TL;DR
- 88% of organizations use AI in at least one function and 79% use generative AI, while scaled agent use stays in the single digits across nearly all functions.
- Published agent adoption figures range from 23% to 74%, because each survey measures something different on a different sample.
- Reported ROI falls from 74% to roughly 5% as the definition of a result gets stricter
- Scaled agent use concentrates in the technology sector: 24% in software engineering, 22% in IT.
- 74% of organizations plan to use agents by 2027; 21% have mature governance in place.
- Gartner projects 40% of enterprise applications will carry agents by the end of 2026, and separately that more than 40% of agentic projects will be canceled by the end of 2027.
- Measured AI productivity gains range from +200% to −19% depending on the task.
As of 2025, 88% of organizations use AI in at least one business function, up from 78% a year earlier. And 79% regularly use generative AI in at least one function, compared with 71% in 2024 (Stanford AI Index 2026, Chapter 4: Economy, drawing on McKinsey survey data).
The figures for agents are an order of magnitude lower. Stanford reports that in most business functions a majority of respondents use no agents at all, and that scaled use stays in the single digits across nearly all functions. In the two functions with the most activity, IT and knowledge management, roughly two-thirds or more reported no use (Stanford AI Index 2026, Ch. 4).
McKinsey measures the same thing in different terms: 62% of organizations are experimenting with agents, 23% are scaling them in at least one function, and no single function exceeds roughly 10% scaled deployment (McKinsey, State of AI, November 2025).
The sections below cover adoption by company size and function, ROI figures from four independent studies, data on pilot failure and governance, measured productivity results, and projections through 2035. Every number is given with its primary source and publication date.
How many companies actually run AI agents?
The answer depends on what each survey asked. Published figures for 2025 and 2026 range from 23% to 74%, and the spread comes from differences in definition and sample rather than from conflicting data.
Google Cloud reports that 52% of executives say their organization is actively using AI agents in production, and 39% say their company has launched more than ten agents (Google Cloud, ROI of AI Study, September 2025). The survey covered 3,466 senior leaders across 24 countries and was conducted by National Research Group between April and June 2025 (Google Cloud press release, 4 September 2025).
One methodological detail accounts for much of the gap with other figures. Google Cloud surveyed only enterprises with more than $10 million in annual revenue that already had generative AI deployed. The 52% therefore describes a share of companies that had already cleared an earlier stage, not a share of businesses in general.
McKinsey measures depth rather than presence: 23% of organizations are scaling agents in at least one function, while 39% are still experimenting (McKinsey, State of AI, November 2025). Deloitte asks about intent: 74% expect their companies to be using AI agents at least moderately by 2027, and within that group 23% expect extensive use while 5% expect agents fully integrated as a core capability (Deloitte, State of AI in the Enterprise 2026, January 2026).
What each figure actually measures
The practical takeaway for reading any agent statistic: using agents and scaling them are different states, and expecting to use them describes intent rather than current deployment. A figure without its sample and definition cannot be compared with another figure.
Which companies get furthest with AI deployment?
The largest ones by revenue. Stanford publishes the distribution of AI deployment stages by company size, and the relationship is monotonic. Among companies with more than $5 billion in revenue, 39% are at the scaling stage and 10% report being fully scaled. Among companies below $100 million, those figures are 25% and 5% (Stanford AI Index 2026, Ch. 4).
The pattern reverses at the experimentation end. In the smallest revenue band, 39% are still experimenting; in the largest, 17%.
Stanford attributes this directly to resourcing: given the resource and investment demands of integration, larger companies were the most likely to report that their AI programs had reached a scaling phase.
Growth is uneven across regions. Reported AI use reached 88% across all geographies in 2025, with the largest year-over-year increases in China and Europe at 13 and 11 percentage points respectively. More than half of respondents reported AI in three or more business functions (Stanford AI Index 2026, Ch. 4).
Where are AI agents actually deployed at scale?
In the technology sector, in engineering and IT functions. Almost nowhere else.
Scaled agent use reaches 24% in software engineering, 22% in IT and 21% in service operations, and those figures describe the technology sector specifically. Across other industries and functions it stays in the single digits (Stanford AI Index 2026, Ch. 4).
Stanford adds an observation that explains the distribution: the business functions reporting the highest rates of agent use tend to be the same ones with broader, more established AI adoption. Agents are landing where a working base already exists rather than where expectations are highest.
Expectations for the future sit elsewhere. Deloitte finds the highest anticipated agent impact in customer support, with strong potential also cited for supply chain management, R&D and cybersecurity (Deloitte, January 2026).
How fast is agent usage growing inside deployed platforms?
