AI Adoption Numbers Conflict Because Surveys Measure Different Things

The Data Gap Behind the Headlines

Two surveys published in the same week in 2025 reached conclusions that looked irreconcilable: almost no company had AI agents running in production, and almost every company did. Both surveys were real. Both findings were accurate. The conflict came entirely from what each one decided to count as “adoption.”

That definitional chaos is the focus of a reference dataset published September 15, 2026 by bdautomated, a company that builds AI agent systems for business clients. The team traced 75 widely circulated AI statistics back to their original documents – recording sample sizes, question wording, dates, and verbatim quotes – and released the full dataset as a free CSV and JSON download under a CC BY 4.0 licence.

What the Major Surveys Actually Found

McKinsey’s 2025 global survey offers the clearest breakdown of where the numbers actually land. Among the organizations it reached, 62 percent said they were at least experimenting with AI agents. That figure drops to 23 percent when restricted to companies that had scaled an AI agent somewhere in the business – and in any single business function, the share with a deployed agent was no higher than 10 percent. The three numbers describe the same population; the differences come from how far along the adoption curve each threshold sits.

PwC’s April 2025 survey of U.S. executives returned a very different figure: 79 percent said AI agents were already being adopted in their companies. Capgemini ran its own check of what survey respondents actually meant when they said “agent” and found 14 percent had implemented one. For a baseline across all U.S. businesses of every size, the Census Bureau measured 19.8 percent using AI in any business function as of May 2026. The spread between 14 percent and 79 percent is not a sign that one researcher is wrong – it reflects surveys with different questions, different samples, and different definitions of the word “using.”

bdautomated applied four checks to every figure before including it in the dataset: the number had to appear in the original source document; the exact location and a verbatim quote had to be recorded; the measurement had to be described in plain language, specifying who was asked, how many people, and when; and it had to be shown alongside conflicting sources rather than averaged with them. Market-size forecasts were excluded entirely because the underlying reports are behind paywalls and could not be independently verified.

The “95 Percent Failure” Figure, Explained

Among the statistics the dataset addresses, the most consequential for market sentiment in 2025 was MIT Project NANDA’s claim that 95 percent of organizations were getting zero return from AI. That figure moved through the press in ways that significantly overstated its scope.

What MIT Project NANDA actually measured was profit-and-loss impact within approximately six months of a pilot deployment. The sample consisted of 52 interviews, 153 conference survey responses, and 300 public deployments. The authors of the study described the findings as preliminary. The statistic does not say 95 percent of AI projects fail – it says most pilots had not produced a measurable P&L impact within a short measurement window, in a particular and limited sample.

Gartner’s Cancellation Forecast and What Counts as Canceled

A separate figure circulating from mid-2025 onward comes from Gartner: a prediction that more than 40 percent of agentic AI projects will be canceled by the end of 2027. The bdautomated analysis flags this as a forecast issued in June 2025, not a count of actual cancellations. As of the dataset’s publication date, no count of project cancellations exists. The figure describes what Gartner’s analysts expect to happen, not what has already happened.

That distinction carries weight when the same statistic appears in planning documents or budget presentations. A forecast from a research firm about future cancellation rates is a different category of evidence than a survey of organizations that have already shut down AI initiatives. Treating the two as equivalent produces the same distortions that inflated AI adoption figures in the first place – the measurement and the thing being measured drift apart until the number no longer describes reality.

The dataset’s release comes at a point when corporate AI spending decisions increasingly reference analyst statistics without checking the conditions under which those statistics were produced. A business evaluating whether to deploy an AI agent faces a choice between a PwC headline suggesting 79 percent of competitors have already moved and a Capgemini figure suggesting fewer than one in seven has actually implemented anything. The gap between those two numbers – 65 percentage points – could shape investment decisions in entirely opposite directions.

bdautomated’s explanation for building the dataset was direct: “Two headlines in the same week said almost nobody has AI agents running and almost everybody does, and both were quoting real surveys. We wanted the page we could not find: what each survey actually asked, so a business owner can tell which number is about a company like theirs.” The page includes four embeddable charts and a full table of all 75 figures with sources, sample sizes, dates, and quotes. The corrections address is public.

A Reference Tool, Not a Verdict

The dataset does not attempt to declare a “real” AI adoption rate. Its structure is built around showing disagreement rather than resolving it. Each figure is tagged with what it counts, which makes it possible to identify whether a given statistic describes companies similar in size, industry, and deployment stage to the one doing the research.

What the analysis makes visible is the underlying problem: the term “AI agent” has no shared definition across the survey industry, which means figures from different publishers are not measuring the same object. McKinsey’s 10 percent for any single business function, Capgemini’s 14 percent for organizations that implemented an agent after re-checking the definition, and PwC’s 79 percent for executives who said adoption was happening – all three could be simultaneously accurate, describing different slices of the same business population at different points on a spectrum that has no agreed endpoints.

The Census Bureau’s figure – 19.8 percent of all U.S. businesses using AI in any business function as of May 2026 – is the only one in the dataset that covers the full range of business sizes rather than a corporate or executive subsample. That makes it the most inclusive number available, and also the least specific: “any business function” and “any AI” set a lower bar than most enterprise AI deployments would recognize.

Gartner’s 40 percent cancellation forecast, set against the MIT Project NANDA preliminary finding about short-term P&L impact, suggests the industry is heading toward a period of consolidation – but neither figure tells a company whether its own AI initiative is likely to survive or produce returns. The question of whether 2027 cancellations will cluster in specific industries, deployment types, or vendor relationships remains entirely unaddressed by the existing data.