The AI is a pile of applied stastistic models. The humans in charge of training it, testing it and acting on its input have full control and responsibility for anything that comes out of it. Personifying or otherwise separating an AI system from being the will of its controllers is dangerous as it erodes responsibility.
Racist cops have used “I go where the crime is” as an exuse to basically hunt minorities for sport. Do not allow them to say “the AI model said this was efficient” and pretend it is not their own full and knowing bias directing them.