Only 28% of AI projects deliver ROI. The number behind the number.
Only 28% of AI projects deliver the expected ROI — and 20% fail outright, per Gartner (2026). The number makes an easy headline. What it hides is more useful than what it shows.
The lazy reading is “AI in retail does not work”. The honest reading looks at the 28% that deliver, and asks what they did differently.
What the study measures
It measures projects that did not deliver the return promised at approval. It does not measure technology that does not work. It measures the distance between what was promised on the slide and what showed up in the operation.
That distinction matters. Most of these projects had a model that ran. What was missing came before and after the model, not in it.
The number behind the number
Add two others. McKinsey (2026) estimates 88% of companies already use AI in at least one function. And MIT (2025) finds 95% of generative-AI projects deliver no return: only 5% capture substantial value.
Read together, the three tell the story. Almost everyone is doing it. Most do not deliver the promise. And a small minority captures real value. The difference is not in access to the technology, which is the same for everyone. It is in the criterion.
Of those that deliver, the pattern repeats: they measured the success criterion before the first line of code. They knew, in week one, which number would say whether it was working.
The operational distance
Between “88% are doing it” and “5% capture value” there is a gap, and it has a name. It is the absence of a written criterion measured early. The project starts from the technology, runs, dazzles in the demo, and nobody can say whether it changed the operation, because nobody defined what to watch.
It is the same gap that sank Walmart’s Jetblack and that separates internal team from implementer on the first project.
This quarter’s decision
The number does not change the year. It changes one requirement at the next budget approval: no AI project enters without a success criterion written in one sentence and measurable in the first week.
Think about the AI report that circulated through your leadership and became no decision. Almost all of them cite the failure percentage and none say how to land on the right side of it. That is the one worth translating.
Send me the number pushing you to invest in AI. In one hour I will send back a one-page brief: three actions for this quarter, each with the criterion that puts it among the 28% that deliver and not the rest. If one becomes a project, the two-week Diagnóstico turns the action into scope: criterion, timeline, and cost, signed before the code.