AI Adoption Index: How Fast Are American Businesses Actually Using AI?

The question the Adoption Index is built to answer

There is a wide gap between how often AI is mentioned in earnings calls and how often it is actually running inside US firms. The AI Adoption Index is designed to close that gap. It combines three primary measurement series, the US Census Bureau Business Trends and Outlook Survey (BTOS), the Stanford AI Index annual enterprise chapter, and the McKinsey State of AI panel, into a single weighted score that can be compared across sectors, firm sizes, and quarters.

The reason a composite is needed is that no single series is sufficient on its own. BTOS is the most representative source for the United States because it samples roughly 1.2 million businesses on a biweekly cadence. It is also the only series with full firm-size and sector coverage. But BTOS asks a deliberately simple yes or no question about whether the firm used AI to produce goods or services in the last two weeks, which understates both depth and the difference between pilot use and production use. The Stanford AI Index aggregates a wider set of vendor and survey data, including job-posting signals and capital spending, but its sector coverage is uneven and its update cycle is annual. The McKinsey State of AI panel is the deepest on use-case detail and on differences between high-performing and low-performing adopters, but its sample is self-selected and skews toward larger enterprises.

Reading any one of these in isolation produces a misleading picture. The Adoption Index averages them, with weights documented below, so a user can see both the headline rate and the disagreements between sources, which is itself informative.

What BTOS shows about the headline rate

The cleanest single number for US AI adoption is the BTOS share of firms reporting they used AI to produce goods or services in the prior two weeks. That figure rose from roughly 3.7 percent in late 2023 to a little under 7 percent by mid-2025, with the most recent waves indicating continued growth concentrated in larger firms and in a small number of sectors (Census BTOS AI module). The Federal Reserve Bank of St Louis maintains a useful time series visualization that tracks the BTOS AI question against employment, output, and capital expenditure.

The sectoral pattern is sharp. Information, professional and technical services, and finance and insurance lead by a wide margin. Accommodation and food services, retail, and construction sit near the bottom. The gap is roughly four to five times. That is the single most important fact about US AI adoption right now: the headline rate of about seven percent hides a top sector running at 25 percent and a bottom sector running at five percent.

Firm size matters at least as much as sector. BTOS shows adoption rising sharply with employment count. The largest firms, those with 250 or more employees, are roughly three times as likely to be running AI in production as firms with fewer than five employees. That gradient is consistent with the long-running stylized fact that productive technology diffuses from large frontier firms to smaller firms over a multi-year horizon, a pattern the OECD documented at length in its 2023 Productivity Outlook.

What the Stanford AI Index adds

The Stanford AI Index, produced annually by the Stanford Institute for Human Centered AI, layers on three series that BTOS does not capture cleanly. The first is private investment in AI companies, which is a leading indicator of the productive capacity available to deploy. The second is the volume of AI-related job postings on aggregators such as Lightcast, which captures hiring intent before it shows up as adoption. The third is patent and publication output, which captures the technical pipeline behind future deployments.

In the most recent edition, private investment in generative AI alone exceeded 25 billion US dollars for a single year, and AI-related job postings remained at multiples of the 2020 baseline despite the broader tech-labor slowdown (Stanford AI Index 2024). These indicators are not direct adoption measures, but they form the supply side that determines how fast the BTOS rate can rise.

The Adoption Index uses the Stanford series as a forward-looking adjustment. When private investment and job postings move together with BTOS, the headline rate is treated as reliable. When they diverge, for example when job postings collapse but BTOS still rises, the Index flags the period as one where the simple headline rate is likely to mean-revert.

What McKinsey adds: depth, not breadth

McKinsey's annual State of AI survey is the most often cited business-side source. It is also the most often misinterpreted. The McKinsey sample is self-selected and global, and the headline number ranging from roughly half to two-thirds of firms reporting regular AI use is much higher than the BTOS US-only measure for a reason. McKinsey counts any reported use, including pilots and one-off experiments, while BTOS asks specifically about producing goods or services.

What McKinsey adds, that no other source matches, is depth on use cases and on the difference between high-performing and low-performing adopters. McKinsey 2024 reported that high-performing adopters are far more likely to have established roles for AI governance, to use a dedicated MLOps platform, and to track AI-specific KPIs. They are also far more likely to source data centrally rather than letting each business unit assemble its own pipeline (McKinsey State of AI 2024).

The Adoption Index uses McKinsey as a depth multiplier. A sector where BTOS shows 12 percent adoption but where the McKinsey panel shows that high-performing adopters in that sector have moved past pilots into production gets scored higher on the depth dimension than a sector where the BTOS rate is the same but McKinsey indicates use remains experimental.

How the composite is built

The Adoption Index is the weighted average of three normalized sub-scores.

