A claim circulates in vendor decks and trade press that smaller companies are adopting artificial intelligence faster than large ones, because they are nimble and unencumbered. The official statistics say the opposite. The US Census Bureau's Business Trends and Outlook Survey, after it broadened its question in November 2025 to cover AI use in any business function rather than only in producing goods and services, found that about eighteen percent of firms reported using AI. A Census Bureau working paper on the same survey, by Kathryn Bonney and coauthors, reports that the figure rises to thirty-two percent when each firm is weighted by its employment. A higher employment-weighted rate than firm-weighted rate is the statistical signature of large firms adopting more, not less. The reading of the data that does not flatter the mid-market is the more useful one, because it points to where the real advantage lies, which is not the rate at which a firm starts but the speed at which it can finish.
What the survey actually measures
The BTOS is a large, biweekly panel that the Census Bureau fields across the universe of US employer firms, which is what gives its AI numbers more weight than the self-selected vendor surveys that produce the higher headline figures. The distinction between the two rates it reports matters. The firm-weighted rate counts every business equally, so a sole proprietor and a thousand-person manufacturer each count once, and that rate sits near eighteen percent. The employment-weighted rate counts a firm in proportion to how many people it employs, so it describes the share of the workforce that sits inside an AI-using business, and that rate sits near thirty-two percent. If small and large firms adopted at equal rates the two numbers would converge. They diverge by roughly fourteen points, which means the larger an employer is, the more likely it is to report using AI.
The gap by size band
Reading the rate by firm-size class makes the pattern concrete. The BTOS breakdowns place firms with two hundred fifty or more employees near thirty-seven percent, mid-market firms of one hundred to two hundred forty-nine employees near thirty-two percent, and firms with four or fewer employees under twenty percent, which puts the smallest firms close to the national average rather than far beneath it. Adoption climbs with headcount through the upper bands, and the one bend in the pattern sits at the very bottom. The Federal Reserve, in an April 2026 FEDS Note by Jeffrey S. Allen, compared AI-adoption signals across multiple surveys and found the same shape, the largest firms with the highest adoption rates and the smallest adopting more strongly than size alone would predict, which is the cross-survey corroboration that turns a single data point into a finding.
The trajectory does carry one piece of comfort for smaller firms. The US Small Business Administration's Office of Advocacy, in a September 2025 report titled "AI in Business: Small Firms Closing In," documented that the gap between large and small adopters has narrowed over recent quarters, and the Census Bureau's own earlier Research Matters analysis, "Is AI Use Increasing Among Small Businesses?", tracked the same upward trajectory by firm-size class. So the smallest firms are catching up, from behind rather than ahead, and the mid-market sits in the middle of that distribution, ahead of the smallest firms and behind the largest.
Adoption rate is the wrong metric for the mid-market
The size gradient describes who has started. It says nothing about who can reach a working result quickly, and that is the metric an operator should care about, because a reported adoption is not a production system. The thirty-seven percent of large firms reporting AI use includes a great many pilots that never reach the profit-and-loss statement, a pattern we examine in the state of enterprise AI in 2026. Reporting that you use AI and running a process on it every month are different facts, and the survey captures the first.
Time to production is where firm size cuts the other way. A large enterprise that decides to put AI into its accounts-payable function must route that decision through an approval chain with several layers, integrate across a dozen overlapping systems accumulated through acquisitions, and clear a governance process that may involve a data committee, a security committee, a vendor-risk review, and a model-risk function. Each step is reasonable on its own. Stacked, they push the interval between decision and deployment into quarters. A mid-market manufacturer or distributor running on QuickBooks or a single legacy ERP has a shorter approval chain, often reaching the president directly, fewer systems to integrate, and no multi-committee governance apparatus, because it never grew one. The same structural traits that hold its adoption rate below the largest firms, namely smaller scale and fewer resources, are what let a committed mid-market firm move from decision to a running deployment faster than a large enterprise can convene its first review.
Why the structural edge is durable
This advantage does not depend on the mid-market being more enthusiastic or more technical. It is a property of the organization's shape. The fewer the layers between the person who decides and the system that changes, the shorter the path to a result, which is why we have argued that the mid-market is the customer all four delivery models leave behind and also the one best positioned to act once a delivery model fits it. The recent fall in model cost, which we covered in what changed in the AI cost curve, removed the budget objection that used to keep these firms out, so the binding constraint is no longer money or capability but the time and the path to a deployment that holds. On that constraint the mid-market's structure works in its favor.
A reasonable counter, answered
A reasonable counter is that short approval chains cut both ways, that the absence of a governance process means a mid-market firm has no model-risk review, no formal vendor vetting, and no committee to catch a bad deployment before it ships, so speed buys exposure. There is real force in this. A firm that can deploy in days can also deploy something wrong in days. But the governance apparatus at large firms is not primarily a quality control on AI outcomes. It is a coordination cost imposed by scale and by the number of stakeholders who can block a change, and most of what it produces is delay rather than diligence. The controls that actually protect a deployment, the scoping of what an agent may do and the human review of its exceptions, do not require a committee structure, which is why a focused mid-market firm can install them inside a fast engagement, the same way that services-as-software firms scale where slower delivery models cannot. The mid-market starts behind on adoption. It can close that gap quickly once it commits, and it can do so for the same structural reasons that put it behind in the first place.