In 2025 a figure appeared that became a headline: the vast majority of organizations get no return from generative AI despite massive investment. It should be cited carefully —it comes from a preliminary MIT report (Project NANDA, Challapally et al., 2025), not peer-reviewed, and the correct figure is about organizations with zero return, not “95% of pilots with no impact”, as it circulated—. But what matters is not the number. It is its diagnosis.
The report attributes the failure not to the models or the budget, but to an organizational learning gap: the systems that get deployed do not retain context or learn from operations. In other words: the bottleneck is sociotechnical, not technological. It is, almost in those words, the thesis I have been holding for years —the problem was never the technology; it is the organization that cannot absorb it—. The divide separates the firms that treated the matter as an engineering problem from those that treated it as one of organizational learning. The bridge lives on that second side.
Where it does pay off: in the person
That most do not capture value does not mean AI is useless. It means the value plays out in a precise place —the person— and there the evidence is hard. A study published in a top journal (Brynjolfsson, Li and Raymond, 2025) measured 15% more cases resolved per hour in a support center, with a revealing distribution: the least experienced workers improve far more than the experts. It is AI leveling up those who knew least, rather than merely speeding up those who already knew. It is what I call the multiplier with a sign: AI amplifies what it finds, and the sign is set by the human reading that steers it.
The distance between “we deployed AI” and “we captured value”
The contrast that closes the argument comes from another 2025 survey: according to McKinsey, though nearly all organizations deploy AI, only about 6% are the ones capturing a real, measurable impact on their bottom line (an AI-attributable impact of at least 5% of EBIT). That distance —between “we deployed AI” and “we captured value”— is exactly the one that separates the vanity metric (models deployed, chatbot queries, “AI usage”) from the value metric (an hour freed, a better decision, a gap closed).
The macro figure does not replace an honest measurement dashboard: it justifies it. If at global scale most do not capture value, measuring well stops being tidiness and becomes the difference between being in that 6% or the rest.
What to do with this
- Measure the learning gap, not just buy models. The MIT report is clear: what is missing is not a better model, it is an organization that retains context and learns from operations.
- Set the unit of value in the person, before building. What concrete decision changes if this system works? Without that question, the deliverable is decoration.
- Tell vanity from value from day one. Agree on two or three measurement axes at the start —do not pick the best-looking metric afterwards—.
One caveat I do not drop: these are global evidences from other contexts. They illustrate mechanisms and give scale; they do not promise a result to a local SME, whose anchor remains regional evidence (CEPAL, nadIA, ILIA 2025). But the underlying message travels well: the 95% that captures no value is not waiting for a better model. It is waiting for someone to treat the problem as what it is —sociotechnical— and not as a purchase.
The framework is developed in The sociotechnical bridge and in My approach. Related essays: The productivity that dims tomorrow’s experts · Digital transformation is not a technical problem.