Subscribe
AI
7 min read

Every AI Adoption Number Is Correct. None Are Comparable

The enterprise AI numbers do not add up, and that is becoming the most reliable finding in the entire literature. In October 2026, Wharton Human-AI Research and GBK Collective published the fourth annual installment of their AI adoption study. Among roughly 1,400 US decision-makers and employees, 62 percent now use AI at work daily and 91 percent use it more than weekly. Eighty percent report positive returns on their AI investments (full report, press release). Six weeks earlier, McKinsey's s

Every AI Adoption Number Is Correct. None Are Comparable

The enterprise AI numbers do not add up, and that is becoming the most reliable finding in the entire literature.

In October 2026, Wharton Human-AI Research and GBK Collective published the fourth annual installment of their AI adoption study. Among roughly 1,400 US decision-makers and employees, 62 percent now use AI at work daily and 91 percent use it more than weekly. Eighty percent report positive returns on their AI investments (full report, press release).

Six weeks earlier, McKinsey's state of AI survey (1,719 respondents across 97 countries, fieldwork May 4 to June 8, 2026) found that nearly nine in ten respondents report regular AI use in at least one business function. But only 37 percent attribute any impact on earnings before interest and taxes to AI, and the share of so-called AI high performers, those crediting at least 5 percent of EBIT to AI, has sat flat at 6 percent for two consecutive years (McKinsey).

The US Census Bureau, whose Business Trends and Outlook Survey is a nationally representative count of actual businesses, found that 19.8 percent of US firms used AI in the two weeks ending May 3, 2026 (Census). Ramp's AI Index, which counts observed card payments to AI vendors, put adoption at 55.9 percent of the 85,000-plus US businesses on its platform in September 2026 (Ramp). And MIT's Project NANDA reported in August 2025 that roughly 95 percent of enterprise generative AI pilots produced no measurable impact on the P&L (Fortune).

Daily use at 62 percent. Firm adoption at 19.8 percent. Pilot failure at 95 percent. All of these numbers are honest. None of them are comparable, and the gap between them is where the actual information lives.

Who is in the denominator

The first thing each survey counts is a different unit. Wharton counts people: managers and employees at US companies with at least 50 million dollars in revenue. McKinsey counts respondents, then weights results by each country's contribution to global GDP; 36 percent of its sample work at organizations with more than 1 billion dollars in revenue, and one respondent standing in for a 200,000-person employer is still one respondent. The Census Bureau counts firms, one establishment one vote. Ramp counts businesses that actually paid money for AI that month. MIT counts pilot projects.

The denominators are not just different in size, they disagree about what deserves to count. The Census Bureau's 2026 AI supplement, its first attempt at a structured national measurement, resolved the apparent paradox between low firm-level adoption and high everyday visibility by measuring both. In the November 2025 to January 2026 reference period, 18 percent of US firms used AI in a business function. Weighted by employment rather than firm count, the same data reads 32 percent, with usage reaching 50 to 60 percent among very large firms in information, professional services, and finance (Census research brief).

That is the whole reconciliation in one pair of numbers. AI adoption is concentrated in the employers that hire the most people. A majority of American establishments do not use AI, and it is still true that a third of the workforce works somewhere that does. When a vendor cites adoption from a survey of executives at large companies, they are reporting the employment-weighted view with the firm-count caveat quietly removed.

Four gauge dials of different sizes and readings mounted on one control panel

What counts as use

The second disagreement is depth. The measures form a ladder, and each rung is easier to claim than the one above it.

An observed payment (Ramp's bar) clears as soon as a company buys a subscription, whether or not anyone logs in. A two-week recall window (the Census question) clears when staff touch AI on any task, including free tools that never generate a transaction. "Regular use in at least one business function" (McKinsey) requires only one function, not the enterprise. Wharton's 62 percent daily figure measures personal habit among surveyed managers, not organizational integration. MIT's bar, production deployment, is the hardest: of companies evaluating custom enterprise AI tools, about 60 percent evaluated, roughly 20 percent reached a formal pilot, and only about 5 percent reached production (report write-up).

