UI/UX for an AI-First Platform: What Building Expona in Next.js Taught Me About Design
Tracy Thayne7/7/2026·6 min read
The most honest user test I ever ran lasted about eleven seconds. Early in building Expona, we did what almost every AI product did in that era: we put a chat box on top of everything and called it an interface. I sat next to a design partner, a sharp marketing consultant who runs campaigns for a living, and watched her stare at the blinking cursor. She typed nothing. Then she asked me the question that reshaped our roadmap: "What am I supposed to say to it?"
She was not confused because she lacked skill. She was confused because we had shipped her a blank page and called it a product. That moment taught me the first law of AI-first design: every ounce of effort you save the model, you quietly transfer to the user, unless the interface absorbs it instead.
This problem is about to hit almost every software team at once. Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025 (Gartner). At the same time, the payoff for getting design right has never been better documented: McKinsey tracked 300 public companies over five years and found top-quartile design performers delivered 32 percentage points higher revenue growth and 56 points higher shareholder returns than industry peers (McKinsey). And the trust problem is measurable too: 53% of consumers say they distrust AI-generated answers, and 61% want the ability to toggle AI features on or off (Gartner). AI features are flooding in, users do not fully trust them, and design is the variable that decides which side of the value gap you land on.
The Blank Box Is a Design Failure, Not a User Failure
Chat was the fastest way to demo AI. It was never the best way to use it.
A prompt box is a blank page with better marketing. It asks the user to know what the system can do, know how to phrase the request, and know what good output looks like, all before anything happens. That is three expert skills disguised as an empty text field. Most users respond the way my design partner did: they freeze, type something timid, get a mediocre answer, and quietly conclude the product is not for them.
Subscribe
Get notified by email when we publish a new post. No spam, unsubscribe anytime.
The fix we landed on at Expona was to lead with the work, not the conversation. Instead of asking "what do you want?", the interface proposes: here is the persona this data suggests, here is the campaign this persona implies, here is the draft this campaign needs. The user's job shifts from authoring instructions to reviewing judgment, which is a job humans are actually good at. AI-first does not mean chat-first. It means the system moves first.
Design for Trust: Show the Work
That Gartner distrust number is not a PR problem. It is a design brief.
When half your users doubt AI output by default, the interface has one overriding job: make the system's reasoning inspectable. In practice that means every generated artifact in Expona carries its receipts. Which sources fed this persona. Which past campaign informed this recommendation. What the system was told, and what it inferred. A user who can see where an answer came from can disagree with it intelligently, and a user who can disagree intelligently starts to trust.
The second half of trust is control, and the 61% who want an on/off toggle are telling you exactly what control looks like. Every AI behavior in the product should be adjustable, overridable, and reversible. Nothing erodes confidence faster than a system that acts on your behalf and offers no way to decline. We treat "undo" as a trust feature, not a convenience feature.
The Interface Is Becoming an Output
The deeper shift is that the UI itself is starting to be generated. Nielsen Norman Group calls this generative UI: interfaces assembled in real time for a specific user and task rather than designed once for everyone.
I find this direction credible, and clarifying. If screens can be generated on demand, then screens are not the moat. What decides whether a generated interface is useful is the same thing that decides whether generated content is useful: the context behind it. A generated dashboard built from thin context is just a prettier blank box. I made this argument about content in Context Is the Whole Game, and it applies without modification to interfaces. The durable layer is the accumulated, structured knowledge of the business.
What Next.js Taught Me About Shipping AI UX
Some lessons only show up in production, and building Expona on Next.js taught me a few worth passing on.
Streaming is a design decision, not a performance detail. An AI response that arrives word by word feels alive; the same response after eight seconds of spinner feels broken. We stream everything we can, and we design skeleton states that say what the system is doing, "retrieving your campaign history," rather than showing a generic shimmer. Server components earn their keep in an AI product because context assembly, the gathering of customer data and history that grounds every generation, belongs on the server, close to the data, leaving the client light and responsive. And motion needs restraint: AI products are tempted to animate everything because everything is dynamic, but respecting users who prefer reduced motion, and defaulting to calm, keeps the product from feeling like it is performing intelligence instead of delivering it.
None of this is exotic engineering. It is the discipline of treating latency, provenance, and control as first-class design materials, because in an AI product, they are the experience.
The Takeaway
The AI-first products that win will not be the ones with the most capable model behind the curtain. They will be the ones whose design absorbs complexity instead of transferring it, shows its reasoning instead of asserting it, and gives users control instead of asking for faith.
The blank box era is ending. The interface is becoming a projection of the context layer underneath it, which means design and data strategy are now the same conversation. Build the context layer well and the interface almost designs itself. Build it poorly and no amount of polish will cover for a system that has nothing specific to say.
Eleven seconds of silence from one honest user taught me more than any framework. Watch where your users freeze. That is where your design is writing checks your context cannot cash.
Tracy Thayne* is the founder of Expona, an AI-powered operational intelligence platform for B2B marketing. Read the Expona founder story or subscribe to the blog (below) for weekly insights on context, AI, and the operating model of the next decade.*
