The AI ROI Gap Starts With the Wrong Question

PwC's 29th Global CEO Survey found that 56 percent of CEOs reported neither higher revenue nor lower costs from AI, while only 12 percent reported both. This article argues that the result is predictable when organizations lead with AI instead of business value. It distinguishes technical pilot success from economic success, explains why measurement must change capital-allocation decisions, and argues that AI should compete for investment on measurable outcomes like any other technology.

Sep 07, 2026 · Blaize Stewart · #ai-roi #ai-adoption #value-realization #artificial-intelligence #enterprise-ai #kpi #portfolio-management #business-strategy

“AI does not yield an ROI.” At least that’s what the headlines proclaim. PwC's 29th Global CEO Survey puts a useful number on a problem that has been visible for some time. The survey, based on responses from 4,454 CEOs across 95 countries and territories, found that 56 percent of CEOs said their companies had seen neither higher revenue nor lower costs from AI over the previous 12 months. Thirty percent reported additional revenue, 26 percent reported lower costs, and only 12 percent reported both. PwC did not measure ROI in the strict accounting sense because the survey did not net those gains against the full cost of the investment. Even so, the result is difficult to ignore. For most companies in the survey, AI activity had not yet translated into a significant financial outcome. The result says little about whether AI can create value in principle. It says much more about whether current enterprise deployments are producing measurable business outcomes. Too many AI programs still begin with the technology by asking what can we do with AI, where can we add a copilot, which model should we adopt, or which workflows can we automate? Those are implementation questions, and they come too early. The better question is what business value needs to be created and how it will be measured. The trend is nothing new.

PwC: The latest result reinforces the gap between AI experimentation and financial impact. Organizations are investing heavily in AI, but measurable returns remain inconsistent.
MIT NANDA: Its preliminary 2025 report became known for the claim that 95 percent of custom enterprise GenAI pilots were failing, although that figure has often been generalized beyond the report’s scope. The more useful finding was that many pilots were not reaching sustained deployment or measurable P&L impact, with success appearing to depend more on integration and implementation approach than on raw model capability.
McKinsey, 2025: Its global AI research found that tracking well-defined KPIs had the strongest relationship with bottom-line impact among the adoption and scaling practices it examined. Yet fewer than one in five respondents said their organizations were actually tracking those KPIs.
McKinsey, 2026: McKinsey extended that work with a measurement framework connecting technical performance, adoption, operating outcomes, strategic outcomes, and financial impact.

Across the research, the pattern is consistent. AI activity is abundant. Evidence of business value is much scarcer. That gap exists because a pilot can succeed technically while failing economically. But it is not all bad news.

AI can be a powerful catalyst. It can expose inefficient workflows, make costly automation practical, enable new product experiences, and change how people interact with information. But a catalyst still needs something worthwhile to act on. Starting with AI reverses the normal logic of investment. A value-led approach starts with a measurable outcome: reduce claims processing time, shorten sales response cycles, improve recommendations, cut reconciliation effort, or reduce service rework.

The purpose of a pilot should be to test a hypothesis, and the hypothesis has to include business value. Technical feasibility still matters. The organization needs to know whether the model can perform the task with acceptable accuracy, latency, security, and reliability. Those measures establish whether the system can work. Economic viability requires additional evidence. The economic hypothesis should be defined before the pilot begins. It should establish the baseline, target KPI, expected cost, attribution method, and success threshold. Deloitte’s 2026 CFO guidance makes the same point, calling for locked pre-deployment baselines, predefined success criteria, methods to isolate AI’s contribution, and hurdle rates for underperforming initiatives. There are three possible outcomes:

One possible outcome is that the use case produces no material value. The model may work and users may even like it, but the business KPI does not move. That is a valid result. The organization learned something and should usually stop funding the use case.
Another outcome is that measurable value exists, but the value is too small to justify the cost. Perhaps an AI workflow saves time, but the released capacity is not usable. Perhaps a feature improves conversion slightly, but the infrastructure, licensing, integration, and operating costs consume the gain. That result may justify redesigning the use case, choosing a cheaper model, replacing AI with conventional automation, or simply stopping.
The third outcome is the one every pilot hopes to find. The business metric moves enough to exceed the full cost and the result holds under credible measurement. At that point the use case has earned the right to scale. The discipline is in accepting all three outcomes. An AI program that can only produce success stories is not measuring investments. It is collecting demonstrations.

PwC's later 2026 AI Performance Study adds an important dimension to the CEO survey. The study of 1,217 senior executives found that 20 percent of companies captured 74 percent of the AI-driven economic value measured in the research. The strongest performers redesigned workflows around AI, pursued growth opportunities, built stronger foundations, and scaled proven uses. PwC found that these leaders were twice as likely to redesign workflows to incorporate AI instead of simply adding AI tools. Significant value is being produced, but it is highly concentrated. The more plausible reading is that value depends heavily on where AI is applied, how the work changes around it, how the result is measured, and whether the organization has the discipline to scale what works. I have written separately about the mechanics of, but the management principle is simpler than the measurement framework. But the bottom is that AI should have to compete for capital the same way any other investment does. PwC's finding that most CEOs still report no revenue or cost benefit from AI should create urgency. Lead with value. Define the outcome first. Establish the measure before implementation. Give the pilot permission to prove that AI is the wrong answer. Scale only when the economics justify it.

Further reading