Prompts aren't interfaces: why designers still matter in the age of AI agents

The pitch was seductive: no more complex menus, no more navigation hierarchies, no more multi-step flows. Now the user just types what they want. The prompt is the interface.

Except it isn’t.

What we’re seeing in practice, as agentic AI tools mature, is a return to UX fundamentals that many thought were retired. Not because the technology failed. But because a conversational interface alone doesn’t solve what design solves.

The myth of the universal interface

The idea that a free-text field replaces intentional design assumes something false: that users know exactly what they want and can articulate it clearly.

In reality, most users don’t know what’s possible, don’t understand the system’s limits, and lack the vocabulary to describe what they need. That’s not user failure. That’s the normal condition of using a tool to solve a problem, not to understand the tool itself.

When you put someone in front of a blank text box with the instruction “ask anything,” you’re not simplifying the experience. You’re creating a blank-canvas problem—the same paralysis that freezes designers facing an empty file. The person doesn’t know where to start, doesn’t know what to ask, and often gives up or underuses the system.

What conversation can’t solve

Conversational interfaces work well for some tasks. They systematically fail for others.

Where conversation works

• Direct factual questions
• Tasks with a clear objective
• Users experienced in the domain
• Single, isolated interactions

Where conversation fails

• Exploring possibilities
• Comparing options
• Multi-step workflows
• New users to the system

If you need to compare three product options, a linear conversation is a terrible way to do it. If you’re exploring what a system can do for you, “ask anything” is less useful than a well-organized menu.

This isn’t a limitation of AI. It’s a limitation of the format. A visual list communicates options in parallel. Conversation is sequential by nature. For some tasks, sequential is the wrong path.

Affordances still exist

Don Norman popularized the concept of affordance in design: a property of an object that signals how it can be used. A doorknob suggests you should turn it. A button suggests you should press it.

A blank text box suggests… what?

Well-designed systems with AI are reintroducing affordances that seemed outdated. Suggested action buttons. Prompt examples. Visible scope constraints. Task templates. Not because AI needs them, but because the user does.

Claude from Anthropic shows “prompt starters” across its interfaces. ChatGPT displays continuation suggestions. Notion AI presents contextual actions based on content type. None of these systems rely on the text box alone. All layer design on top of conversation.

Design as curation of possibilities

One of the central functions of interface design is curating possibilities. Not showing everything the system can do, but showing what makes sense now, for this user, in this context.

That’s the exact opposite of “accept any input.”

When an AI agent can theoretically do anything you ask it to, the design question shifts. It’s no longer “what does the system allow?” It becomes “what helps the user succeed?” And that question demands the same tools as always: user research, journey mapping, information hierarchy, usability testing.

The classic patterns that came back

Anyone tracking AI product evolution sees an interesting movement: the reintroduction of UX patterns that seemed obsolete.

Progressive disclosure—revealing complexity gradually—is everywhere. Systems that start simple and reveal advanced options as the user demonstrates need.

Wizards and guided flows, which many considered outdated, reappeared as “flows” or “workflows” in AI automation tools. Because breaking a complex task into manageable steps still beats asking the user to describe everything at once.

Visual feedback about system state, one of Nielsen’s heuristics, gained new importance. When the agent is “thinking,” the user needs to know. When there’s uncertainty in the response, it needs to be visible. Transparency about AI’s internal process is a design problem, not an engineering one.

What this means for builders

If you’re building something with AI, the temptation is to let conversation do the heavy lifting. It’s quick to ship, feels modern, and the tech actually works.

But the tech working doesn’t mean the experience works.

• Does the user know what they can ask the system to do?
• Are there clear paths for the most common tasks?
• Does the system communicate its limits before the user hits them?
• Is there visual feedback during processing and uncertainty?
• Can new users succeed without training?

If the answer to any of these is “no,” you have a design problem. Not a language model problem.

Design isn’t about prettifying

There’s persistent confusion between design and aesthetics. Design isn’t “making it look nice.” It’s solving usage problems. It’s creating clarity where there was confusion. It’s cutting friction where there was abandonment.

Agentic AI doesn’t eliminate these problems. In some cases, it amplifies them. A system that can do many things needs more design, not less, so the user navigates those possibilities without getting lost.

Conclusion: the prompt is just one layer

The conversational interface is a tool. One layer of interaction. Not a substitute for intentional design.

The products that win in this phase aren’t the ones with the most capable model. They’re the ones that translate that capability into navigable experience. That takes research, information architecture, visual hierarchy, testing with real users. It takes designers.

The blank text box will keep existing. But it won’t be enough.