Users need to be a central part of your AI feature design–especially evaluating AI output

Users need to be a central part of your AI feature design–especially evaluating AI output. A story from early in my career:

My first-ever AI/ML feature project was conceived and run entirely by engineers. We were working on a platform where users entered a lot of data to accomplish information tasks, such as planning trips, or researching large purchases. These tasks could be fairly involved, and information would get lost.

The feature concept was simple: use AI to cluster information together by task, so that users could see everything in one place and receive appropriately tailored new content.

The engineering tests showed that it worked perfectly–recipe ingredients were clustered together (“meal planning”), electronics reviews in a different cluster (“researching electronics”), and so on.

But user testing told a different story.

The AI clustering, though logical, didn’t map to how our users thought about their tasks:

  • 🐱 Users weren’t just “meal planning,” they were doing multiple other food-related tasks such as planning outings, or even researching pet food safety.

  • 🕹️ They weren’t “researching electronics,” they were looking for vintage gaming systems they could refurbish with a specific type of hardware.

The AI had grouped the data in ways that appeared useful, but it made real users’ workflows harder, not easier. If we’d shipped this feature we’d have wasted engineering time and frustrated our users.

This story goes to show that siloed work leads to blind spots in AI feature design–you need more diverse perspectives.

My engineering team did everything right, but no amount of engineering testing could have caught these problems. It took rigorous experimentation with actual users about their real-world tasks to see the AI’s limitations. And this illustrates why AI feature design requires a new level of partnership between UX and engineering.

🧩

I’ve been getting more requests lately for training on how to effectively research AI products and features, and it all starts with a deeper integration between UX research and engineering.

So I’ve been building a new workshop to help folks make this transition in their own work. It’s designed for researchers, designers, PMs, and founders who want to learn how to involve users in AI feature design and iterate on AI outputs in ways that ensure users can complete their goals.

If this resonates with you, I’d love to hear your thoughts:

  • ❓What’s your biggest question about researching AI features?

  • ❓If your team is already building AI features, what’s the toughest challenge you’re facing with research and strategy?

  • ❓If you’re wanting to do more AI feature research but hitting a wall, what’s your biggest obstacle?

Drop me an email, or let’s grab 30 minutes to chat!