Is AI adoption slowing? There has been a growing unease in recent weeks that perhaps something has changed, whether it's the it-isn't-AGI-yet meh reaction to GPT-5; or a sense that large language models don't scale as well as once thought; or perhaps Nvidia results that were really good, but not really-really good. A recent graph from Apollo has added to this, showing that large companies are reversing their recent rate of adoption for AI.
Here is the graph:
You should, as a discerning data consumer, have a few initial questions:
- Where did this survey data come from?
- What survey question(s) were asked?
- Given the data source and question, is the result credible?
Where did the data come from?
The data came from the U.S. Census Bureau's Business Trends and Outlook Survey (BTOS), which it conducts every two weeks. It is a panel study with 1.2 million companies as rotating respondents, 200,000 of whom are asked to respond in any given two-week period.
What question(s) were they asked?
The survey question was this: “In the last two weeks, did this business use Artificial Intelligence (AI) to produce goods or services?” Helpfully, the Census Bureau gave examples of what it means by "AI". It offers, parenthetically, "Examples of AI: machine learning, natural language processing, virtual agents, voice recognition, etc.)”.
Is the survey result credible?
This is where things get complicated. The Census Bureau is good at surveys, given that collecting lots of data is what it does, after all. The trouble is partly analytical and partly the question itself.
The question is, in a word, strange. It describes AI in terms that wouldn't have been out of place pre-ChatGPT, which, you may have heard, is the fastest-growing technology product in history. There is virtually zero chance that ChatGPT usage at most large companies is this low, let alone declining, so whatever respondents at large companies are saying is declining, it is not that.
And this leads us into the methodological weeds, but I'll keep it simple. While the survey is sent to 1.2 million businesses, it only goes to 200,000 at a time in any two-week period. But the fraction of U.S. companies with more than 250 employees is small, on the order of 0.4-0.5%. That means that in any given period, the 250+ employee group, given typical response rates of 20% or so, might be as few as 100 companies. This, in turn, leads to more variance in the estimate.
We can estimate what the error rate is. The above graph suggests that AI adoption by large companies declined from 14% to 12%. But survey statistics tell us the margin of error on a small sample of large companies is on the order of 5%. You can shrink the error somewhat, as the above graph does, by doing a rolling average across the last few panels. That, however, only gets you to about a 2.8% error. In other words, the recent "decline" in large company AI adoption—even if you ignore how bizarre the BTOS question is—is well within the margin of error.
What BTOS Really Says is Going on With AI Adoption
So, while large companies' AI adoption may be declining, as the above graph shows, it is less likely than it might seem. First, the question captures a strange sense of what AI is, and second, the error on such a small sample of large companies doesn't let us say much about what large companies are doing with respect to using AI.
We can, however, get around this by slicing the BTOS data differently. Instead of breaking it down by company size, it can be broken down by sector. And that, which I have done, tells a different story.
Most sectors are seeing sharply increasing AI adoption, with the unsurprising exceptions of construction and manufacturing. But those two sectors make up a large share of the 250+ employee cohort, as much as 40%. So the original results by employee count are likely skewed badly by companies with the lowest AI adoption rates of any form, further helping explain the "AI decline" graph, in addition to measurement error.
Simpson's Paradox
This is a kind of example of Simpson's Paradox, where multiple subgroups within a larger population might show an improving trend, while a line of best fit through all the groups shows a declining one.
In the following figure (adapted from Wikipedia's entry), the green line of best fit through all the data shows a decline, but all the subgroups—you might think of them as AI adoption by various sectors—show increasing adoption.
Conclusion
Does a Simpson-ian paradox explain what is going on in AI adoption data? It is possible, and it does fit the industry adoption vs size story well. Either way, it is almost certain that AI adoption, far from declining, is advancing sharply in companies of all sizes—outside, perhaps, some deadbeat companies in construction and manufacturing, which we knew already.