IBM, a technology and consulting company, said that large enterprises report an average return on investment (ROI) from AI of just 17%, while internal friction consumes roughly one-fifth of the potential value organizations could be getting from their AI investments.
The 17% figure in “Is your AI paying off?” comes from an unpublished IBM Institute for Business Value (IBV) survey of 1,250 IT executives conducted from June through August 2026.
The estimate that about one in five dollars of AI value is lost to internal friction comes from IBM’s earlier “Redesign for enterprise AI” research. The new report identifies fragmented processes, inconsistent measurement, poor visibility and technical debt as factors eroding AI returns.
These problems show up across the AI portfolio, with nearly two-thirds of AI initiatives fail to meet their expected objectives, IBM found, leaving a relatively small number of successful projects to generate a disproportionate share of realized value.
The report argues that improving returns therefore depends on more than choosing a better model: companies also have to see what AI is costing them, account for the technology needed to support it and measure which investments are actually working.
Cloud costs outrun projections
IBM’s “2026 Tech Leader Study” found that 85% of technology leaders lack real-time visibility into AI spending, while organizations report cloud costs running nearly 50% above initial projections as AI workloads consume more infrastructure than expected.
Without visibility, IBM said leaders struggle to tell whether delivering a unit of AI value is becoming cheaper or more expensive. The report points to models, inference, data, infrastructure and vendor relationships as parts of the AI cost base that need to be considered together rather than treating the model or application as the whole investment.
IBM’s “The tech debt reckoning” found that almost 70% of executives expect technical debt to make at least some AI initiatives financially untenable. At the same time, nearly 60% of executives report pressure to reduce non-AI technology spending to free up money for AI, creating a tension between funding new AI projects and maintaining the technology foundation those projects depend on.
Organizations that explicitly include modernization and technical-debt remediation in their AI business cases project almost 30% higher AI ROI than organizations that treat modernization as a separate investment.
IBM’s “The enterprise in 2030” found that organizations planning to use smaller AI models or a mix of custom and foundation models expect 24% greater productivity gains and 55% greater improvement in operating profit margins by 2030 than organizations relying mainly on large pretrained models.
The study describes those figures as executives’ expectations, not observed results.
Measurement determines where money goes
Just 25% of executives said their organizations consistently measure business value and AI ROI, according to IBM, making it harder to decide which projects should receive more money and which should be scaled back.
A separate IBV study, “Solving the AI ROI puzzle”, found that nearly 70% of chief AI officers launch AI initiatives even when they lack a reliable way to assess or measure their impact. In another unpublished 2026 survey cited in the new report, organizations that use AI performance data to guide funding and talent decisions reported 112% higher AI ROI than their peers.
Only 6% of organizations direct their five largest AI investments toward the five business functions generating their highest AI returns.
The organizations report 22% higher function-specific AI ROI than the rest, suggesting that the problem is not simply how much businesses spend on AI, but whether they can identify where the spending is producing value and move resources accordingly. “Is your AI paying off?” combines findings from separate IBM Institute for Business Value studies with different respondent groups, fielding periods, geographies and research designs. The samples were not pooled or reweighted into one dataset, so comparisons across them should be read as directional patterns rather than causal relationships.