Fast, when measured through a single platform's telemetry. Microsoft reports 15x year-over-year growth in active agents across Microsoft 365, rising to 18x in large enterprises (Microsoft, 2026 Work Trend Index, May 2026). An agent counts as active if it has at least one day of user-initiated usage in a 28-day period or completes at least one autonomous run in that period.
This figure does not compare directly with the McKinsey or Deloitte surveys. Telemetry measures activity inside one vendor environment among customers who have already bought and deployed the product, while the surveys measure the share of organizations in the market. The 15x growth describes intensity of use within an installed base rather than the spread of agents across the economy.
What ROI do companies report from AI and agents?
Reported figures run from 74% down to 5%, depending on who asked and what counts as a result. This is the section where cross-source comparison carries the most risk, so each figure below comes with its sample and its caveats.
- The optimistic end. Google Cloud reports that 74% of executives achieved ROI from generative AI within the first year. Among those reporting revenue gains, 53% cite growth of 6-10%, and among those reporting productivity gains, 39% saw productivity at least double (Google Cloud, ROI of AI Study, September 2025).
- The middle. McKinsey measures results at the level of financial reporting, and the picture changes. 39% of organizations report enterprise-level EBIT impact from AI, and for most of them that impact is below 5%. The share McKinsey classifies as AI high performers, meaning they attribute 5% or more of EBIT to AI, is 6% (McKinsey, State of AI, November 2025).
- A widely cited but older figure. An IDC study sponsored by Microsoft puts the return at $3.70 for every $1 invested in generative AI, rising to $10.3 among leaders. It draws on interviews with more than 4,000 business leaders (Microsoft news, January 2025).
- The pessimistic end. MIT, in The GenAI Divide: State of AI in Business 2025, reports that roughly 95% of organizations see no measurable P&L impact from generative AI, with around 5% of integrated pilots extracting significant value. The work draws on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public deployments (MIT NANDA, July 2025, via Fortune).
Why do ROI figures disagree so much?
Because three different things all go by the name ROI.
- The first is who answers. An executive assessing their own initiative and a set of financial statements give different answers. The 74% from Google Cloud is a self-report of perceived ROI; the 39% from McKinsey is claimed to have an EBIT impact. These are different levels of rigor rather than contradictory findings.
- The second is what counts as a result. A productivity gain for an individual worker, a revenue gain for a business unit and a change in enterprise EBIT pass through different numbers of filters. The higher the level, the smaller the share of companies that reach a measurable result.
- The third is who was asked. A sample restricted to companies with generative AI already deployed and revenue above $10 million will structurally return higher figures than a general sample.
An ROI figure without its sample, its definition of a result and the identity of whoever commissioned the study should not be carried into your own business case.
How many AI agent projects get canceled?
More than 40% by the end of 2027, according to Gartner. Three causes are named: escalating costs, unclear business value and inadequate risk controls (Gartner, 25 June 2025).
Gartner characterizes the state of the market at the time of the forecast as follows: most agentic AI projects are early-stage experiments or proofs of concept, largely driven by hype and often misapplied. The analysts add that most agentic propositions lack meaningful value or ROI, because current models do not have the maturity and agency to autonomously achieve complex business goals or to follow nuanced instructions over time. And separately: many use cases positioned as agentic today do not require an agentic implementation.
Company positioning in early 2025 was cautious. A Gartner poll of 3,412 webinar attendees in January 2025 found that 19% had made significant investments in agentic AI, 42% had made conservative investments, 8% had made none, and the remaining 31% were taking a wait-and-see approach or were unsure.
Why does the 40% cancellation figure keep appearing as new?
Because 2026 coverage cites it without a date. Gartner published the forecast on 25 June 2025 (Gartner). A year later, on 7 July 2026, Forbes published an analysis revisiting that forecast and examining why these projects actually die (Forbes, 7 July 2026).
From there the figure entered circulation detached from its 2025 origin, and much mid-2026 coverage presents it as fresh research. That does not make the forecast wrong, but its age matters: the prediction window is already half spent, and a reader who takes the figure as new will misjudge how much time remains before it can be tested.
A working rule for any statistic in this field: if a source does not date a figure, the figure is almost always older than it looks.
What governance do organizations actually have in place?
Considerably less than they have deployment plans. Deloitte finds that 74% of organizations expect to use agents at least moderately by 2027, while only 21% report having a mature governance model for agentic AI (Deloitte, State of AI in the Enterprise 2026, January 2026).
Deloitte adds an observation about the organizations that succeed: the most successful companies take a measured approach, starting with lower-risk use cases, building governance capability and scaling deliberately.
