Summary
- Stanford's 2025 AI Index prints 78% as the share of surveyed respondents who say their organisation uses AI in at least one function, up from 55% in 2023, and sources Figure 4.4.1 to "McKinsey & Company Survey, 2024".
- The chapter describes the 2024 McKinsey analysis as "based on two surveys spanning 2,854 respondents", but the wave behind the 78% drew 1,491 participants in July 2024. The 2,854 is my sum of McKinsey's two printed sample sizes, not a split the chapter publishes.
- The question, the eleven business functions, the fieldwork dates and McKinsey's own note that "adopted" was left undefined appear in McKinsey's publication. In the two chapter figure pages and three index pages I read on 19 November 2026, they do not appear beside the figure.
The chart, in its own words
Figure 4.4.1 in the 2025 edition carries this label: "Share of respondents who say their organization uses AI in at least one function, 2017-24." Under it sits this line: "Source: McKinsey & Company Survey, 2024 | Chart: 2025 AI Index report." The annotations read 78% for AI and 71% for generative AI (2025 AI Index, Chapter 4: Economy).
The body sentence is more careful than the chart: "The latest McKinsey report reveals that 78% of surveyed respondents say their organizations have begun to use AI in at least one business function, marking a significant increase from 55% in 2023 (Figure 4.4.1)."
Three things in one sentence, and each one carries weight. The unit is surveyed respondents. The threshold is at least one business function. The measurement belongs to McKinsey; what the index adds is the verb "reveals" and the arrow from 55 to 78.
The base printed beside the value is not the value's base
In the same section, the chapter says the 2024 McKinsey analysis "is based on two surveys spanning 2,854 respondents across various regions, industries, company sizes, functional areas, and tenures" (Chapter 4). That is a real number about a real sample. It is not the denominator of the 78%.
McKinsey's July 2024 survey — the one that produced the 78% — was in the field from 16 to 31 July 2024 and drew 1,491 participants across 101 nations (The state of AI, March 2025, web edition). Its early-2024 predecessor ran from 22 February to 5 March 2024 with 1,363 participants (The state of AI in early 2024).
1,363 + 1,491 = 2,854. That sum is mine. The chapter prints the total, never the split, and never says which of the two waves the 78% comes from. The base printed next to the value is roughly 1.9 times the base of the survey the value comes from (2,854 ÷ 1,491 = 1.91, also mine).
None of this makes the 78% wrong. Respondents were surveyed, 78% of them answered a particular way, and the figure reports that answer faithfully. What it makes is a reading problem: the printed base belongs to the study, and the printed percentage belongs to one wave of it.
Two editions, two verbs, two units
Turn one year back. The 2024 edition's Figure 4.4.1 read "Share of respondents who say their organizations have adopted AI in at least one function, 2017-23", sourced "McKinsey & Company Survey, 2023 | Chart: 2024 AI Index report" (2024 AI Index, Chapter 4: Economy). The verb was "have adopted". A year later it is "uses", and the source year has advanced from 2023 to 2024.
The text moved too, and not only in verb. The 2024 chapter's sentence reads: "55% of organizations surveyed have implemented AI in at least one business unit or function, marking a slight increase from 50% in 2022 and a significant jump from 20% in 2017." There the unit is organisations surveyed. In 2025 it is respondents. Those are different denominators: one counts firms, the other counts the people who answered about firms, and a survey that returns several responses per firm converts one into the other only through a weighting rule that neither chapter states.
This is the drift two careful publishers produce independently and no reader can reconcile, because neither document derives its unit from the other.
Where the question lives
McKinsey, not the index, printed the instrument: "In which business functions has your organization adopted AI (for example, machine learning, computer vision, natural-language processing)?" Eleven business functions were offered as answer choices. The exhibit carrying the 78% is titled "Organizations that use AI in at least 1 business function, % of respondents" (March 2025).
And McKinsey states that "adopted" was left undefined, with reported use able to span early experimentation by a few employees through to AI embedded across multiple business units (paper edition).
