Where does the AI value gap lie?

Five research efforts, run by different organisations on different populations of companies using different methodologies, converged in 2025 and 2026 on roughly the same finding. MIT's Project NANDA reported that ninety-five percent of organisations deploying generative AI saw zero measurable return, not a low return, zero.1 Boston Consulting Group's September 2025 survey of 1,250 companies found sixty percent generating no material value despite continued investment, with only five percent creating value at meaningful scale.2 RAND Corporation put the outright failure rate of AI projects above eighty percent, roughly double the failure rate of comparable non-AI IT initiatives.3 Morgan Stanley found that only twenty-one percent of S&P 500 companies could point to a measurable AI benefit at all.4

Independent surveys studying different populations do not converge on the same order of magnitude by coincidence. One unflattering survey is noise. Five, run by separate organisations with no shared incentive to agree, are a structural finding. And the structure they describe is not that the underlying technology underperforms. It is that the technology is consistently outperforming the deployment built around it.

Most of the failure happens after the pilot already worked

The more revealing number sits earlier in the pipeline than "did this create value." Gartner's enterprise deployment research found that roughly seventy to eighty percent of AI pilots get initiated, but only twenty to thirty percent ever reach meaningful production deployment, and production rollouts routinely take twelve to eighteen months longer than the pilot's success seemed to promise.5 Separately, S&P Global found that forty-two percent of companies scrapped most of their AI initiatives in 2025, after the pilot stage, not before it.6

This is the detail that gets lost in headline failure statistics. Most of these systems are not failing a technical test. They pass the pilot. The model does what it was asked to do in a contained demonstration. What collapses afterward is everything the pilot didn't have to deal with: integration into a live workflow, data that wasn't curated for a demo, a review process sized for a tenth of the eventual volume, and a definition of success that was never specified precisely enough to measure against once the system left the sandbox.

What the rare successes have in common

The same body of research is unusually consistent about what separates the exceptions. McKinsey's 2025 analysis found that organisations reporting significant financial returns from AI were twice as likely to have redesigned their workflow before selecting a tool, rather than selecting a tool and asking afterward what workflow it should fit into.7 Gartner projects that sixty percent of AI initiatives lacking AI-ready data infrastructure will be abandoned through 2026, which is another way of saying that the precondition for value was set before the model was ever chosen, not after.8

This also explains why broad, organisation-wide "AI transformation" initiatives underperform narrowly scoped ones with almost mechanical regularity across these studies. A narrow target can be specified precisely enough to build a data pipeline and a review process around. A transformation mandate, by design, cannot be, which is why the data infrastructure and the control architecture so often get built after the fact, if they get built at all.

The precise percentage in any single study should be read as directional rather than exact, since "measurable return" and "material value" are not defined identically across them. What is robust, because it appears independently across five unrelated studies, is the order of magnitude and the direction: a large majority of current enterprise AI spending is not converting to value, and the minority that is converting was not spending more. It was specifying what the system had to do, and how its output would be checked, before the model was chosen at all.