Four years after ChatGPT went mainstream, the corporate AI story has reached its most uncomfortable chapter: the bill is arriving, and almost nobody can read it. A cluster of reports out this week, from consultants, investment banks and analysts, converge on the same awkward truth. Businesses are now convinced AI is making them more productive. They simply cannot show how, or how much.
The clearest snapshot comes from a new EY survey of UK chief executives. Half of them named AI as one of the three biggest contributors to productivity gains over the past year, ranking it ahead of old standbys like redesigning business processes. Yet only 14 percent said they had full, real-time visibility over what their AI actually costs, and roughly one in three admitted they cannot properly assess its return on investment. The conviction is running well ahead of the accounting.
Investors are starting to notice the gap. According to Goldman Sachs, just 2 percent of S&P 500 companies have quantified the earnings impact of AI efficiency gains, and only 11 percent have cited a specific use case, even though half of them spent their last earnings calls promising fatter margins from the technology. As Morgan Stanley's Ellen Zentner put it, AI's next wave will reward companies that use it to lower costs and strengthen margins, not just those building the infrastructure. Earnings season, now under way, is shaping up as the moment investors demand real numbers rather than aspiration.
Underneath the measurement problem sits a harder one: the technology is colliding with budgets. At a recent Dell symposium, a company specialist described Bay Area customers "running through their token budget in a matter of weeks that they had allotted for months." The response is a quiet migration back toward on-premises hardware and open-weight models, used alongside frontier systems in the cloud to take the pressure off runaway inference bills. The economics of running AI, it turns out, look nothing like the economics of buying it.
The infrastructure was not built for this. One industry analysis notes that enterprise data centres were optimised for training models, a one-time capital event, not for inference, the recurring cost that scales with every query and every agent. Gartner found that only 28 percent of AI projects fully succeed and meet their ROI targets, while 20 percent fail outright. The most common reason for failure was expecting too much, too fast from infrastructure that could not support the workload.
Put those findings together and a pattern emerges. The constraint is no longer access to powerful models; almost every large company has that now. As one enterprise software chief argued this week, the real shortfall is "agency," the capacity to turn a model's recommendation into a governed, accountable decision that actually changes a business outcome. Intelligence without context, he wrote, is expensive noise. The winners will not be the organisations with the most AI, but the ones that can show, in numbers, what they got for it.