How The AI Adoption Rate Is A Clock
Ajay Pundhir, senior AI leader & founder of AskAjay.ai. Believes AI should amplify human expertise, not replace it.
gettyTwo statistics in the same Federal Reserve note describe the same American economy: 18% of firms have adopted AI, and 78% of the labor force works at a firm that has.
Both numbers appear in an April 2026 FEDS Note from the Federal Reserve Board, which compiles three separate federal instruments and reports what each found. Neither number is wrong. They count different populations. One counts firms, and most firms in America are small. The other weights by employment, and large firms adopt earliest and most. A third figure in the same note, drawn from a household survey, puts work-related generative AI use at about 41% of the workforce.
Read the first; you are ahead of the field. Read the third; you are behind. Same month, same government, same underlying reality.
I have watched this play out in enough investment committees to know that the adoption statistic is never really the justification. Nobody approves a three-year AI program because a survey says 88% of companies have started one. The business case carries the justification.
What the statistic does is start a clock.
That distinction matters more than it might seem. The clock is what buys the concessions. When a plan opens on a slide showing that most of your industry has already moved, the argument shifts from whether to fund to how fast, and speed is expensive in ways that never appear as a line item. Compressed vendor selection. Diligence deferred to phase two. A pilot was promoted before its own evaluation finished. Those are all purchases, and the currency is the urgency premium sitting on slide two.
So the question worth asking about a peer-adoption figure is not whether it is accurate. It is what pace it licensed and whether the population it counted resembles the company it is being compared to.
Here is how much a measured adoption rate can depend on a choice nobody in the boardroom made.
Until late 2025, the U.S. Census Bureau’s Business Trends and Outlook Survey asked firms whether they had used AI “in producing goods or services.” On November 17, 2025, according to the Bureau’s own record of the question change, that became whether they had used AI “in any of its business functions.” Cognitive testing had found the old phrasing misfiring: Some firms did not see themselves as producers of goods or services and skipped the question, while others said no and then described, later in the same interview, using AI in hiring, accounting or project management.
The wording changed. Behavior did not. And the number moved enough that the Census Bureau took the unusual step of declaring the old and new figures incomparable, creating a new time series from December 4, 2025, and filing the earlier data under historical.
A statistical agency broke its own continuity because a single clause moved the answer. That is the strongest available evidence that adoption rates measure a question as much as they measure an economy.
The damage is not confined to the benchmark slide. It reaches the objective.
Consider a board that adopts “reach peer adoption within eighteen months” as a strategic goal, having sourced peer adoption from a survey where a firm counts as an adopter if any function anywhere uses AI at all. Now consider that the same board’s internal reporting counts a function as adopted only when a system is deployed, integrated and measured.
Those two definitions are separated by an enormous amount of work. A company can pass the survey’s bar in a quarter and spend three years reaching its own. Set the target against the looser count and measure against the stricter one, and the organization will run permanently behind a goal it could never have reached, which is a reliable way to keep buying speed it does not need.
In March, in my own column on how the AI market is splitting, I used industry adoption and profit-impact figures as peer context for an argument about which companies were pulling ahead. I did not name the population those figures counted. I have also, on my own site, told executives to bring exactly that kind of peer benchmark into a board conversation.
The figures were real, and the argument holds. The practice was sloppy, and it is the practice this piece is about.
The most quoted adoption figure of the past year, that 88% of organizations now use AI in at least one function, comes from a McKinsey & Company survey series. Stanford HAI’s AI Index reproduces it, credits it clearly and adds its own caution in the same chapter: Figures like these “are self-reported and should be viewed as directional rather than comprehensive.”
The caveat travels with the number in the source. It falls off somewhere between the report and the deck.
None of this argues for slower AI investment. Some companies should be moving faster than they are. It argues that the speed should come from a decision rather than from a statistic whose population nobody checked.
One sentence in the investment paper is enough. Name the survey, name who it counted and state what pace that justifies for this company. If the number will not survive being described precisely, it should not be setting the schedule.
A benchmark whose population nobody in the room can name is not a schedule, and the pace it produced is not a plan.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
