A new global study published today by SAP and Oxford Economics shows that businesses are reporting and projecting higher returns from artificial intelligence, even as they acknowledge major gaps in readiness and oversight. The Value of AI Report 2026 surveyed 2,600 business leaders across 13 countries and found that expected ROI from AI has climbed to 21% this year — equivalent to roughly $6.3 million on average — up from 16% in 2025. Looking two years ahead, respondents expect that figure to reach 38%, or about $15.9 million.
Agentic AI is the main driver of the more optimistic forecasts. Average expected returns from agentic AI alone are projected to more than quadruple, rising from $4.3 million last year to $17.6 million within two years. Eighty-three percent of respondents said agentic AI has moderate to very high potential to transform their organizations. Only 3%, however, said they are fully prepared for it.
The same research highlights a persistent lag on the control side. Just 12% of businesses said their skills or processes and frameworks are fully ready to govern AI effectively. Nearly seven in ten (69%) either agreed or were unconvinced that they are deploying agents faster than they can govern them. Thirty-eight percent reported having no human-in-the-loop process for agentic workflows, 37% lacked permission and access controls for agents, and only 44% maintained a registry of agents in use.
“AI has moved from experiment to execution, and that’s beginning to show real returns,”
“But there’s still a long way to go. Because AI that lacks context—whether that’s processes, data, or governance—at best creates activity without outcomes and at worst creates risk.”
Sean Kask, SAP’s Chief AI Strategy Officer
The study paints a picture of rising business confidence in AI’s financial returns that is outpacing the maturity of the systems meant to keep those systems accountable.
Conditions Driving the Event
Global companies are reporting higher expected returns from AI investments, with average projected ROI rising from 16% last year to 21% this year and 38% within two years, according to the SAP and Oxford Economics survey of 2,600 business leaders.
Average annual AI spending per business has increased modestly to $28 million, while the expected financial return on that spending has grown more sharply, creating stronger board-level pressure to demonstrate value.
Agentic AI has become the primary source of future ROI optimism, with projected returns from agentic systems more than quadrupling to an average of $17.6 million within two years.
Nearly one-third of all business tasks (30%) are already supported by AI, with respondents expecting that figure to rise to 48% within two years, accelerating the shift from pilots to production use.
Strategic investment in AI has almost doubled year-over-year to 17%, yet 41% of organizations still describe their approach as piecemeal rather than coordinated.
Only 3% of businesses say they are fully prepared for agentic AI, while 83% rate its transformative potential as moderate to very high, creating a wide gap between ambition and operational readiness.
Data quality remains the most frequently cited barrier, with 73% of companies reporting incomplete data and 79% experiencing rework, delays, or backlogs caused by low-quality AI outputs.
Governance maturity is lagging significantly: only 12% of respondents said their skills, processes, and frameworks are fully ready to govern AI effectively.
Nearly seven in ten organizations (69%) either agree or remain unconvinced that they are deploying agents faster than they can govern them.
Key control mechanisms are missing at scale: 38% lack a human-in-the-loop process for agentic workflows, 37% lack permission and access controls for agents, and only 44% maintain a registry of agents in use.
Shadow AI use continues to rise, with 69% of businesses reporting that it occurs at least occasionally, further expanding the ungoverned surface area.
Workforce readiness has not kept pace with tool evolution, as 78% of respondents are either unsure or agree that company upskilling is falling behind the speed of AI change.
What AI ROI Looked Like Before
Until recently, most enterprises treated AI return on investment as a soft estimate rather than a continuous measurement system. Finance and leadership teams typically relied on vendor case studies, employee surveys about perceived time savings, and simple calculations that divided total AI spend by rough hours saved. Clear baselines were almost never established before tools were deployed, so there was no reliable “before” state to measure against.
Cost tracking was incomplete. Organizations usually counted only cloud bills, API tokens, and licenses. Shadow AI purchases, untagged infrastructure, governance overhead, rework from poor outputs, and the cost of reviewing agent actions were rarely included. As a result, the ROI numbers shown to boards were almost always overstated gross returns rather than risk-adjusted figures.
Usage data was equally thin. Companies often knew how many licenses they owned but had little visibility into who was actually using the tools, how deeply they were using them, or whether the usage produced finished work. Value was described in vague terms such as “better decisions” or “higher productivity” without linking specific AI activity to concrete outcomes like cycle-time reduction or cost avoided. AI ROI was largely a storytelling exercise supported by incomplete data.
What AI ROI Looks Like Now
The picture is starting to shift, though it remains incomplete. The SAP and Oxford Economics Value of AI Report 2026 shows expected AI ROI rising from 16% last year to 21% this year, with a projected jump to 38% within two years. Agentic AI is the main driver of that optimism, with projected returns more than quadrupling. At the same time, the research makes clear that most organizations still lack the systems needed to turn those expectations into reliable numbers.
Leading teams are beginning to demand full spend visibility (including shadow tools), proper cost attribution, observed adoption and depth of use, and a direct link from usage to business outcomes. Some are also starting to subtract the cost of failures and governance overhead so the return figure is risk-adjusted.
Yet the same data shows how far most companies still have to go. Only 12% say their skills and processes are fully ready to govern AI. Nearly seven in ten report deploying agents faster than they can govern them. Data quality problems remain widespread, with 73% facing incomplete data and 79% experiencing rework or delays from low-quality outputs. These gaps undermine any ROI claim, because returns that cannot be measured cleanly or defended under scrutiny do not survive board or regulatory review.
The current state of AI ROI is defined by a widening gap: rising expected returns on one side, and immature measurement infrastructure on the other.
Our Take
AI ROI Take
The SAP and Oxford Economics findings make one thing clear: expected returns from AI are rising faster than the systems that can prove them. Organizations that treat ROI as a storytelling exercise — relying on vendor case studies, self-reported time savings, and incomplete cost figures — will continue to report higher expected numbers while remaining unable to defend those numbers under real scrutiny.
The practical work now is to close the measurement gap before the expectations gap becomes a credibility problem. That means building four capabilities in order. First, full spend visibility that captures every AI cost, including shadow tools, untagged infrastructure, and governance overhead. Second, clear cost attribution so spend can be assigned to teams, use cases, and outcomes rather than sitting as a shared black box. Third, observed adoption and usage data that shows who is actually using the tools and how deeply, not just how many licenses were purchased. Fourth, value realization that connects usage to concrete business results and subtracts the cost of failures and rework.
Until those four layers exist, any ROI claim remains a gross estimate rather than a risk-adjusted figure that finance and the board can trust. The companies that install this measurement infrastructure now will be able to answer the simple question the SAP data leaves hanging: what did the AI spending actually produce? The ones that do not will keep reporting rising expected returns while the gap between ambition and evidence continues to widen.
Agentic AI is accelerating both the opportunity and the risk. Organizations that deploy agents faster than they can measure and govern them are not generating returns. They are accumulating untracked liability. The next phase of AI value will belong to the teams that treat measurement as a prerequisite for scale rather than an afterthought.
