U.S. businesses’ predictions of their near-term AI use became more accurate by late 2025, but the reasons early adopters gave for using AI did not consistently correspond with later industry-level outcomes, according to a research spotlight by the U.S. Bureau of Economic Analysis (BEA).
Tracking expected versus actual AI use
The July working paper, “AI Expectations and Outcomes”, compares Census Bureau Business Trends and Outlook Survey estimates of expected AI use with actual use six months later.
In the third quarter of 2023, businesses expected the share using AI to grow over a quarter over six months, but the share rose 14%; by the fourth quarter of 2025, over 20% expected to use AI within six months and just over 21% reported doing so in the second quarter of 2026.
AI adoption initially ran below expectations, followed by a period when it grew faster than expected, before expected and actual use aligned more closely by late 2025. The authors said the results are consistent with businesses learning to assess their AI use cases, although a longer time series is needed to confirm the trend.
A shift in Census survey methodology
The comparison also spans a Census survey change. In late 2025, the AI question broadened from use “in producing goods or services” to use in “any of its business functions,” almost doubling reported usage between surveys and creating a break that prevents direct comparison across the wording change. The Census Bureau’s current BTOS data also flags the November 2025 wording revision.
Connecting adoption motives to industry outcomes
For outcomes, the researchers used the Census Bureau’s 2019 Annual Business Survey, which asked adopters about their motivations for using AI during 2016-2018, then compared those motivations with later industry-level production data. The most common reasons were improving process quality or reliability, upgrading outdated processes, automating labor tasks and expanding the range of goods or services.
The strongest association appeared where expanding goods or services was a relatively strong motivation. The regression associated that motivation with relatively faster growth in total factor productivity and labor productivity, while the other motivations were less clearly associated with the outcomes the researchers could measure.
A stated motivation to automate labor tasks, for example, was not associated with a statistically significant relative decline in labor’s contribution to output growth, although it was associated with higher capital use and lower intermediate-service inputs. Industries focused on upgrading processes or improving quality also showed some statistically significant changes, but the authors described the links between those motivations and measured outcomes as unclear.
Data limitations and takeaways for enterprise planning
The authors cautioned that limited detail in the available data means some results may not be robust and are sometimes inconclusive.
For enterprise planning, the paper separates two questions: whether businesses’ aggregate expectations about their own near-term AI use match later reported use and whether stated adoption motives correspond with later industry-level outcomes.
Aggregate expected and actual use had become more closely aligned nationally by late 2025, but the second relationship remained difficult to establish. Fourth-quarter 2025 expectations in the professional, scientific and technical services sector and the finance and insurance sector still underestimated AI-use growth reported by the second quarter of 2026.