A new reference dataset from bdautomated is challenging widely repeated claims about AI agent adoption by tracing 75 statistics back to their original surveys and reports.
Published on September 15, 2026, the project documents the question asked, the respondent group, sample size, timing and precise wording behind figures reported by 18 publishers. The accompanying data is available in CSV and JSON formats under a CC BY 4.0 licence.
Different definitions produce very different figures
The review found that prominent estimates often measure different stages of adoption. McKinsey’s 2025 global survey reported that 62% of organizations were at least experimenting with AI agents, while 23% had scaled an agent somewhere in the company. Within any individual industry, however, the share that had scaled an agent was no higher than 10%.
Other surveys produced larger figures because they included adoption plans or broader forms of use. In a PwC survey conducted in April 2025, 79% of U.S. executives said agents were already being adopted at their firms. A Capgemini survey that more closely examined what respondents meant by “agent” found that 14% had implemented one.
Separate data from the U.S. Census Bureau showed that 19.8% of companies across industries and size categories were using artificial intelligence in at least one business function as of May 2026. The figure covers AI use generally, rather than AI agents specifically.
Dataset puts disputed claims in context
The project also examines a widely circulated MIT Project NANDA finding that “95 percent of organizations are getting zero return.” According to the review, that result measured profit-and-loss impact within roughly six months of a pilot. It was based on 52 interviews, 153 conference survey responses and 300 public deployments, and the authors described the findings as preliminary.
As a result, the statistic does not establish that 95% of AI projects fail. The dataset also distinguishes between observed results and forecasts. Gartner’s prediction that more than 40% of agentic AI projects would be cancelled by the end of 2027 was issued in June 2025; the cancellations have not yet been counted.
Each entry in the database was checked to confirm that the number appears in the original document, with its context and a verbatim quote recorded. The review also identifies who was surveyed, how many respondents participated and when the research was conducted. Conflicting figures are displayed alongside one another instead of being averaged.
Market-size forecasts were excluded because the underlying reports were paywalled and could not be independently checked. The website includes a corrections address, four embeddable charts and a table covering all 75 figures.
“Two headlines within the same week said virtually no one has AI agents working and virtually everyone does, and each was quoting real surveys,” a bdautomated spokesperson said. The company said the project was intended to show businesses what each survey actually measured before applying a statistic to their own circumstances.