• Breadth (BTOS): percentile of the sector's adoption rate within the most recent four BTOS waves, weighted 0.5.
• Supply pressure (Stanford): percentile of the sector's private-investment plus job-posting share against the cross-sector median, weighted 0.2.
• Depth (McKinsey): percentile of the sector's reported use-case maturity score, weighted 0.3.

The choice of 0.5 weight on BTOS reflects that breadth in production is the most policy-relevant signal. The 0.3 weight on McKinsey depth captures that adoption rates without depth are misleading, because a 20 percent breadth rate with no production use is less consequential than a 10 percent breadth rate with mature use. The 0.2 weight on Stanford is the smallest because supply pressure is leading rather than current.

Every weight is published. Users who care more about leading indicators can re-weight toward Stanford. Users who care only about current production can collapse to BTOS.

What the Index says today

Three structural findings come through any reasonable weighting.

Information, professional and technical services, and finance and insurance form a leading cluster. They are above the BTOS sector median on breadth, dominate Stanford private investment, and show the deepest use-case maturity in McKinsey. The Index puts them at 70 to 85 on the 0 to 100 scale.

Health care, manufacturing, and wholesale trade form a middle tier. BTOS adoption is in the mid single digits, but supply-pressure indicators are strong, and McKinsey depth signals indicate selected production use in specific functions, customer service in health care, predictive maintenance in manufacturing, demand forecasting in wholesale. These sectors score 40 to 60.

Construction, accommodation and food services, retail, agriculture, and personal services form a low tier. BTOS adoption is the lowest, supply pressure is weak, and McKinsey depth signals are negligible. These score below 25. This is consistent with the long-running observation in the Acemoglu macroeconomic AI estimates that productive AI deployment is concentrated in sectors with high baseline information intensity.

Why state-level adoption looks different

A US Adoption Index for businesses can be re-cut at the state level by joining the BTOS sector breakdown to the state-level industry mix from the Bureau of Labor Statistics Quarterly Census of Employment and Wages. States with large information, finance, and professional-services sectors, California, Washington, New York, Massachusetts, Virginia, and the Texas triangle, score highest on a sector-weighted Adoption Index even before any state-specific innovation policy is added. States dominated by hospitality, agriculture, and construction score lower for structural reasons unrelated to ambition.

The state question is structurally similar to the one covered in our companion piece on AI readiness by state, which looks at infrastructure, talent, and policy. Readiness is what the state can support. Adoption is what is happening today.

Caveats and reading guidance

Three caveats matter for anyone using the Index in a policy or investment context.

First, BTOS is fast but shallow. A respondent reporting "yes" on the AI question may be using a single off-the-shelf generative tool by a marketing team. That counts in the breadth measure but is not equivalent to integrating AI into a production process. The Index addresses this through the McKinsey depth weight, but the underlying ambiguity remains.

Second, the McKinsey sample is global and self-selected. We use McKinsey only as a relative ranking of depth across sectors, not as an absolute level. Treating the McKinsey headline percentage as the US adoption rate is the most common misreading in the press.

Third, the Stanford series is annual. Quarterly users will see the supply-pressure sub-score move slowly. When private-AI funding cycles shift sharply, the Adoption Index will lag those shifts by up to nine months.

For broader context on what these adoption patterns mean for workers, see our companion pieces on the AI Workforce Risk Index and AI job displacement by county. For the structural side, see State AI Readiness Rankings. For the underlying technology, the foundations explainer on what AI actually is and Large Language Models, Explained Without the Hype give the necessary background.

Where to go next

Researchers should download the BTOS public-use tables directly and join them to QCEW state-level industry mix to build their own state Adoption Index. The Federal Reserve Bank of Richmond has published a useful guide on linking BTOS to FRED time series for productivity work. The OECD AI Policy Observatory maintains a cross-country adoption tracker for those who want comparative context outside the United States.

The honest reading of the Adoption Index is that the United States is in the early-middle phase of an enterprise-AI diffusion cycle. The headline rate is rising steadily, the sectoral pattern is sharp, and the gap between leading and lagging sectors is widening rather than narrowing. Policy that wants to broaden adoption beyond the leading cluster needs to address the binding constraints in mid- and low-tier sectors: data integration capacity in health care, workforce skills in manufacturing, and capital in small retail and hospitality. The Index is the diagnostic. The interventions are sector-specific.

Sources

• US Census Bureau, *Business Trends and Outlook Survey AI Supplement*.
• Stanford Institute for Human Centered AI, *2024 AI Index Report*.
• McKinsey & Company, *The State of AI in 2024*.
• OECD, *Productivity Statistics* and AI Policy Observatory.
• Acemoglu, *The Simple Macroeconomics of AI*, NBER 2024.
• Bureau of Labor Statistics, *Quarterly Census of Employment and Wages*.
• Federal Reserve Bank of St Louis, *FRED Blog*.