The problem
Brands have hundreds of products, organized in hierarchies — milk, for example, splits into infant, children, adult, and specialty, each with its own sub-attributes:
We work with eCommerce data across millions of SKUs, and the obvious move is to ask an AI model to place each one. The problem is AI hallucinates and is inconsistent on anything ambiguous. Take “PediaSure® 3+ Complete Nutrition for Children Aged 3 to 10” above — is that infant nutrition (usually ages 1–5) or children's nutrition (usually 5+)? One model run says one thing, the next says another. My own team couldn't agree consistently either. What works when you adapt by hand for 1 client falls apart when you try to standardize across 10.
What I did
Instead of asking the model to place a product directly into a category hierarchy in one shot, I redesigned the task as a questionnaire — a path of yes/no (or ‘so-so,’ which routes to a clarifying follow-up) questions leading to the final category:
Yes/no is a far more objective thing to evaluate than ‘which category is this,’ and it gives the model one clear decision point at a time instead of one big ambiguous one — so each pass makes the model more reliable, not less. The most fun part was pulling in people from totally different teams to help shape the questionnaire — some had personal experience with these exact product categories through their own families, which turned out to be genuinely useful domain knowledge.
What happened
The structure held up in a way one-off manual adaptation never could — the same yes/no pathway now runs unattended across a much larger, messier product catalog, flagging category, differentiating pack size, and helping bundle products, and the model keeps getting more accurate with every pass instead of drifting.