Which Industries Are Adopting AI Fastest? (US Census Data, 17 Sectors)

Construction is adopting AI faster than any other US industry. It nearly doubled in nine months, from 7.4% to 14.9% of businesses — the clearest signal in the US Census Bureau's data, at 6.7 standard errors. Meanwhile Information, still the highest-adoption sector at 45.0%, has stopped growing measurably. Eleven of seventeen sectors grew over the period. None declined.

Nearly every AI adoption statistic in circulation comes from a vendor survey of a few hundred executives who agreed to answer a vendor survey. The sample is self-selected, the sponsor sells AI, and the number is almost always large.

The US Census Bureau runs a different instrument. The Business Trends and Outlook Survey samples roughly 1.2 million businesses, splits them across six panels, collects every fortnight, and asks a plain question: "in the last two weeks, did this business use Artificial Intelligence in any of its business functions?" It publishes standard errors alongside every estimate.

It is barely mined, because it arrives as multi-tab spreadsheets with no documentation of which tab holds what. We pulled the sector series and tested every movement against its own error bars. The full dataset is available as CSV.

Eleven sectors grew. None declined.

Between the cycle collected 17–30 November 2025 and the one collected 27 July – 9 August 2026, eleven of seventeen sectors moved by more than 1.96 standard errors. Six did not move detectably. Nothing went backwards.

The interesting number is the one that did not move

Information is the highest-adoption sector in the survey at 45.0%. Its nine-month change of +4.0pp does not clear its own standard errors. Educational services, second-highest of the conventional sectors at 36.7%, is in the same position.

Meanwhile construction went from 7.4% to 14.9%. Manufacturing added 6.6 points. Wholesale trade added 5.3. Transportation added 3.3 from a very low base.

The shape of that is a diffusion curve doing exactly what diffusion curves do. The sectors that adopted first are approaching their ceiling; the growth has moved to the sectors that had not started. Anyone still describing AI adoption as a technology-sector story is describing 2023.

This matters commercially in a specific way. If you sell into information or professional services, you are selling into a market that is largely already using something, and the conversation is replacement. If you sell into construction, manufacturing or wholesale, the market is still in first-purchase, and it is moving quickly.

What the standard errors delete

Read the raw table without error bars and there are two obvious headlines: utilities and accommodation both went backwards. Both are noise.

Utilities' change of −0.5pp carries a standard error of about ±6.3pp. The sector has few enough businesses that the survey cannot resolve a movement of that size. Accommodation's −0.6pp is likewise well inside its error. Reporting either as a decline would be inventing a finding.

We treated the two cycles as independent when combining their errors. Because the panels overlap, the true error is somewhat smaller than that, so the test is conservative: it makes significance harder to claim, not easier.

Six weeks when nobody was counting

The published AI series has a hole in it. Cycles 202521, 202522 and 202523 — covering 6 October to 16 November 2025 — do not exist, because of the federal funding lapse. Collection stopped.

It is worth stating plainly because it is the sort of gap that gets quietly interpolated over. During a period of rapid change in business AI use, the United States has no measurement of it for six weeks. Our window starts immediately after that gap for exactly this reason.

Method

• Source: US Census Bureau Business Trends and Outlook Survey, `Sector.xlsx` and `National.xlsx` from the BTOS download directory, retrieved 22 August 2026.
• Question 7: "In the last two weeks, did this business use Artificial Intelligence (AI) in any of its business functions?" Percentages are the share answering Yes.
• Window: cycle 202524 (collected 17–30 November 2025) to cycle 202616 (collected 27 July – 9 August 2026).
• Significance: change divided by the combined standard error of the two estimates, treating them as independent, tested against 1.96.
• Suppressed cells are carried through as missing rather than as zero.
• "Multi-sector companies" is the Census category for firms operating across several sectors — not an all-sectors total.

Limitations

This is a nine-month window, not a year. The AI question is only populated in the published series from cycle 202524 onward, so a twelve-month comparison is not available from this file, and we would rather report nine months honestly than stitch together sources measuring different things.

"Using AI" is self-reported and undefined in scope. A business answering yes may be running a large deployment or may have a member of staff using a chatbot. The question captures breadth of adoption, not depth of it, and the trend is more reliable than the level.

This is the United States only. The UK's counterpart, the ONS Business Insights and Conditions Survey, asks about AI less consistently, so a like-for-like transatlantic cut is not available from public data at this granularity.

The line worth taking to a board

Most AI business cases are argued from vendor projections. This one can be argued from a government instrument with a 1.2-million-business sample: your competitors in construction went from 7.4% to 14.9% in nine months. In manufacturing, 12.7% to 19.3%. In wholesale, 12.6% to 17.9%.

Those are the sectors where adoption is still a competitive differentiator rather than table stakes, and where the window in which it is one is visibly closing. The useful question for a business in one of them is not whether to adopt but which process to start with — which is the first thing worth settling in any AI automation engagement, before anything gets built. For construction, the sector moving fastest here, we have written that out as five workflows worth automating and the ones that are not.