The Census supplement puts a number on shallowness that headline adoption rates usually hide: among firms that do adopt, 57 percent confine AI to three or fewer business functions, and 65 percent restrict it to three or fewer worker tasks. Wharton's own 2026 data says the same thing from the other direction. Daily use is 62 percent, but AI is fully embedded in core workflows at only 28 percent of organizations. Use is nearly universal and mostly shallow, simultaneously.

One more ladder worth naming: Stanford Digital Economy Lab's Adoption Monitor, launched June 10, 2026, tracks worker adoption of generative AI for work at 62 percent as of May 2026 while survey firms expect a net move of only 6 percent from non-use to use over the next three years (Stanford DEL). Workers adopt faster than institutions authorize. The MIT report had the same observation, calling it shadow AI. The gap between the two numbers is where compliance teams now live.

What counts as value

The third disagreement is the one that matters most for budgets: the return figures are self-reported attributions, not measurements.

McKinsey's 37 percent who see EBIT impact are respondents crediting AI for earnings changes they cannot decompose. Its 6 percent high performers are respondents who both claim a 5 percent EBIT threshold and describe the impact as significant, a definition with two self-graded gates. Wharton's 80 percent positive returns is likewise an opinion with a receipt attached. The same Wharton survey shows why attribution is so cheap to grant: departments that evaluate AI on growth outcomes report strongly positive returns at 35 percent, while departments that judge AI purely on efficiency report strongly positive returns at 13 percent (study summary). The scoreboard changes the score.

MIT's 95 percent is the counter-extreme: pilots that could not show a measurable P&L effect at all, across roughly 30 to 40 billion dollars of enterprise investment. Read as a failure rate for AI it is wrong, and read as a measurement rate it is brutal. Most pilots were never instrumented to detect financial impact, which is exactly the finding the researchers stress. The number describes the accounting, not the models.

Two years into the argument, the pattern across sources is stable. Individual-level benefit is real and broad: McKinsey finds 80 percent of respondents report improved personal productivity. Organizational-level financial impact is narrow and flat: the 6 percent figure has not moved. Wharton's tracker caught sentiment finally splitting from usage, the first year in the series the two moved in opposite directions: daily use rose 16 points, but executives describing themselves as excited fell 8 points while cautious and skeptical each rose. Adoption is not consent, and enthusiasm is not impact.

The measurement is getting professional

The most useful response to the mess has been institutional. Stanford Digital Economy Lab's AI Economic Indicators (June 2026, led by Erik Brynjolfsson and Connacher Murphy) publish monthly dashboards that separate adoption surveys from administrative data, and print their own disagreements: the Adoption Monitor explicitly notes that Hartley et al. find AI work adoption decreasing while Gallup and Bick, Blandin and Deming find it rising toward 50 percent (launch note). The Census 2026 AI supplement is the first nationally representative attempt to measure firm use, function-level deployment, and worker task use as three distinct layers.

Administrative series also come with their own traps. The Census Bureau broadened its BTOS AI question in November 2025, from AI used to produce goods or services to AI used in any business function, and measured adoption jumped. Part of that apparent surge is questionnaire design (Motley Fool review of Census data). A series can move without behavior moving.

State of AI's October 2026 review of the evidence settled on the right standard: every percentage needs a population, a definition, and a period, and until it has all three it is not ready to support a market-wide claim (State of AI). That is a low bar and most published adoption numbers still miss it.

How to read the next adoption headline

Before citing any of these figures at a board meeting or in a vendor pitch, run five questions:

  1. Who is counted: people, firms, or transactions? Executive panels overstate the firm-level view; employment-weighted data overstates the typical company.
  2. What cleared as use: a payment, a two-week touch, regular use in one function, or production deployment? The same economy reads anywhere from 5 percent to 91 percent depending on the rung.
  3. How deep does it go: how many functions and tasks? Most adopting firms are at three or fewer.
  4. Who graded the return: attribution is not measurement, and the gap between 80 percent positive and 6 percent high-performing is the gap between vibes and books.
  5. Did the question change mid-series? The Census rewording jump and Wharton's 2026 addition of employee perspectives both break naive year-over-year comparisons.

None of this makes enterprise AI look small. It makes it look concentrated, shallow, and under-measured, which is a different and more useful claim. The technology diffuses at the speed of habits and the economy measures at the speed of accountants. For now, the honest number is a range with four names attached, and anyone who quotes just one of them is selling something.

Sources

AIAnalysis