider or a color picker could simply remove.

The second half is output. Text is a serial medium: your brain reads one word at a time to extract meaning, which takes time and effort. That is fine for legal analysis or a nuanced medical history, but it is a tax when the answer is a number or a shape. A doctor checking vital signs needs a clear numerical display, not a narrative; a trader watching for a price spike needs a line graph, not a paragraph. The key point is that visual methods allow parallel processing — you can spot a pattern in a chart in under a second — while defaulting to text turns a quick glance into a reading assignment.

The practical heart of the article is a pair of lightweight tools for deciding modality) before writing code: a Task Audit and an Input/Output Alignment Matrix. The Task Audit walks teams through a "physical reality check" (are the user's hands free? where are their eyes? how loud is the room?), a "cognitive baseline" (how dense is the reading, how reversible are the actions?), and a "handoff map" of how the user moves between environments during a single task.

One insight worth pulling out beyond Yocco's own framing: this is fundamentally a question of who pays the "adaptation load." A do-it-all chatbot looks like a blank slate to the product team but is actually a tax-shift — the system's flexibility becomes the user's job to specify, every single time. Menus, buttons, sliders, and dashboards are pre-computed answers to the question "what can I do here?" The chat box forces every user to re-derive that menu from scratch. Modality matching is really just relocating cognitive work back to the machine, where it is cheap, instead of leaving it on the human, where it is expensive.

Second, modality-first design is quietly an accessibility win. When you stop defaulting to a text box and instead ask "which sense and which limb is actually available right now," you end up designing for the traveler with coffee in one hand, the warehouse worker in thick gloves, and the screen-reader user all at once. Yocco flags accessibility explicitly, but the deeper point is that designing for the constrained case tends to produce interfaces that work better for everyone — the classic curb-cut effect applied to AI.

There is also a timing angle worth noting. As the industry shifts from chatbots to agents — software that acts on a user's behalf rather than merely answering — modality becomes more consequential, not less. An agent that is rebooking a flight or reordering inventory is not really "conversing"; it is executing. The interface it presents should look less like a message thread and more like a control panel with confirmations, status, and undo affordances. Yocco's matrix is a useful bridge from today's chat-centric demos to tomorrow's agentic products.

To be clear, this is not an argument against chat. Yocco is explicit that the text interface is a viable and powerful option for many tasks — it is one tool in an expansive toolkit. The error is defaulting to it unthinkingly because the underlying model happened to be trained on dialogue. The right question is never "chat or not chat," but "what is this user trying to do, where are they, and how much cognitive effort do they already have in reserve?"

The broader framing ties it all together. For the past couple of years the industry's energy has gone almost entirely into training smarter models, and Yocco's reminder is that a brilliant model wrapped in a lazy text interface still fails. Jakob Nielsen has been making a parallel argument at the Nielsen Norman Group — that intent-based outcome specification is the first genuinely new UI paradigm in 60 years, replacing the command-based interaction we have used since the dawn of the graphical interface. Both point in the same direction: the next frontier of AI quality will not be measured in benchmark scores, but in how gracefully an interface meets a person in the middle of a real, messy moment.