Best answer
It does not need AI. Three button questions and a branch per outcome will do it, and the result is predictable in a way a model is not.
The structure:
- Entry: a comment keyword, a story reply, or a button from your bio link.
- Question one, the one that splits your catalogue in half. For skincare that is usually skin type, not budget.
- Question two, the refining one.
- A branch per combination, each sending one recommendation with a button to that product.
- A tag per answer, applied as you go.
Two questions with three options each is nine endings, which is already more than most catalogues need. Three questions is twenty seven and you will not write them all, so keep it at two and let the product page do the rest.
The tags are the part that pays for the build. After a month you know how many people have oily skin and never bought, which is a segment you can message with consent, and you know which combination gets recommended most, which is a stock decision.
Where AI earns its place is the free text case: someone who ignores the buttons and types a paragraph about their skin. An AI step can read that and route it into one of your existing branches. That runs on your own provider key, so you control the spend, and the AI prompts and BYOK article covers how the key and the caps are set.