The three most cited AI ROI benchmarks of 2026 are 21 percent, 11 percent, and 12 percent, and they are not measuring the same thing. SAP and Oxford Economics report that companies expect a 21 percent return on AI this year. Kyndryl reports that 11 percent of organizations have achieved both of their top two AI goals. PwC reports that 12 percent of CEOs say AI has delivered both cost and revenue benefits. Each figure is credible, well sampled, and honestly published. All three are self-reported. None of them observed a single hour of actual work.
Two things are true at once. The evidence that enterprise AI is producing real value is stronger this year than last. And almost none of the numbers being circulated as proof are measurements in the sense a CFO means the word.
What are the 2026 AI ROI benchmarks?
Here are the three headline figures side by side, with what each one is actually built from.
| Source | Sample | Headline figure | What the number is | Basis |
|---|---|---|---|---|
| SAP and Oxford Economics, Value of AI Report 2026 | 2,600 business leaders, 13 countries | 21 percent ROI this year, up from 16 percent, rising to 38 percent in two years | Expected return on an average AI spend of 28 million dollars | Forward expectation, self-reported |
| Kyndryl, 2026 People Readiness Report | 1,100 senior business and technology leaders, 8 countries | 11 percent have achieved both of their top two AI goals | Attainment against objectives the organization set for itself | Achievement, self-reported |
| PwC, 29th Global CEO Survey | 4,454 CEOs, 95 countries | 12 percent say AI has delivered both cost and revenue benefits | Perceived benefit realization at the CEO level | Perception, self-reported |
Read the middle column again. One number is a forecast. One is a scorecard against goals nobody outside the company has seen. One is an executive impression. They are three different instruments, and they are routinely quoted in the same paragraph as though they were three readings of the same gauge.
Three questions, three numbers
The spread is not a contradiction. It is what you should expect when the questions differ.
Ask a leader what return they expect and you get an optimistic figure, because expectation is cheap and the investment is already committed. SAP and Oxford Economics found that global AI spend rose to 28 million dollars per company on average, up from 26.7 million dollars in 2025, and that expected ROI climbed to 21 percent from 16 percent, per the Value of AI Report 2026. Expected ROI in two years is 38 percent.
Ask whether they hit the goals they set and the number collapses. Kyndryl's 2026 People Readiness Report found that 32 percent of organizations had achieved at least one of their top two AI goals and 11 percent had achieved both, from a study of 1,100 senior leaders across eight countries.
#### The verb is the methodology
In survey research the verb carries the epistemics. Expect, achieve, and perceive are three different claims about the world, and only one of them requires anything to have happened yet. When a benchmark travels from a report into a board deck, the verb is usually the first thing to fall off. What arrives is "AI ROI is 21 percent," which is not what anyone published.
Why do executives report satisfaction and doubt at the same time?
The most revealing pair of findings in the SAP research sits in one sentence. Sixty-nine percent of businesses say they are satisfied with their current AI ROI, and more than two thirds are not convinced AI is achieving its full potential. Both are true because satisfaction is being assessed against expectations, not against evidence.
SAP Chief AI Strategy Officer Sean Kask named the underlying problem directly: businesses "must understand AI often provides value that is harder to measure than expected, and risk that moves faster than most governance can keep up with." That is a vendor executive saying, in a report about rising returns, that the returns are hard to measure. It is worth taking seriously.
Kyndryl's data shows the same tension from the workforce side. Deployment climbed, with 57 percent saying AI is embedded in core business processes or deployed broadly, against 35 percent a year earlier who said AI was fully integrated across their organizations. Those two phrasings are not identical constructs, which is itself the point. Readiness moved the other way on a consistent question: 23 percent say their workforce is fully ready for AI, a six-point drop from 2025, and 79 percent agree the speed of AI will outpace their workforce, governance, and operating models.
What the operational data says underneath the ROI headline
The same SAP study reports that 79 percent of businesses experience rework, delays, or backlogs from low-quality AI outputs, that 73 percent report challenges with incomplete data, and that 12 percent say their skills or their processes and frameworks are fully ready to govern AI. A 21 percent expected return and 79 percent of businesses cleaning up after their own AI outputs are not incompatible, but they cannot both be the whole story. One of them is being counted more carefully than the other.
What would a measured AI ROI number look like?
A measured number starts from work, not from a questionnaire. It requires four things that a survey response does not.
- A baseline. What the work looked like before the tool arrived, described in observed behavior across the systems where the work happens, not in recalled impressions.
- A comparison group. Teams that adopted the tool set against comparable teams that did not, matched on tenure, role mix, work type, and tool stack.
- A stated confidence level. A number that carries its own reliability, so a finance leader can see how much weight it will bear.
- A disclosed limit. An explicit statement of what the analysis cannot establish. Levos does not claim full attribution, and says so on its measurement methodology page.
That is controlled cohort analysis with confidence scoring. It produces a smaller, less quotable number than a survey does, and it is the only kind of number that survives an interrogation. It is also the basis of the AI Impact signal family, which treats AI adoption and its downstream effect on work as a measurable signal in its own right rather than a line item in a software budget.
Why the existing stack does not produce this number
The tools that could answer the question each see one slice. The finance system knows what was spent. The HRIS knows who is employed. The engagement platform knows how people say they feel. The AI vendor's own dashboard knows how many prompts were issued in its own product, which is a usage metric reported by the party being evaluated. None of them sees the work itself, across tools, before and after.
Vendor-supplied adoption dashboards have a structural conflict of interest, whatever the vendor's intentions. A measurement layer that sits above the stack and reads signals from where work actually happens does not. That is the position a Human Capital Operating System occupies: the intelligence layer above the existing tools, aggregating the signals that none of them sees alone.
How should a CFO read an AI ROI benchmark?
Use them as context, not as evidence. Four questions to put to any figure before it enters a decision:
- Who was asked, and how many? Sample size and geography change what a number can support. A survey of 4,454 CEOs and a survey of a hundred HR leaders cannot carry the same weight.
- Expected, achieved, or perceived? Find the verb in the original report, not in the article quoting it.
- Who sponsored it? Research published by a firm that sells the category being measured is not disqualified, but it should be read with that in view. This applies to us as much as to anyone.
- What is our equivalent number? If your organization cannot produce its own version of the figure from its own data, that absence is the more important finding.
Match Group CFO Steve Bailey put the consequence plainly to CFO Dive, saying he now requires a "business case with clear impacts either in the form of cost savings or efficiency gains" for material AI spending, because "a blank check for AI makes that very difficult to do." A benchmark cannot supply that business case. Only internal measurement can.
Frequently asked questions
What are the 2026 AI ROI benchmarks? The three most cited figures are 21 percent, 11 percent, and 12 percent, from SAP with Oxford Economics, Kyndryl, and PwC respectively. They are a forward expectation, goal attainment, and perceived benefit realization. All three are self-reported.
Why do AI ROI benchmarks disagree with each other? Because they ask different questions of different populations, and none of them observe the work. The gap between 21 percent and 11 percent is the distance between what leaders anticipate and what they can currently confirm.
Is a self-reported AI ROI number useless? No, but it should be labeled. Survey data is good evidence of direction and market shape, and weak evidence of a specific return at a specific company.
What would a measured AI ROI number look like? It would start from observed behavior in the systems where work happens, compare adopting teams to comparable non-adopting teams, carry a confidence score, and disclose what it cannot establish.
How should a CFO use an AI ROI benchmark? As a range to position against. Ask who was surveyed, whether the figure is expected or achieved or perceived, who sponsored the research, and what the equivalent internal number is.
Ready to produce your own number?
The benchmarks tell you where the market thinks it is. They cannot tell you what your AI investment returned. That number has to come from your own work.
Design partner cohort today: 150 to 500 employees. Expanding to 500 to 2,000 in the second half of 2026.
