What happened
In “AI adoption in 2026: use, spending, and returns,” Nathan Benaich of Air Street Capital compares three estimates reviewed on October 31, 2026. The report’s central point is that each estimate observes a different population and behavior. They are separate observations, not a league table or a time series, and the percentages should not be averaged into a global adoption rate.
McKinsey’s 2026 survey found that nearly nine in ten respondents reported regular AI use in at least one business function. It surveyed 1,719 respondents in 97 countries, with fieldwork from May 4 to June 8, 2026; results are weighted by countries’ contributions to global GDP. Thirty-six percent of respondents worked at organizations with revenue above $1 billion. The measure captures respondent-reported organizational use. It is not a count of all firms and does not establish enterprise-wide deployment.
Ramp’s AI Index reported 55.9% for September 2026. It covers more than 85,000 businesses in Ramp’s US transaction sample and counts adoption when a positive payment for an AI product or service is observed during the month. That offers evidence of buying among the businesses visible to the index, not proof that a paid tool is actively used, deployed in production, or generating a return. Free tools and purchases made outside Ramp may be missed.
The US Census Bureau’s Business Trends and Outlook Survey (BTOS) reported 19.8% for the period ending May 3, 2026. It is a nationally representative US business survey asking about AI use in a business function during the previous two weeks. The report cautions that this is a dated observation, not an October estimate. The question broadened in November 2025, breaking comparability with earlier estimates focused on AI use to produce goods or services.
The comparison explains why the figures diverge: respondents in a global survey, customers visible in a US payments sample, and businesses in a national survey are different denominators. A business may use a free tool without an observed payment, pay for a subscription that sees little use, or deploy AI in one team without changing the wider organization. Measurement periods differ too: a monthly payment observation is not interchangeable with a two-week use question. The page complements the State of AI Report’s annual business analysis of spending intensity, inference economics, and workflow change; it is a methods comparison, not a new survey. Its update policy says the source reviews the comparison monthly and will record definition changes rather than splice unlike observations into a seemingly continuous series.
What the archive adds
An earlier Telegram post (April 19) argued that adoption is a proxy for activity, not evidence that work has changed. This comparison adds a specific measurement framework: separate reported use, observed spending, and financial impact, each with its own population and observation window. It is a new fact on a known measurement concern, not a restatement of the same data.
Trend signal
No supplied SITE_PATTERNS entry directly addresses the distinction between adoption measures, denominators, and business returns. The report reinforces a measurement discipline, but linking it to a broader agent or software pattern would overstate the connection.
Why now
The comparison was reviewed on October 31, 2026, and incorporates September Ramp data alongside the 2026 McKinsey survey and the Census observation ending May 3. Those dates make the comparison timely, but they do not make the three estimates simultaneous measurements of the same thing.
Why it matters
A team choosing an AI budget or reporting progress can get a misleading picture if it treats a survey response, a vendor payment, and a production workflow as equivalent evidence. These measures answer different questions: whether respondents say their organization uses AI, whether a payment was visible in one platform’s sample, and whether businesses reported recent use. None alone establishes that a workflow works reliably or pays for itself.
For a deployment, the report recommends tying use to a defined task and outcome. Track completion quality, elapsed time, human review, error rates, and full operating cost, then compare with the previous workflow under similar conditions. A payment or a larger volume of generated tokens is evidence of activity, not sufficient evidence of value. For market claims, publish the population, period, and definition beside the percentage; for internal decisions, measure the workflow the team actually intends to improve.
Limits and uncertainty
This page compares findings attributed to McKinsey, Ramp, and the Census Bureau; it does not collect new survey data or independently establish the underlying estimates. The source notes that Ramp’s live index changes as its data is updated, and that its transaction coverage can miss free tools and purchases outside Ramp. The McKinsey figure is respondent-reported, while the Census figure covers a specific two-week period and follows a questionnaire change in November 2025. The report does not show that any one measure predicts successful deployment, nor does it prove that AI caused reported financial outcomes. McKinsey’s finding that 37% of respondents attributed some positive impact on earnings before interest and taxes to AI is not a causal estimate.
Skill to build now
Before presenting an adoption percentage, write down its denominator, observation period, and counted behavior. Then choose a separate outcome measure for the workflow: completion quality, time, review burden, errors, and full cost. Keep the baseline comparable. This small measurement card helps prevent “paid,” “used,” “deployed,” and “valuable” from collapsing into one claim.
Reader takeaway
There is no meaningful single AI adoption rate without a definition. Treat reported use, observed purchases, production deployment, and financial returns as different stages of evidence—and choose the measure that answers the decision your team needs to make.