What happens when a totally AI-powered project leads you straight back to a solution only a human can solve? That’s exactly what happened when we tried to implement an AI-powered product selector for a client.
We wanted to upgrade a years-old product selector on our client's website to an AI-powered version. Instead of forcing users through a rigid decision tree, we envisioned a natural-language interface where users could simply describe their needs or pain points, and the AI would match them to the right product(s) based on the site’s existing content. A natural-language layer on top of their product data could make finding products efficient, scalable, and smart.
Except it didn’t work.
Time after time, the AI returned inconsistent product matches for the same user queries. We reengineered the system prompt, refined the architecture, and iterated multiple times—but the results never improved meaningfully.
That’s when I realized the problem wasn’t technical. It was human.
The AI wasn’t misinterpreting users; it simply couldn’t find meaningful distinctions between the products themselves. AI can only reason with what it’s given and what it found was content that failed to help each product stand for what it truly offered. The AI became a mirror, reflecting not a flaw in the model but a weakness in how the brand expressed its own product differentiation.
Here’s the subtle insight: no user will ever tell you that they browsed your site and couldn’t find the right product. They might not even realize it; they just leave or move to a competitor with clearer messaging. Analytics won’t surface this either: a long session duration might seem positive, but it could actually represent confusion. The AI, however, had to parse the content at scale and in doing so, it surfaced this gap in a way no other method could.
At one point, the stakeholder wondered if perhaps an AI-powered selector wasn’t a good fit for their complex products. My response: 0-click search and AI-first discovery are already here. If your content can’t help an AI distinguish between your offerings, it won’t help human users or algorithms elsewhere either.
A tool like this becomes a diagnostic lens, revealing how people describe their needs, what they struggle to find, and what language they use to search. That insight is gold for improving messaging, clarity, and product positioning. In fact, every product owner should test whether their site content is truly AI-ready by building a selector like this based solely on existing copy.
This experience revealed an irony: working with AI often brings us back to fundamentals—clear thinking, distinct language, and human insight. Sometimes, the smartest AI solution is realizing what only great content can do.