AI Adoption Statistics 2025–2026: Two Census Measures, Two Different Questions

A slide that says “AI adoption rose from 5.4% to 18%” looks useful for a budget meeting. It is also a comparison the U.S. Census Bureau says you should not make. The 5.4% estimate asked whether a firm used artificial intelligence (AI) to produce goods or services in the previous two weeks. The later 18% estimate asked whether a firm used AI in any business function. The question changed, and so did the measured series. This October 1, 2026 revision replaces this page’s older, unverified cross-industry rankings and return claims with the source documents a leader can inspect.

What the older figure measured

A 2024 Census working paper analyzed the Business Trends and Outlook Survey from September 2023 through February 2024. It reported that 5.4% of U.S. firms used AI in producing goods or services during the prior two weeks by February 2024, up from 3.7% in September 2023 under that question. That is a narrow, dated measure of firm use. It does not tell you that 94.6% of firms had never used an AI assistant for writing, search, or internal analysis.

The old version of this article presented precise rates for several industries and specific return ranges without primary sources beside those claims. Those numbers are removed here. A mix of vendor cases, executive polls, spending forecasts, and Census responses cannot be turned into one industry league table merely because each contains a percentage.

What the newer figure measured

The April 2026 Census working paper reports that 18% of U.S. firms used AI in a business function during its November 2025–January 2026 reference window. When firms are weighted by employment, the figure is 32%. That second number gives larger employers more influence; it does not mean 32% of individual workers personally used AI. The paper also reports much higher use among very large firms in information, professional services, and finance. A small company should not adopt a large-company rate as its own target.

The paper describes use across firms, business functions, and worker tasks, and reports that AI-related employment decreases occurred in 2% of firms. That is the share reporting a decrease in this survey, not a forecast of layoffs at your company. The document is a Census working paper, which the Bureau says has not undergone the review of an official Census publication. Its value is a nationally representative, dated measurement with visible questions and methods—not a prediction of return on a particular software purchase.

Why the two figures cannot make one growth line

On November 17, 2025, Census changed its core question from AI use “in producing goods or services” to AI use “in any of its business functions.” Census’s wording note says the change produced a level shift and that the Bureau created a new time series beginning with data released on December 4, 2025. A chart joining 5.4% and 18% as if only time had changed would assign the wording effect to adoption growth.

You can compare like questions within their own series. You can also discuss why a broader business-function question finds uses the production question missed. What you cannot calculate from these two headline figures is a clean growth rate. A vendor may cite a larger executive survey number; ask who answered, what “use” meant, and whether the result counted firms, employees, or tasks before placing it beside Census.

Use national adoption as context, then measure your job

The national figures are useful for describing where business use exists and for challenging the claim that every firm has already automated its work. They are weak evidence for approving your own 500-seat contract. For that decision, the operations owner should choose one recurring job, count eligible volume, measure current time and error, name the accepted output, and run a limited comparison. Finance should count review and implementation cost and keep released capacity separate from any actual budget reduction.

• When quoting an adoption statistic, write down the source, geography, survey dates, people or firms counted, exact question, and weighting method.
• When a vendor result appears beside it, label that result as a provider claim or customer report and check whether it measures the same population and outcome.
• When deciding to buy, use a local baseline and a stop rule. A national use rate is not a return estimate.

If the deck still draws a line from 5.4% to 18%, correct the slide before the budget vote. Then bring the workflow evidence to the meeting. The current State of AI Adoption report keeps the two Census papers separate and pairs them with the tool and price data this site tracks.

Useful next steps

• Open the current State of AI Adoption report — See the two Census measures beside the site’s tool catalog.
• Read the 2026 Census paper in context — Understand why the newer question changed the reading.
• Build a measured AI business case — Base spending on your own workflow, not a national average.