The foundation underneath is weaker than commonly assumed. Gartner reports that 63% of organizations either do not have the right data management practices for AI or are unsure whether they do (Gartner, 26 February 2025). The same research projects that organizations will abandon 60% of AI projects that are not supported by AI-ready data through 2026.
How much does AI actually improve productivity?
Stanford aggregates results from several studies, and the pattern holds: gains are largest where work is structured and measurable and where output is easy to monitor (Stanford AI Index 2026, Ch. 4).
The strongest measured results (Stanford AI Index 2026, Ch. 4):
- Customer support agents using a conversational AI assistant resolved 14–15% more issues per hour
- Developers using GitHub Copilot completed 26% more pull requests
- Marketing teams using multimodal AI for ad creation saw a 50% increase in output per worker
- Accountants using AI-based tools gained 55% in throughput
The downside is measured too. The most widely cited example is the METR study, in which experienced open-source developers became 19% slower with AI assistance, with a gap between how helpful they believed the tools were and how they actually performed. Stanford adds a qualifier that rarely travels with the figure: METR has not been able to replicate the result in later work, largely because developers grew reluctant to work without AI, and because developers in late 2025 were likely sped up by AI relative to the original study period (Stanford AI Index 2026, Ch. 4).
Learning is a separate case. Engineers who relied heavily on AI while picking up new libraries showed no measurable speed improvement and ran into what researchers call learning penalties.
Does this show up in macroeconomic data?
Partly, and not yet conclusively. A study of 12,000 European firms found a 4% increase in labor productivity from AI adoption, with training strengthening the outcome. In the United States, productivity growth reached 2.7% in 2025, nearly double the 1.4% average of the previous decade (Stanford AI Index 2026, Ch. 4).
Stanford cites one explanation for this: it may reflect the early stage of a J-curve, where organizations absorb the costs of adoption first and the larger productivity gains appear in the numbers later. OECD projections for G7 economies estimate annual productivity gains of 0.2 to 1.3 percentage points over the next decade.
There is a signal pointing the other way. A survey of 6,000 executives across four countries found widespread adoption alongside minimal realized productivity gains and a projected 0.7% reduction in employment over the next three years.
Stanford states the position on the evidence directly: productivity gains are measurable within narrow tasks, while at the macro level the evidence remains early and mixed.
What do analysts project for AI agents through 2035?
Gartner's projections describe agents becoming embedded in enterprise software over the next decade. These are forecasts rather than observed data, and they should be read with that in mind.
By the end of 2026, 40% of enterprise applications will be integrated with task-specific AI agents, up from less than 5% at the time the forecast was published (Gartner, 26 August 2025). Gartner analyst Anushree Verma describes the trajectory as agents evolving rapidly, moving from the basic assistants embedded in enterprise applications today to task-specific agents in 2026 and eventually to multiagent ecosystems by 2029.
Two further figures sit at the 2028 horizon: at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from 0% in 2024, and 33% of enterprise software applications will include agentic AI, up from under 1% in 2024 (Gartner, 25 June 2025).
The furthest estimate concerns revenue. Under Gartner's best-case scenario, agentic AI could drive roughly 30% of enterprise application software revenue by 2035, surpassing $450 billion, up from 2% in 2025 (Gartner, 26 August 2025).
Worth holding alongside these figures: the same analysts project that more than 40% of agentic AI projects will be canceled by the end of 2027 (Gartner, 25 June 2025). Both forecasts can hold at once, because agents shipping inside vendor products and individual enterprise projects succeeding are separate things.
Where is AI investment actually going?
Here the data is observed rather than projected. Global corporate AI investment reached $581.69 billion in 2025, a 129.9% increase over the previous year. Private investment made up the largest share at $344.66 billion, up 127.5% (Stanford AI Index 2026, Ch. 4).
Generative AI companies accounted for $170.9 billion of that, nearly half of all private investment and an increase of over 200% from 2024.
The breakdown by focus area shows where the money concentrates. The largest share went to AI infrastructure, models, research and governance at $143.2 billion. AI agents as a focus area drew $8.02 billion, per the report's focus-area breakdown.
Investment is flowing mainly into the layer agents run on rather than into agent products themselves. That tracks with what the adoption data shows: infrastructure is being built faster than agents are reaching scale.
Conclusion
The 2026 data describes a market where the infrastructure is already in place and agents have reached scale in a narrow set of technology-sector functions. Figures diverge because of differing definitions and samples rather than conflicting sources, so any adoption or ROI number is only meaningful alongside what was measured and who was asked.
If you are working out where agents fit into your own processes and want a clear-eyed assessment rather than market averages, talk to the Tensorway team.
