A year ago, the enterprise AI question was whether the technology worked. It does. The uncomfortable new question, aired repeatedly this week across surveys and conference stages, is why so much of it still does not pay. Report after report lands on the same gap: adoption is racing ahead, returns are limping behind, and the reason has almost nothing to do with the models themselves.
The figures are sobering for anyone who wrote a triple-digit return into a business case. IBM's poll of roughly 1,000 executives put average AI return on investment at about 55 percent, well below what many expected a year ago. PwC's 2026 Global CEO Survey found only 12 percent of chief executives could point to both revenue growth and cost reduction from AI, while 56 percent reported no significant financial benefit at all. On the Indian GCC circuit, 71 percent of centers now use generative AI, but only 27 percent say it has produced tangible results. Goldman Sachs, for its part, argued the broad enterprise productivity payoff has simply not arrived yet.
What is striking is how consistently the blame lands not on the algorithms but on the organizations around them. IBM's field CTO Sunil Murthy estimates that 21 percent of potential productivity is lost to friction between technology and business teams. His diagnosis is a memorable one: a model that gives a wrong answer is often not missing intelligence, it is missing context. Ask whether to walk or drive to somewhere 100 meters away and the obvious answer is walk, unless you happen to be heading to the car wash.
That context problem turns up everywhere in the data. Research by Alteryx found 77 percent of IT leaders call business context critical to accurate AI output, yet 53 percent struggle to feed that context into their systems, partly because only 18 percent of firms give business users self-service access to their own data. The rules and definitions that make a company run mostly sit in silos the AI cannot reach, so the model falls back on generic assumptions and produces confident, useless answers.
The mood at the AI4 conference in Las Vegas captured the shift. Buyers arrived with less patience for vague claims and more interest in governance, cost control and proof of production deployments. “FinOps for AI,” managing token spend the way firms once learned to manage cloud bills, was often the opening question rather than an afterthought. Gartner has warned that by 2027, 40 percent of enterprises will demote or decommission autonomous agents after governance failures surface in production.
The throughline is almost old-fashioned. The technology keeps improving at a startling pace; the organizations absorbing it do not. As Murthy put it, the model was never really the constraint. The constraint is whether the architecture, the data, the decision rights and the operating model get redesigned to match. The companies still treating AI as something you buy, rather than something you rebuild around, are the ones quietly discovering that a better model does not fix a broken workflow.