Ir para o conteúdo
ZTHEX View on the site →
INSIGHTS · AI ROI TERRITORY

What did the preliminary MIT NANDA report find about AI returns?

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.

Methodology and findings: Project NANDA, The GenAI Divide (2025). Projected cancellation of agentic projects: Gartner, junho de 2025.

// related: the full study, with sources ranked · what a memory layer is · the layer's architecture

TALK TO US ALL READS
NEXT →