How I decided where to put AI in an educational product (and where not to)

The full article is being translated from the Spanish original. What follows is the short version: the shape of the decision, without the numbers.

The decision

I built an exam-prep platform where AI powers exactly the features that justify their cost, tiered by pricing plan, and stays out of the ones that would burn margin for marginal value.

What each interaction costs

Every AI feature was priced per interaction and mapped against the plan that carries it. Features that couldn’t cover their own cost at scale didn’t ship as AI features.

What a simpler system could do

We deferred retrieval infrastructure, not because we couldn’t build it, but because the numbers said “not yet.” A simpler system delivered most of the value while the economics matured.

What waiting cost, and what it didn’t

Deferring was a decision with a review date, not a rejection. The framework behind these calls is published at apat.io/framework.