The AI feature we talked a client out of
A conversational product finder, fully specified and budgeted. We recommended against it, wrote down why, and the client agreed. Six months on, here is the follow-up.
The brief was clear and the budget was real: a conversational assistant on a home goods storefront, able to answer questions about products and guide people to a purchase. It is the sort of project a studio is supposed to be delighted by.
What the data said
We spent two days of the framing week in their analytics before designing anything. The pattern was unambiguous. Customers were not failing to find products through lack of guidance — they were failing because 38% of the catalogue had no dimensions recorded, and dimensions were the single most common filter attempt.
An assistant sitting on top of incomplete data would have confidently answered questions with information that was not there. The most likely outcome was a well-reviewed feature that quietly increased returns.
What we recommended instead
- A data remediation project to complete the attribute coverage, mostly supplier chasing rather than engineering.
- A rebuilt faceted search using those attributes.
- Revisiting the assistant afterwards, with the underlying data good enough to support it.
That recommendation was worth roughly a third of the original budget. We wrote it down and sent it anyway.
Six months later
Search-led revenue is up substantially and returns attributable to size and fit are down. The assistant is now back on the roadmap for next year, and this time it will have something reliable to talk about.
The general point is not that AI features are bad. It is that a language model on top of bad data produces confident, fluent, wrong answers — and that is worse than no answer at all.
