As more AI products launch, saying “it works” will mean very little. We will need better ways to test reliability, bias, usefulness, and failure before systems reach people.