Scope an AI feature by what happens when it is wrong

When scoping applied-AI work, the demo nearly always works. The real question is the shape of the failure: how often is it wrong, how visible is the error, and who pays for it.

Features where a wrong answer is cheap and obvious — a draft a human reviews — ship quickly. Features where a wrong answer is expensive and invisible need evaluation, guardrails and a way to surface uncertainty before they earn their place in the workflow.