AI Adoption, Drawn to Scale
Two pictures of how far AI has actually spread, drawn one dot at a time, with every figure sourced and dated on the artifact.
Most of the world has not used AI, and most large enterprises have not moved it into production. Both of those are measurable, and both are measured by named institutions on a published schedule. A number without a source and a date is a slogan. These two charts carry theirs in the footer, because that is the difference between a picture you can defend in a room and one you cannot.
Everyone, drawn to scale
Every one of the 2,500 dots below is about 3.3 million people, and all 8.2 billion of us are on the page. Nothing is cropped and nothing is hidden in an "other" bucket, which is what makes the proportions arguable rather than decorative.
The two groups usually collapsed into one are worth separating. About 2.2 billion people are offline entirely, so AI is not a choice they are declining. Another 4.8 billion are online and have not used it, which is a very different fact and a much larger number than most adoption conversations assume. Meanwhile the entire debate about frontier capability is happening among the last 27 dots on the grid.
Every large enterprise, drawn to scale
Same treatment, different population: 2,000 dots, one dot per large enterprise, colored by how far AI has actually travelled from pilot to production.
A quarter of large enterprises have genuinely moved AI into production. Slightly more than half expect to be there within two quarters. The distance between those two bands is the entire enterprise planning problem: budgets, roadmaps and vendor commitments are routinely built on the expectation rather than on the achievement.
A separate survey points the same way without being folded into the same chart. McKinsey's November 2025 State of AI found 62 percent of organizations experimenting with AI agents and fewer than one in ten scaling them in any single function, with 23 percent scaling an agentic system anywhere in the enterprise. Two different instruments, two different populations, one consistent direction: the experimenting is broad and the scaling is not.
Where every number comes from
These are the sources behind the two charts, with how often each one actually publishes. The cadence matters as much as the figure: a chart refreshed faster than its sources is theatre.
Facts and Figures. People online and offline worldwide. The hollow dots come from here.
AI Diffusion Report. How many people have used AI, by country. The dataset is public and MIT licensed, which is why this is the anchor.
Weekly active users and paying consumer subscribers, from the company's own announcements.
Developer Survey, with SlashData for the developer population. The share of developers using AI tools, converted to a headcount.
The State of AI in the Enterprise. The pilot-to-production split behind the second chart. It absorbed the quarterly Generative AI series it grew out of, so it publishes annually now.
Business Trends and Outlook Survey. The share of US businesses using AI to produce goods or services, and the only genuinely high-frequency official source in this set.
World Population Prospects and the International Data Base. The denominator.
Four rules that keep these honest
Monthly I check only what actually moves monthly: vendor announcements and the Census survey. Quarterly that now leaves the Microsoft release on its own, because Deloitte absorbed its own quarterly series into an annual report, and I found that by running the check rather than assuming the cadence held. The rest is annual. Re-dating a chart whose underlying data has not moved would look like currency and be the opposite.
They are separate survey questions with overlapping populations. Stacking three of them into one ladder produces a chart that looks authoritative and is arithmetically meaningless. The second chart uses a single question family whose answers are mutually exclusive and sum to 100, which ruled out several more dramatic framings.
Forty million people is about twelve dots at this scale. Where rounding flatters a figure, the legend states the real number next to it so the rounding cannot overstate the picture.
A widely repeated claim says 45 percent of Fortune 500 companies run AI agents in production, usually credited to McKinsey. McKinsey's own numbers run the other way: 62 percent of organizations are experimenting with agents and fewer than one in ten is scaling them in any single function. The Fortune 500 framing traces back to a vendor's platform-adoption statistic, which counts something else entirely. It would have been the best number on this page, and it is not on this page.
These figures are not interchangeable, and mixing them is the most common way this subject gets misreported. The Census Bureau counts a business as using AI only when AI helps produce goods or services, which puts US adoption near 17 percent; vendor surveys asking whether anyone uses AI tools at all report figures several times higher. Neither is wrong. They are answers to different questions.
The same trap applies on the people side. "Has ever used AI" is not "uses AI weekly," and a single vendor's user count is not a market total. Where a tier could only be sourced from one company, the chart says so rather than implying it covers everyone.
What would change the picture
The connectivity figure moves slowly and predictably, so the interesting volatility is all at the bottom of both charts. A step change in low-cost access across the Global South would move the largest band on the first chart, which nothing else plausibly can. On the enterprise side, the honest test is whether that 54 percent expecting production within two quarters actually arrives; if the next Deloitte edition shows the achieved share flat while the expectation stays high, the gap stops being a transition and starts being a pattern worth naming.
Both charts are rebuilt from a single renderer, so the version on a slide and the version on this page cannot drift apart. The method behind the sourcing, and the system that runs it, are on the other practice notes: the operating system behind this research and why polish is not verification.
Charts drawn by Travis Havens from the published sources named in each footer. Figures are reproduced as data with attribution; no source's own artwork is used. Percentages are rounded for display and the underlying values are stated in each legend.
Working out where your organization actually sits on the second chart? No pitch. A short conversation about what matters first for your situation.
Find me on LinkedIn