Twenty-three percent. That is the share of organizations that told McKinsey they were scaling an AI agent system anywhere in their business, in its State of AI survey fielded in mid-2025 and published that November. Another 39% were experimenting. Put those together and you get the headline number vendors like to quote, 62% engaged with agents. Separate them and you get a more useful picture: most companies are testing, a minority are scaling, and in any single business function, no more than 10% of respondents reported scaling agents.
That gap, between interest and deployment, runs through every credible data set on AI agents. This page collects the AI agents statistics that come from named surveys and analysts, with dates, and flags where definitions differ. Use them to calibrate claims you hear from vendors, from your board, or from your own team.
How many companies use AI agents
Adoption numbers vary widely, mostly because surveys define “agent” and “use” differently.
McKinsey’s “The state of AI in 2025: Agents, innovation, and transformation,” published November 5, 2025, surveyed respondents from June 25 to July 29, 2025. It found 62% of organizations at least experimenting with AI agents. Of those, 23% of all respondents said they were scaling an agentic AI system somewhere in the enterprise, meaning expanding deployment within at least one business function. The other 39% had begun experimenting. Most of those scaling said they were doing so in only one or two functions.
PwC’s AI agent survey of 308 US business executives, run April 22 to 28, 2025, produced higher numbers. Seventy-nine percent said AI agents were already being adopted in their companies. Of those adopting, 66% said agents were delivering measurable value through increased productivity. Eighteen percent said their companies were not using agents at all, and the top reason was a lack of clear use cases or business value.
A follow-up PwC analysis published in August 2025 broke adoption down further: more than a third of companies had adopted agents broadly, another 27% in a limited way, and 17% had adopted them fully across the company.
Why the gap between McKinsey and PwC? PwC itself noted that many employees are using agentic features built into enterprise apps for routine tasks like updating records and answering questions. A survey that counts those features as adoption will report higher numbers than one asking about scaled agent systems. Neither is wrong. They measure different things.
Experimenting versus scaling
The more telling statistics are about depth, not breadth.
In McKinsey’s data, no single business function had more than 10% of respondents reporting scaled agent use. Agent use was most often reported in IT and knowledge management, functions where tasks tend to be well documented and errors are easier to catch.
McKinsey also identified a group of AI high performers, about 6% of respondents, who reported significant value from AI. These organizations were more likely to redesign workflows rather than add AI to existing ones, and to scale faster.
PwC reached a similar conclusion from the other direction. Fewer than half of the companies adopting agents in its survey were fundamentally rethinking operating models, and it warned that broad adoption “doesn’t always mean deep impact.”
Spending intent is high anyway. In PwC’s survey, 88% of executives said their team or function planned to increase AI-related budgets in the next 12 months because of agentic AI. And 46% said they were concerned their company might be falling behind competitors in adopting agents. Fear of missing out is a real driver in these numbers.
Forecasts and failure rates
The most quoted forecasts come from Gartner, in a press release dated June 25, 2025.
Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. Its senior director analyst Anushree Verma said that most agentic AI projects at the time were early-stage experiments or proofs of concept “driven by hype” and often misapplied.
In the same release, Gartner made two growth predictions. At least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024. And 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.
Those three numbers can look contradictory. They are not. Gartner’s view is that many early projects will fail while the technology still spreads into software products. A company can cancel its own custom agent project and still end up using agentic features inside its CRM or service desk software.
Read failure-rate figures with care. Of all the AI agents statistics on this page, the cancellation forecast is the one most often quoted without context. A canceled pilot is not always a loss. Some pilots exist to learn what does not work, and stopping them on schedule is the right outcome.
Effects on jobs and headcount
Workforce statistics about AI agents are thinner and noisier than adoption statistics, and most surveys measure AI broadly rather than agents specifically.
In McKinsey’s 2025 survey, a plurality of respondents expected little or no change in total headcount over the following year. But 32% predicted an overall reduction of 3% or more, and 13% predicted an increase of that size. Respondents at larger organizations were more likely to expect reductions.
