According to an MIT study covering 300 public deployments, 52 organisational interviews and 153 executive responses, the difference between what works and what doesn't isn't model quality or regulation: the tools that fail don't retain context, don't adapt and don't improve with use.
If it's not the model, what is it?
It's the approach. A generic tool performs well in a demo and breaks in the real workflow, because the real workflow has exceptions — and exceptions need context the tool doesn't keep.
Why did the budget go to the wrong place?
Because it followed visibility. 50 to 70% went to sales and marketing, where the return turned out lower; the return showed up in operations and back-office, the less visible areas.
What does the study itself recommend?
Fund the memory layer and demand pricing by learning milestone, not by seat licence. It's a description of a product that retains context — written by a third party.
The study didn't conclude that AI doesn't work. It concluded that it works where there's context — and most companies bought the tool before they had context to give it.