95% of AI initiatives fail, according to MIT

Everyone quotes the MIT number. 95% of enterprise AI pilots deliver no measurable return. Almost nobody quotes what the study said the reason was. It is not the models. So here is my question.

What is the best fix you have actually seen work?