Hi I'm Natalie,
a UI/UX designer specializing in UX for AI products and applied AI for UX workflow

Research and what I learned

Talking to drive-thru staff and regular customers surfaced the same pattern: people don’t distrust AI ordering because it’s AI. They distrust it because they can’t tell whether it heard them right. The fix wasn’t a smarter model — it was making the order visible and checkable while it was still being spoken, not after.

Failure and disambiguation

When the model isn’t confident, it asks instead of guessing: “Diet Coke or regular?” Every item in the running order stays changeable, so fixing a misheard item is one correction, not a restart of the whole exchange.

Process and iteration

I started from the competitive review, mapped the happy-path journey. The useful work was mapping the unhappy one: misheard items, background noise, someone changing their mind mid-order. Nearly every feature in this case study came out of that second map. Wireframes went straight into moderated tests with five people ordering out loud to a screen.

The first version of the AI Host spoke every confirmation out loud, on top of showing it on screen. In testing, people found the voice confirmations slower and more annoying than useful — it made the wait feel longer, not shorter. I cut the spoken confirmation and kept the visual one, and let the AI Host's presence (not its narration) carry the reassurance that it was listening.

What I'd do differently

I’d bring in real background noise earlier. I ran the early sessions in a quiet room, which is the one condition a drive-thru is never in — a real lane would almost certainly surface failure rates the quiet tests never showed me. I’d also test the AI Host’s personality against a plain progress indicator, since I never proved the character adds trust on its own versus the transparency features doing the work.

Research and insights

Interviews and desk research pointed the same direction: awareness wasn’t the blocker. People already knew littering was a problem. What stopped them was that picking up trash alone felt pointless — they wanted proof the effort added up to something, and a reason to come back next weekend.

Approach and iteration

With 24 hours on the clock we sketched three concepts in the first two hours and voted as a team before anyone opened Figma — cheaper to kill an idea on paper than in a file. The leaderboard came first. The map came out of a later pivot, once we realized people needed to find an event, not just see a score.

Our first leaderboard ranked users by total pieces picked up. That works for whoever started earliest and nobody else — a newcomer can’t catch someone with a six-month head start, so the ranking stops being a reason to participate and becomes a reminder that you’re behind. We switched to weekly resets, which puts everyone on the same clock every Monday.

Iteration and refinement

We put so much of the 24 hours into research, engagement mechanics and user flows that the visual execution never caught up — and against the finalists, that showed. The judges responded to the thinking, but a competition is also judged on what’s on the screen at hour 24.

Next time, I would plan it out better: lock the flows by hour twelve and give the back half to visual design. Getting the concept right doesn’t count for much if it isn’t finished.