There is the whole measurement, in the measurer's own words. A company whose only AI use is a pilot run by three people in one department sits inside that 78%. A company running models in production across a dozen functions sits inside the same 78%. The percentage is not misleading about this; "in at least one function" does exactly what a label should do. It is the reason a reader who meets the number without the question will assume something narrower.
The middle point the index drops
McKinsey's own sentence: "In the latest survey, 78 percent of respondents say their organizations use AI in at least one business function, up from 72 percent in early 2024 and 55 percent a year earlier" (web edition).
Three points, not two. The index prints the endpoints. Between them the movement runs 55 to 72 — seventeen points — then 72 to 78, six points. The 23-point jump the index reports is real, and it contains a seventeen-point step that happened upstream between a survey fielded 11-21 April 2023 with 1,684 participants (The state of AI in 2023) and the one fielded in February and March 2024 (early 2024).
The fieldwork dates matter for a plain reason: the two points the index compares were gathered roughly fifteen months apart. "55% in 2023" and "78% in 2024" are not two calendar-year observations. They are two surveys, and the calendar labels on them are conventions.
What the landing page does to it
The 2025 report's landing page renders the finding in one line: "AI business usage is also accelerating: 78% of organizations reported using AI in 2024, up from 55% the year before" (The 2025 AI Index Report).
No survey organisation. No sample. No question. And the unit has quietly changed back from respondents to organisations — the move the chapter's own label avoids. The sentence most readers meet first is the one with the fewest handles on it.
Stanford University, whose Institute for Human-Centered AI publishes the index, has its own entry in the BTW directory; this piece is about its Economy chapter, not its network.
What a reader can recompute
From index-printed values alone: 78 − 55 = 23 points. That is the whole list. The chart supplies a share, a year and a source name, and nothing that behaves as a base.
With McKinsey's documents open, three sums become available. 1,363 + 1,491 = 2,854 reproduces the chapter's combined base. 2,854 ÷ 1,491 ≈ 1.91 measures how far the printed base sits from the value's own. The 55 → 72 → 78 sequence restores the upstream step the index compresses. All three are arithmetic, labelled mine, performed on other people's printed numbers.
What no document in front of me allows is rebuilding the 78% itself — taking a count of respondents and a count of answers and dividing. That number is not a derivation. It is a response share, reported by the organisation that asked the question.
What a compiler owes
A compilation is not a scandal because it cites. The AI Index does the honest thing at figure level: it names McKinsey in the source line and in the body, and it never claims to have fielded the survey. That is more than plenty of charts in this genre manage.
But a compiled percentage carries four things, or it carries only a citation:
The question — because "uses AI", "has adopted AI" and "has implemented AI" are three thresholds wearing one noun. The option set — because eleven business functions with "adopted" undefined is a different instrument from a binary question about production use. The base of the value, not a base adjacent to it — because a study-level total cannot answer a wave-level question. And the fieldwork dates — because a fifteen-month gap and a two-year gap are not the same claim about acceleration.
Print those four and a reader can audit the number without fielding a survey. Omit them and the reader keeps a citation, a percentage, and the reasonable but wrong impression that the adjacent base is the value's own.
The limit on all of this
Eight documents read on 19 November 2026: the 2025 and 2024 Economy chapters, three AI Index pages, three McKinsey publications. Within the 2025 chapter's section 4.4 and the 2024 edition's Figure 4.4.1, the question wording, the eleven functions, the fieldwork dates, the per-survey sample sizes and McKinsey's note on the undefined word do not appear. That is a statement about those passages, not about the whole report.
The full 414-page 2025 report would not open through my route — it returned a response too large for the retrieval adapter. The question and the appendix may well sit somewhere in a document I could not read, and a summary chapter is under no obligation to reproduce every methodological footnote. I did not contact the index's authors either. Grant all of that, and what remains is the reason this was worth writing: the number, its label, its source line and its printed base sit within a few lines of each other, and the base is not the number's own.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






