Survey says AI ROI expectations hit 17%, but the gap between reported and measured value remains wide
The SAP Value of AI Report 2026 finds that AI now supports roughly 30% of tasks in the average organization, up from 25% a year ago, and that expected ROI for agentic AI has climbed from 10% to 17% in the same period. Those are the headline numbers the report uses to frame enterprise AI as delivering on its promise. The framing deserves scrutiny.
The source is a vendor-commissioned report produced with Oxford Economics, based on a survey of 2,600 business leaders across 13 countries. SAP is also the source of the product claims cited throughout the article, including the performance metrics for its accruals accounting agent, the scale of its knowledge graph, and the capabilities of its AI Agent Hub. The report identifies three broad barriers to AI value: strategy, data, and governance. Each of those barriers maps directly to SAP product positioning. Data context maps to SAP Business Data Cloud. Agent discovery and governance map to the AI Agent Hub. This is not a coincidence, but the article does not flag the alignment between the challenges named and the products presented as answers.
The ROI figures require careful reading. The report finds 69% of businesses say they are satisfied with their AI ROI, and 67% remain unconvinced AI delivers its full potential. These are not contradictory findings if satisfaction refers to proof-of-concept returns and the unconvinced share refers to the gap to scalable value. But the 17% and 21% ROI figures are expectations, not measured outcomes. The source presents them as evidence that AI is generating real returns, but the underlying data reflects what respondents expect or intend, not what they have independently verified. Whether those expectations survive contact with production workloads, data quality issues, and governance failures is a separate question the report does not resolve.
The data quality finding carries genuine weight. Seventy-three percent of respondents name data quality and availability as the primary reason they are not extracting more value from AI, and 79% report rework, delays, or backlogs from low-quality outputs at least occasionally. That is a constraint the source does not soft-pedal. Sean Kask, chief AI strategy officer at SAP, frames it as a context problem: extracting data from an ERP system breaks semantic relationships that generative AI depends on. This is a coherent technical argument, though the source does not compare it against alternative architectural responses or independent assessments of whether knowledge graphs resolve it at scale.
The governance findings are more alarming in scope than the article emphasizes. Only 12% of businesses say they are fully prepared to govern AI. Sixty-nine percent acknowledge occasional to frequent use of unapproved shadow AI tools. The AI Agent Hub, which SAP presents as the governance solution, has surfaced thousands of agents inside customer landscapes that those organizations did not know they had. Kask draws a parallel to employee onboarding, which is a reasonable analogy for access control. Whether it adequately addresses the harder problem of behavioral auditing, failure attribution, and cross-agent dependency chains is not evaluated in the source.
The 400 AI use cases SAP has shipped and the beta accruals accounting agent (reducing a task from approximately 12 hours per month to 2 to 3 hours) are specific claims. The source does not independently validate the efficiency numbers or describe how they were measured. The math on scaling that reduction across processes assumes uniform applicability and ignores variance in process complexity, exception handling, and audit requirements. Those assumptions are common in vendor ROI projections and worth noting when they appear in a report the vendor produced.
The broader workforce finding deserves attention: almost 80% of respondents agree that maximizing AI value requires more than technical upskilling, and 75% are planning to reskill employees. This aligns with what independent research on enterprise AI adoption has found, but the source does not cite external validation. The framing that the conversation has shifted from job replacement to human-AI collaboration is plausible but not tracked longitudinally in the source.
The survey structure itself warrants caution. Self-reported ROI expectations, satisfaction ratings, and preparedness scores are subject to optimism bias and framing effects. The source does not describe how respondents were recruited, whether they represent a random sample of enterprise AI adopters, or what baseline they use for the 16% to 21% general AI ROI range. A survey of 2,600 business leaders is a substantial sample, but sample size does not substitute for methodological transparency when the commissioner is also the subject of the favorable findings.
The bottom line is that this report provides useful survey data on enterprise AI sentiment. The gap between reported satisfaction and reported under-delivery is a real tension worth tracking. The governance preparedness numbers are worth watching. But the source presents vendor framing as market evidence, uses survey expectations as proof of ROI, and positions its own product portfolio as the solution to the problems its own research identified. Readers treating the report as independent validation of AI value delivery will come away with a cleaner picture than the source actually provides.