The 2025 reality check
Three findings from 2025 are worth a board’s attention. A preliminary MIT NANDA study reported that roughly 95% of organisations were seeing no measurable profit-and-loss return from generative AI[1]. S&P Global found the share of companies abandoning most of their AI initiatives had jumped to 42%, up from 17% a year earlier[2]. Gartner predicted more than 40% of agentic AI projects would be cancelled by the end of 2027[3].
Those failure rates sit against near-universal adoption. McKinsey found 88% of organisations were using AI regularly by late 2025[5]. That level of use has not produced returns that reach the earnings line.
Why the gap is real even if 95% is arguable
The 95% figure needs a caveat. That study was preliminary, not peer-reviewed, used a short measurement window and drew criticism on its sample[1]. It is best read as a rough indicator. The broader conclusion still holds, because the other evidence is independent and points the same way. S&P’s abandonment data, Gartner’s cancellation forecast and BCG’s finding that 74% of companies struggle to scale AI value are all consistent with it[2][3][4].
The causes line up too. RAND’s analysis of failed AI projects traced them to misunderstood requirements and poor data rather than the model itself[6]. BCG reaches a similar split, putting success at roughly 70% people and process, 20% technology and 10% algorithms[4]. In both, technology is the smaller part of the problem.
The learning gap
The NANDA report’s diagnosis is more useful than its headline number. It found that pilots stall when the tool does not learn from feedback, or does not fit the way people already work. A demo that performs well in a controlled setting turns into shelfware once it meets a live workflow and its exceptions. The report also found that bought-in solutions and genuine partnerships tended to outperform ambitious internal builds, because they crossed that workflow gap more often.
What the minority who succeed do differently
The organisations that get a return are more disciplined about a few specific things. Better models are rarely the reason.
- They pick one high-value use case and define the business metric before evaluating any tool
- They fix the data and the workflow the AI has to live inside, rather than bolting it onto a broken process
- They assign one named owner to the business result the tool is meant to deliver
- They measure against the metric they set, and they will stop a pilot that misses it
This is the same discipline that separates the successful fifth from the failing majority in the broader AI record, which we cover in our companion guide on why 80% of AI projects fail. Different studies report different failure rates, but they keep coming back to the same cause.
The board’s job
For a board, the trap is being reassured by the wrong metric. Adoption is the easy thing to report, and it can look like progress. Counting seats deployed, prompts run and staff trained says nothing about whether the spending has come back. Ask instead for the profit-and-loss impact of each material use case, and fund the less visible work, the data and process redesign, that decides whether the tool ever pays back.