More recent Gartner surveys point to reductions already happening in specific functions. A Gartner survey published November 5, 2026, found that 21% of customer service and support leaders have reduced human agent headcount because of AI. Another Gartner survey, published October 28, 2026, found that 47% of supply chain organizations reduced headcount, by an average of 10.5%, over the past two years, primarily due to AI.
Note the wording. These surveys attribute changes to AI in general, which includes agents but also chatbots, analytics and automation tools. Be cautious about any statistic claiming to isolate the job impact of agents alone.
Customer service is where most observers expect the effect of agents to show first, because so much of the work is repeatable and measured: ticket volume, resolution time and satisfaction scores. That makes it the function to watch for honest AI agents statistics on jobs. If headcount reductions in service teams hold steady while customer satisfaction holds or rises, other functions will follow. If satisfaction drops, expect some of those reductions to reverse.
Agent washing and the vendor market
One of the most useful numbers for buyers is about vendors, not users.
Gartner estimated in June 2025 that only about 130 of the thousands of agentic AI vendors are real. The rest, it said, were engaged in “agent washing,” rebranding existing products such as AI assistants, robotic process automation and chatbots as agents without substantial agentic capabilities.
For anyone evaluating vendors, that estimate argues for a simple test. Ask what the product does when a step fails. A real agent plans, takes actions across systems, checks results and adjusts. A relabeled chatbot answers questions and hands off. The demo will not always reveal the difference. A pilot on your own messy data will.
How to read an AI agent survey
Most of the confusion around agent statistics comes from four questions that a headline number does not answer. Check each one before you repeat a figure in a board deck or a press release.
First, who was asked? McKinsey surveys a global mix of respondents across company sizes and roles. PwC’s agent survey asked 308 senior US executives. A survey of executives at large companies will report higher adoption than a survey of all businesses, because large companies buy more enterprise software with agentic features built in.
Second, how was “agent” defined? Some surveys describe agents as systems that plan and carry out multi-step tasks on their own. Others leave the term open, which lets respondents count a chatbot or a writing assistant. The looser the definition, the higher the number.
Third, what counts as use? Experimenting, piloting, adopting in one team, and scaling across functions are different stages. McKinsey separates them. Many vendor surveys do not.
Fourth, when was the survey in the field? McKinsey’s data was collected in June and July 2025, and PwC’s in April 2025. In a market moving this fast, a year-old survey can understate current use, and a prediction made in 2025 may already be revised. Always cite the field dates, not just the publication date.
A fifth check is worth adding when a survey comes from a company that sells agents. That does not make the numbers false, but it does mean the questions may be framed to produce a strong adoption figure. Look for the sample size and the exact question wording. If neither is published, treat the number as marketing.
The Agent Reality Gap, and what it means for brands
To read any agent statistic quickly, we use a single ratio we call the Agent Reality Gap: the share of organizations experimenting with agents divided by the share scaling them. In McKinsey’s 2025 data, that is 39% experimenting against 23% scaling, a gap of about 1.7 to 1 across the whole enterprise. Inside any single business function, where scaling never exceeded 10%, the gap is far wider. When a vendor or survey quotes a single adoption number, ask for both halves of the ratio. If they can only give you the combined figure, the gap is probably large.
There is a second implication for marketing and communications teams. As agents move into purchasing and research tasks, some of your future buyers will be software reading the web on a person’s behalf. An agent asked to shortlist vendors will draw on the same sources AI assistants already cite: independent coverage, review sites, and clear, consistent facts about what your company does. Brands that are hard to describe accurately today will be hard for agents to recommend tomorrow.
Treat every new batch of AI agents statistics the same way: split the headline number into experimenting and scaling before you believe it. Watch the gap between experimenting and scaling over the next two survey cycles, because when it closes, the agents reading about your company will matter as much as the people.