Rewording a Survey Question Lifted the Share of Companies Using AI from 12% to 40%

The figure looks like a corporate AI salesperson's dream: among American companies with 250 or more employees, the share using the technology has reached 40.7%. A year earlier, the figure was around 12%. One could declare a nearly fourfold increase and get a convincing slide for investors.

But these two numbers measure different things.

The US Census Bureau regularly surveys businesses as part of the Business Trends and Outlook Survey. Until November 2025, companies answered whether they used AI in producing goods or delivering services. Then the question was broadened: it now covers the use of AI in any business function. Marketing, programming, customer support, forecasting and internal analysis also began to count.

One Term, Two Different Metrics

Consider an online store. It does not use machine learning to build its assortment, manage its warehouse or handle delivery, so it answers "no" to the old question. But its marketers draft ads with a language model, and its developers use an assistant to write code. To the new question, the same company answers "yes."

The business did not change overnight. What changed is the condition by which it is classified as an AI user.

That is exactly why the earlier 12% and the new 40.7% cannot be placed on one line and interpreted as continuous growth. After the switch to the broader question, the share of large companies was about 30% in February 2026, and by October it had risen to roughly 40%. That segment does allow talk of growth, provided the methodology stayed unchanged within the period.

The author of the original analysis separated the old series from the new one and calculated a moving average over six consecutive survey waves. This approach smooths out fluctuations but does not remove the main break: an average of different questions is still an average of different indicators.

Why This Matters to Product Teams

The same mistake occurs in product analytics. For example, a team changes the "active user" event: previously a user had to create a project, now simply opening the app is enough. After the release, the metric grows by tens of percent. This is not necessarily product success: part of the gain comes from the new definition.

Therefore the wording of the question, the event description, filters and calculation version should be stored next to the metric itself. When a definition changes, it is useful to start a new series, mark the date of the break, and recalculate history only when the source data truly allow a single rule to be applied to the entire period.

There is also a second change. Since October 2026, the Census Bureau, citing an executive order of the US president, replaced Artificial Intelligence with Super Intelligence in the questionnaire, explaining that this is now the name for what used to be called AI. The application examples remained recognizable, but the unfamiliar term could shift respondents' answers: some may decide it refers to a more advanced technology.

Practical Takeaway

An analytics developer should version event definitions and not stitch together incompatible periods. A product manager should check the cause of a jump before presenting results to leadership. A founder should not take a loud adoption figure as proof of a market until the question, the sample and the date of the methodology change have been examined.

A metric without the history of its definition is not a fact but a number that has lost its context. In this case, the data do show the spread of AI in American business, but the honest conclusion is more modest: after the question was broadened, use of the technology in any function of large companies grew from roughly 30% to 40.7% during 2026.