Census AI-Adoption Figures Compare Three Different Questions

A September 2025 chart from Apollo Global Management appeared to show large U.S. companies slowing their AI adoption. But the Census survey behind the chart changed its question wording twice. The rise from roughly 12% to roughly 40% is therefore not a clean adoption curve: it spans three different wording regimes. ¹

The source is the Census Bureau’s Business Trends and Outlook Survey (BTOS). Because the question asks whether a business used AI, the resulting response is context about reported use—not, by itself, a measure of spending, deployment depth, or how embedded AI is in a company’s operations. ¹

What the Apollo chart and the Census AI adoption data measured

BTOS is conducted every two weeks and goes to 200,000 businesses out of a total of 1.2 million. The panel cycles so that an individual business completes the survey about every twelve weeks. ¹

Before November 17, 2025, the question asked whether a business had used AI in the prior two weeks to produce goods or services. That wording focused on production rather than AI use across all business functions. ¹

The question does not establish the scale or frequency of use, its spending or business impact, or whether use was experimental or routine. Those are limits of what this question measures, rather than evidence against respondents’ answers. ¹

Why the wording change breaks the trend line

On December 3, 2025, the Census Bureau announced a material change: the question shifted from AI use “in producing goods or services” to use “in any of its business functions.” The source reports that the Bureau said responses before and after November 17 were not directly comparable. ¹

Beginning with the October 19, 2026 collection, the wording changed again. The survey began asking about “Super Intelligence” or “SI,” with examples of possible uses. These are three wording regimes, not three points measured with one unchanged question. ¹

Period	What the question covered	Caveat
Before November 17, 2025	AI use in the prior two weeks to produce goods or services 1	Production-focused wording 1
From November 17, 2025	AI use in any business function 1	Broader scope; the source says it is not directly comparable with the earlier wording 1
From the October 19, 2026 collection	“Super Intelligence” or “SI” use in any business function, with examples including data analysis, marketing, product design, coding, customer service, forecasting, pricing, and quality control 1	New wording; the source raises, but does not resolve, whether it affects responses 1

For example, a business whose AI use was limited to marketing could fall outside the production-focused question but within the any-function question. That illustrates how a broader definition could capture activity the earlier wording did not. The source does not quantify how much of the reported increase reflects newly adopted use versus the expanded scope. ¹

Did enterprise AI adoption actually accelerate?

Among the largest businesses, those with 250 or more employees, about 12% answered yes in September 2025 under the production-focused wording. Under the broader any-function wording, about 30% answered yes by February 2026 and about 40% by October 2026. ¹

These figures show an increase in reported use, but the source does not separate changes in business behavior from the effect of the broader question. The figures across the wording change are not directly comparable. ¹

That explains the apparent discrepancy between the September chart and the later, higher figures: they reflect different question wording, so the later rates do not by themselves establish a reversal or acceleration in the underlying adoption trend. ¹

Analysis: Read the rise as a change in reported responses, not as a precise measure of the change in adoption. Both increased use and a wider definition could contribute; the figures cited here do not distinguish between them. ¹

What “Super Intelligence” changes next

Beginning with the October 19, 2026 collection, BTOS changed its wording in response to Executive Order 14434. The revised question uses “Super Intelligence” or “SI” in place of “Artificial Intelligence” or “AI” and provides examples across several business functions. ¹

The October 2026 figures discussed above still reflect the earlier AI wording, according to the source; they are not results under the new SI wording. ¹

Whether the new label changes how respondents answer remains an open question in the source. It says the first results using the new wording were due for publication on the Thursday after its November 2, 2026 publication date. ¹

Analysis: Treat results under the new wording as a fresh comparison point, rather than assuming they continue an unchanged series. A label change could affect what respondents picture when answering, even if their use has not changed. ¹

Three questions to ask before citing AI adoption statistics

• What counted as “use”? Production-focused and any-function definitions may capture different groups of businesses. ¹
• What was the recall window? The question asked about use in the prior two weeks. ¹
• Did the wording or examples change between the points being compared? In this case, both the scope and the label changed. ¹

Use BTOS as directional context on reported business use, not as a substitute for measures of spending, deployment depth, or return on investment. For those questions, companies can pair survey context with internal measures of which functions use AI, how widely systems are deployed, and whether use is recurring or one-off. ¹

Teams that route model traffic through an enterprise AI gateway can draw some of those internal measures from usage analytics rather than relying on survey responses alone.

The same discipline applies to other headline numbers. For two further examples of reading a figure against what it does and doesn't measure, see Google Says PageBreak Found 500+ XSS Bugs—Here's What the Number Doesn't Tell You and Why LLM Buyers Need Behavioral Diagnostics, Not Just Aggregate Scores.

A shift in the headline rate is worth investigating, but it is not, by itself, evidence for or against a company’s deployment plans.

What to watch

Before comparing new results with earlier figures, check the question text and examples. The change to SI wording means the new responses should not automatically be treated as if nothing changed. ¹

The open question is whether businesses answer differently when asked about “Super Intelligence” than when asked about “AI.” The source says this is unresolved. The first release under the new wording can show what respondents report, but the figures alone will not identify whether any difference reflects changed use, changed interpretation, or both. ¹

The defensible takeaway is narrower than an adoption-acceleration headline: the survey tracks changes in reported responses, while the effect of changes in question wording complicates comparisons over time. ¹