In a survey of nearly 6,000 senior executives across the United States, the United Kingdom, Germany and Australia, 89 percent said artificial intelligence had no effect on labor productivity at their firms over the previous three years, even as global AI spending is forecast to reach $2.7 trillion in 2026, according to figures compiled by TokenPost. That forecast is a 49.5 percent jump on 2025, with AI infrastructure alone projected at $1.484 trillion. The same executives reported using AI for an average of just 1.5 hours a week. The gap between what companies are spending and what they can measure has become the defining problem of enterprise AI.
The numbers do not all point one way, and the distinctions matter. A separate tracker found that in August 2026, 40.5 percent of employed U.S. adults used generative AI for work, that AI accounted for 5.3 percent of total work hours, and that workers estimated the time saved at 2.1 percent of all hours. Among 1,719 respondents in a global poll, 80 percent said AI improved their individual productivity, yet only 37 percent reported a positive effect on their organization's earnings, and about 6 percent cleared a high-performer bar. Faster individual work, in other words, is real; it simply is not reaching the income statement.
Why it leaks away is the question several industry analyses converge on. Writing for Forbes, Kiplot chief executive David Edwards invokes Robert Solow's 1987 quip that the computer age was visible everywhere except the productivity statistics, and Erik Brynjolfsson's work on the productivity J-curve: returns come only after a firm rebuilds itself around a technology, which is slow, costly and invisible on the balance sheet. A Cognizant and Google Cloud panel described a "productivity illusion" in which an AI tool speeds one step of a process while human review, system handoffs and fixed queues absorb the gain. As one panelist put it, "the leak isn't the model, it's the system that the model might be dropped into."
A second diagnosis points at the models themselves. Writing in CIO, author Rodrigo Madanes argues that enterprises do not need artificial general intelligence but "artificial specific intelligence," and cites an MIT finding that 95 percent of enterprise AI pilots show no measurable profit. Frontier models are frozen at training time and know the public record, not a company's undocumented, tacit expertise, so they can top every benchmark and still deliver little. His prescription, and that of a separate Forbes piece on data reliability, is unglamorous: narrow the pilot to a workflow you would actually scale, put domain experts beside the engineers, and feed the model trustworthy data, because a more powerful model fed bad data simply generates wrong answers faster.
The surveys measure different things and cannot be read as a single number, a point TokenPost stresses: saved time becomes value only when it produces additional output, and individual speed is not an earnings gain. That caveat is also the way out. The through-line across every account is that AI's returns depend less on model capability than on the surrounding work: redesigned processes, reusable infrastructure, an accountable owner for the business metric, and capacity that is deliberately redirected rather than quietly reabsorbed. Until firms do that rebuilding, the $2.7 trillion and the 89 percent will keep describing the same year.