Ask three reputable sources how many businesses use AI and you get nearly 90 percent, 55.9 percent, and 19.8 percent. None of them is wrong. A comparison published by State of AI, written by Nathan Benaich of Air Street Capital and reviewed on 31 October 2026, lays out why the numbers diverge and what each one can honestly support.

McKinsey’s 2026 survey found that close to nine in ten respondents report their organizations use AI regularly in some part of the business. The survey reached 1,719 people in 97 countries between 4 May and 8 June 2026, and results are weighted by each country’s share of global GDP. Thirty-six percent of respondents come from organizations that bring in more than $1 billion a year. That skews the sample toward large companies, and it measures what a respondent says, not what a census of firms would find.

The Ramp AI Index puts adoption at 55.9 percent in September 2026. Ramp counts a business as an adopter if it paid for any AI product or service in the month, drawing on more than 85,000 businesses in its sample. This is observed buying, which is harder to exaggerate than a survey answer. It also misses free tools, purchases made outside Ramp, and any sign that the paid product gets used.

The US Census Bureau’s BTOS business survey reports 19.8 percent for the period ending 3 May 2026. It is a nationally representative survey of US businesses, and it asks whether a business function used AI during the prior two weeks. The Census Bureau widened that question in November 2025, which breaks comparison with earlier estimates that covered only firms making goods or delivering services. It is also a May reading, so it should not be set beside an October one.

The gap between 19.8 and nearly 90 comes down to who is in the denominator. Managers around the world, customers of one US spending platform, and a representative slice of American firms are three different populations. Averaging them would produce a number that describes nothing. A time window matters just as much: a payment in a month and a use report over two weeks are different events.

Adoption is also separate from payoff. State of AI notes that McKinsey’s survey reports 37 percent of respondents see some positive effect on EBIT (earnings before interest and taxes) from AI. That is a self-reported outcome, not proof of how profits would have looked without it. A company can pay for a subscription nobody opens, or run AI in production on one team while the rest of the business stays unchanged.

For anyone building a market model or a sales deck, the practical fix is cheap. Put the population, the period, and the definition next to every percentage you quote, and pick the measure that matches the question: survey reach for breadth, payment data for paid demand, workflow testing for production use. Benaich’s page states that a figure without those three details is not ready to carry a market-wide claim, and a pitch that quotes “90 percent adoption” without them is exactly that.

Reported by State of AI (stateof.ai), analysis by Nathan Benaich of Air Street Capital, reviewed 31 October 2026.