← All case studies
AI can’t agree with itself.
Brands have hundreds of products in category trees. Milk alone splits into:
Across millions of SKUs, the obvious move is to let AI place each one. But on anything ambiguous, it’s inconsistent. Is PediaSure® 3+, for children aged 3 to 10, infant or children’s nutrition? One run said one thing, the next said another. My own team couldn’t agree either.
Turn one big question into small ones.
Instead of one-shot placement, I redesigned the task as a questionnaire of yes/no questions, with “so-so” routing to a follow-up. Yes/no is far easier to judge than “which category,” and each step gives the model one clear decision.
The best part: people from totally different teams helped shape the questions. Some knew these products from their own families, which turned out to be real domain knowledge.
It gets better with every pass.
The same yes/no path now runs unattended across a much bigger, messier catalog, and gets more accurate with every pass instead of drifting. It went from 1 client to 10+, doing three jobs: