AI Operator

Get Started

AI Operator

AI ROI Statistics 2026: Productivity, Revenue, and EBIT Evidence

AI produces measurable gains in selected tasks, but enterprise ROI remains uncommon and uneven. Research reports productivity improvements around 15% in customer support and meaningful time savings in some knowledge work. Yet McKinsey found only 39% of respondents attributed any enterprise EBIT impact to AI, and roughly 5% met its high-performer threshold for significant value and more than 5% EBIT contribution.

Jul 30, 2026

Key findings

• Task-level productivity evidence is real but highly dependent on the task, worker, tool, and evaluation window.

• In McKinsey’s 2025 survey, 39% of respondents reported any enterprise-level EBIT impact from AI. Only 109 of 1,993 respondents met the report’s AI high-performer definition.

• OECD found 65.1% of GenAI-using SMEs reported improved employee performance, 45.2% reported saving money, and 25.9% reported increased revenue. The survey did not measure the size of those changes.

• A large workplace experiment found active AI-tool users spent two fewer hours per week on email, but it detected no broader change in task quantity or composition.

• Industry correlations can support an investment thesis, but they are not causal ROI estimates for an individual company.

• SMBs should measure ROI at the workflow level before expecting enterprise-wide financial impact.

Methodology and definitions

This evidence review separates five outcome levels:

1. Task performance: speed, quality, or output on a defined activity.

2. Worker capacity: time saved across a person’s work period.

3. Workflow outcome: cycle time, accuracy, volume, cost, or conversion for a repeatable process.

4. Function outcome: cost, revenue, or service improvement across a department.

5. Enterprise outcome: measurable effect on company-wide earnings, usually EBIT or profit.

A result at one level does not prove a result at the next. Saving 20 minutes on a task creates potential capacity. It becomes ROI only if the organization captures that capacity through more output, lower cost, better quality, faster revenue, or avoided risk.

The sources include McKinsey’s 2025 Global Survey on AI, OECD’s representative SME survey, PwC’s labor-market and company-financial analysis, NBER field experiments, and Stanford’s research synthesis. Survey evidence is reported as perception or association. Experimental evidence is reported only for the setting tested.

Enterprise EBIT evidence

McKinsey’s 2025 State of AI report surveyed 1,993 participants between June 25 and July 29, 2025. Thirty-nine percent reported attributing some level of enterprise-wide EBIT impact to AI.

The report defined AI high performers as respondents who said their organizations obtained significant value and more than 5% of EBIT from AI. There were 109 high performers, about 5.5% of the total sample. That supports a restrained conclusion: meaningful enterprise returns existed, but they were concentrated.

McKinsey later summarized the result as 94% not seeing significant value. “No significant value” is not the same as zero benefit. An organization can save time in a function and still fall below a threshold for material enterprise impact.

High performers differed in more than tool choice. They were more likely to pursue growth and innovation alongside efficiency, redesign workflows, deploy across more functions, define human-validation processes, and have senior leaders demonstrate ownership. These are correlations, not proof that any one practice causes EBIT.

SME-reported benefits

The OECD’s Generative AI and the SME Workforce survey covered 5,232 SMEs in seven countries in late 2024. Among firms using generative AI:

• 65.1% said it improved employee performance.

• 45.2% said it helped save money.

• 35.1% said it enabled tasks the firm could not perform before.

• 34.6% said it helped offer new products or services.

• 28.5% said it helped compete with larger companies.

• 25.9% said it increased revenue.

These results show the ordering of perceived benefits. They do not establish magnitude, attribution, or net return after software, implementation, review, training, and maintenance costs. The OECD explicitly notes that a company reporting a benefit was not asked whether it was large or small.

That distinction prevents a misleading statement such as “65% of SMEs achieved AI ROI.” The accurate statement is that 65.1% of GenAI-using SMEs reported improved employee performance.

Task productivity and time savings

Stanford’s 2026 AI Index economy chapter synthesizes studies reporting gains around 14% to 15% in customer support, 26% in software development, and 50% in marketing output. Those values come from different studies, tasks, tools, and outcome measures. They are examples, not a cross-industry ROI table.

