Conversation Design: What UX Researcher Kenneth is learning as AI takes over the interface

For most of my career, the interface was a screen. Buttons, forms, flows, all things you could sketch, test, and iterate on visually. Now a growing share of interactions happen through language, and that's forcing me to rethink habits I'd built up over years of UX work.

The interface disappears

The biggest shift isn't the technology itself, it's what happens to the interface once conversation becomes the primary mode of interaction. There's no layout to fall back on, no visual hierarchy to lean the user toward the right action. The words carry all the weight. Every ambiguity that used to be resolved by a button label or a form field now has to be resolved by phrasing, tone, and structure in the response itself.

That's a different discipline than most of us are used to. In traditional UX, you can often paper over a weak content decision with good visual design. In conversation design, there's nowhere to hide.

Old research instincts still apply

What surprised me most is how much of my existing research process transfers directly. I still start with behavioral data if it exists. I still talk to stakeholders to understand what they think the problem is, and then check that against what users actually do. The psychology underneath conversation design is the same psychology underneath any UX problem: what does someone expect, what do they really need, and where's the gap between those two things.

What changes is the artefact I'm testing. Instead of a wireframe or a prototype, I'm testing a script, a set of intents, or a model's actual output. The moderated sessions I'd normally run to watch someone click through a flow now involve watching someone type a question and react to how the system responds. The medium changed. The underlying discipline of triangulating what people say against what they do hasn't.

Where I've had to unlearn things

A few habits from screen-based design don't hold up well in conversation.

• Assuming one correct path. In a form, there's usually a clear happy path. In conversation, users phrase the same intent a dozen different ways, and a good design has to anticipate that variance rather than funnel people toward one script.
   
• Treating error states as edge cases. In a visual interface, an error state is one screen among many. In conversation, misunderstanding is the default condition you're designing against, not an exception you patch in afterward.
   

• Designing for a single turn. A lot of screen-based thinking optimizes one interaction at a time. Conversation is cumulative. What was said three turns ago still shapes what a good response looks like now, and that memory has to be part of the design, not an afterthought.

   

The research question I keep coming back to

With every new conversation design project, I ask the same question I'd ask on any UX engagement: what is the actual behavioral pattern we're designing for, and do we have evidence for it, or are we assuming it? 

AI makes it tempting to skip that step, because the technology itself feels like progress. But a well-built model wrapped around an untested assumption about user intent is still just a well-built solution to the wrong problem.

The tools are new. The discipline required to use them well isn't.