Case Study

How AI Personas Discovered Seasonal Color Theory

The defects that human testers miss, but AI personas catch.

By Sarah Oates Dec 1, 2025 10 min read

The Project

The brief was straightforward: build a feature that helps women discover which colors complement their appearance and personal style. The target demographic was women aged 20-35, fashion-conscious, actively shopping for clothing online.

This was our side project—the shopping app that eventually led to everything else.

The technical implementation wasn't particularly complex. Color theory has been well-studied. We had data on skin tones, undertones, color harmonies. The UX challenge, we assumed, was equally manageable: show users colors that work for them. Provide visual examples. Make it easy to shop.

We were catastrophically wrong about how our target users actually think about color.

Round 1: The "Technically Correct" Approach

Our initial implementation followed what seemed like logical UX principles. The interface had a color picker organized by hue, filters for skin tone and undertone, and product recommendations.

Clean interface. Technically sound color theory. Responsive design. Production-ready.

The Human Testing: Everything Is Fine

We showed it to my alma mater sorority. Real women in our target demographic. Their initial reaction? "It works fine. I can see the colors and pick things."

No complaints about bugs. No broken interactions. But also no enthusiasm. Just "fine." We didn't realize how damning "fine" was.

The Persona Testing: Everything Is Wrong

Then Anthony built an AI sorority sister. We fed it data about Gen Z women interested in fashion and asked it to evaluate our color matching feature.

The response was immediate and specific:

"I'm confused by this organization. I don't shop for 'blue' or 'green.' I think in terms of my color season. Am I a spring? Winter? Also, I want to see how these colors look in actual outfits on people with similar coloring to mine, not just products on white backgrounds."

We went back to the human sorority sister. "Does this seasonal color thing resonate with you?"

"Oh my god, YES. That's exactly how I think about color."

"Why didn't you tell us?"

"You asked if it worked. It worked. I just didn't like it."

The Defect Nobody Saw

The defect wasn't technical. It was conceptual. We'd built a technically perfect implementation of the wrong mental model.

Implicit knowledge like seasonal color theory is invisible until you make it explicit. Our human tester knew it but didn't think to mention it because she was in "tester mode." The AI persona knew it because it was trained on data from that demographic.

Round 2: Rebuilding Based on Persona Feedback

We completely restructured the feature:

  • Primary navigation by seasonal color type (Spring, Summer, Autumn, Winter)
  • Visual examples showing colors on models with similar skin tones
  • "Find My Season" quiz

We showed the revised version to the same sorority sister. Her reaction: "Oh wow, this is so much better."

The AI persona had told us exactly this on the first try.

What We Learned

  1. Human Testers vs. Persona Agents: Human testers find bugs. Persona agents find conceptual mismatches and terminology gaps. You need both.
  2. Demographic Distance Matters: We weren't women in our twenties thinking about color theory. We were building for a demographic we weren't part of.
  3. "Working" Isn't the Same as "Right": Traditional QA asks "Is it broken?" Empathetic development asks "Is it right?"
That shopping app became a 2/10 that taught us everything. And seasonal color theory became the example we use to explain why empathetic development matters. Because "fine" isn't good enough.