A Stanford labor evidence review describes a customer-support deployment with a 15% average productivity increase. Novice and lower-skilled workers improved about 30%, while highly skilled agents did not improve and experienced a slight reduction in response quality. A workforce average can hide important differences.

The NBER Shifting Work Patterns experiment randomly gave 7,137 knowledge workers across 66 firms access to a generative AI tool integrated into email, meetings, and writing. In the second half of the six-month experiment, the 80% of treated workers who used the tool spent two fewer hours per week on email and reduced after-hours work. The researchers did not detect changes in the quantity or composition of tasks beyond those individual time savings.

A separate Danish labor-market study, revised in March 2026, found task reorganization and reported productivity benefits but no measurable effect on earnings or recorded hours, ruling out effects larger than 2% over two years. These studies illustrate the conversion gap between local efficiency and financial return.

Industry-level economic associations

PwC’s 2025 Global AI Jobs Barometer analyzed close to one billion job advertisements and thousands of company financial reports. It found 27% growth in revenue per employee in industries most exposed to AI, compared with 9% in the least-exposed industries. It also reported that skills sought by employers changed 66% faster in highly AI-exposed jobs.

The analysis is broad and economically relevant, but exposure is not implementation. Industries differ in capital intensity, demand, market structure, digital maturity, and workforce composition. The result is an association consistent with AI-related productivity, not a causal estimate that a specific AI investment returns three times more revenue.

How SMBs should calculate AI ROI

Start with one workflow and one baseline. The existing AI automation ROI calculation guide can be used to structure the arithmetic, but the evidence inputs must come from the company’s own operation.

Measure:

• Eligible cases per month.

• Current handling time and loaded labor cost.

• Current error, rework, delay, or abandonment rate.

• AI-assisted handling time.

• Human review and exception time.

• Software, implementation, training, and maintenance cost.

• Quality change, including false positives and customer-impacting errors.

• Revenue or retention outcomes only when attribution is credible.

Calculate net monthly value after recurring costs, not gross hours saved. Apply an adoption factor to reflect how much eligible work actually uses the workflow. Run a conservative case with lower savings and higher review time.

The strongest first project has high frequency, measurable outputs, bounded risk, and a clear way to redeploy saved capacity. A workflow that saves time no one tracks is harder to connect to earnings.

Limitations

Survey respondents can overstate or understate outcomes. OECD benefits have no effect size. McKinsey’s sample is not a national business census. PwC’s industry associations are not causal. Experiments often test a narrow task, tool, company, or worker population.

EBIT is also a lagging measure. A useful workflow may improve service or capacity before enterprise earnings move. Conversely, a fast task result can disappear after review, integration, and adoption costs.

What would change our view

We would become more confident in broad AI ROI if representative studies linked specific deployments to audited cost, revenue, and earnings outcomes over longer periods, with matched controls and full implementation costs. We would become more cautious if task gains weakened in production or if review, correction, and change-management costs consistently absorbed them.

FAQs

What is the average ROI of AI?

There is no defensible universal average. Studies measure different tasks, costs, benefits, and time periods. ROI should be calculated for a specific workflow using the company’s own baseline and total operating cost.

Do most companies make money from AI?

Many report local benefits, but significant enterprise impact is less common. McKinsey found 39% reporting any EBIT impact and about 5.5% meeting its high-performer threshold.

What AI productivity statistic is most credible?

Credibility depends on the decision. Randomized and field experiments provide strong evidence for their tested setting, such as two fewer email hours per week among active users. They should not be generalized to every workflow.

Why does AI save time without improving profit?

Saved time may remain unused, become unmeasured capacity, be offset by review and correction, or occur in work too small to affect enterprise earnings. Workflow redesign is often required to capture the benefit.

What should an SMB measure first?

Measure case volume, handling time, quality, review time, exceptions, recurring cost, and the share of eligible work using the workflow. Add revenue only when the attribution path is explicit.