Who Adopts AI—and How Deeply?
The distribution of AI adoption across firms is as important as the technology's capability. The key questions are who adopts, which firms and sectors are exposed, and whether initial use remains narrow or becomes a broad operating capability. Answering them helps distinguish a technology that is merely present in many places from one that is changing the competitive capacity of particular firms.
The early evidence does not describe a level playing field. AI use is spreading, but it is concentrated among larger enterprises and knowledge-intensive industries. At the same time, the firms most exposed to generative AI are not simply the largest incumbents. Some are smaller and younger, while also employing the skills and complementary resources that make advanced tools easier to exploit. That combination creates both an opportunity for challengers and a reason to expect uneven gains.
Adoption is widespread in employment before it is widespread among firms
For the U.S. business population covered by the second Census Bureau AI supplement, 18% of firms reported firm-level AI use during the November 2025–January 2026 reference period. Because larger firms employ more people, the same observation represented 32% of employment when weighted by employment. Worker use of AI in work-related tasks was reported by 23% of firms, representing 41% of employment. Expected firm-level use over the following six months was 22%.
These figures should not be read as four competing estimates of one underlying adoption rate. The firm measure records whether an enterprise reports formal use. The worker measure records whether a firm reports AI use in employee tasks. The questions use different reference periods and describe different layers of activity. The employment-weighted figures answer how much employment is located in adopting firms, not what percentage of individual workers personally use AI. Still, the contrast establishes an important distributional fact: a minority of firms can account for a much larger share of economic activity when adoption is concentrated among larger enterprises.
Sector differences reinforce that pattern. Under the survey's revised instrument, current firm-weighted adoption was highest in Information at 38%, followed by Professional, Scientific, and Technical Services at 34%, Educational Services at 31%, Finance and Insurance at 30%, and Real Estate at 24%. The national firm-weighted rate was 18%. Employment-weighted use was higher still in the leading sectors: Information reached 54% of employment and was expected to approach 67% within six months. Other knowledge-intensive sectors were also projected to approach or exceed half of employment in firms using AI.
The contrast with more physical or trade-oriented sectors is consequential, but it should not be turned into a permanent map of winners and losers. Manufacturing, construction, transportation and warehousing, wholesale, and retail reported lower current rates, while some of these sectors expected meaningful increases. The data describe an uneven stage of diffusion, not a technological boundary that those industries cannot cross. The competitive question is whether firms in the lagging sectors can close the gap before early adopters accumulate learning, data, and organizational advantages.
Firm size adds another layer. Firms with 250 or more employees reported a current adoption rate of 31%, while firms with fewer than 20 employees were around 18%. Expectations also widened with size: the smallest firms expected only a modest increase, whereas firms with 100 to 249 employees and the largest firms anticipated more substantial near-term expansion. Employment-weighted estimates indicated that more than half of employment in the largest size class would soon be in AI-using firms.
A firm-size gradient can arise for several reasons. Larger enterprises can spread a fixed investment across more workers and processes. They may have internal technical staff, structured data, procurement capacity, legal support, and managers able to redesign workflows. A small firm may obtain the same cloud service at a low marginal cost but still lack time to identify useful applications, train staff, or verify outputs. Thus, the measured gap is not necessarily caused by the price of an AI tool. It may reflect the complementary conditions required to turn a tool into a reliable business process.
Geography also matters, although it should be interpreted alongside size and sector. The Census analysis finds variation across states and among firms operating across multiple states. Multi-state enterprises and technologically dense regions tend to report stronger adoption, while some regions lag. Employment weighting makes the geographic pattern more even because large employers distribute activity across locations. This is another reason not to treat a map of firm counts as a map of worker exposure or competitive capacity. A region with many small firms experimenting narrowly may look different from a region dominated by a few large employers with the resources to integrate AI deeply.
The time series contains a measurement complication. The Census Bureau revised the core question in late 2025, shifting from AI used in “producing goods or services” to AI used in “any business function.” The revised wording was intended to capture formal uses in areas such as marketing, accounting, project management, and research and development that respondents might not have associated with production. The resulting jump in reported adoption reflects both continued diffusion and the changed instrument, as well as the context created by the supplement. Pre-revision and post-revision levels therefore cannot be treated as a clean trend without adjustment.
Near-term expectations are informative but not definitive. Firms generally expected higher future use than current use, and the expected and realized series moved in similar directions, but expectations can contain optimism and may use a different frame from realized adoption. A projected 67% employment-weighted rate in Information is a signal of intended expansion, not an observed outcome. If those expectations are realized, the sectoral concentration of exposure will become more important; if they are not, the gap will reveal barriers between recognizing a use case and implementing it.
The distribution already has a competitive interpretation. Concentration in larger and knowledge-intensive firms can mean that these firms are moving first because they have more to gain from AI or because they have the assets needed to use it. Early adoption can create a lead through learning and process redesign. But a lower firm-weighted rate in a sector does not prove that it lacks competitive potential. Cloud access, specialized vendors, or a focused application may still allow a smaller firm to improve a particular product or reach a niche. What matters is whether initial use can scale and whether the benefits remain available to rivals.
The distinction between prevalence and exposure is especially important for public interpretation. If a statistic is firm-weighted, it tells us how common adoption is among enterprises. If it is employment-weighted, it tells us how much employment sits inside adopting enterprises. A high employment-weighted rate can therefore coexist with a low firm-weighted rate without any contradiction. The first describes concentration of people in adopting firms; the second describes concentration of adopting firms in the business population.
The diagram is a reading aid, not a causal model. A firm can be large and still integrate AI narrowly, or a small firm can build a valuable broad capability. The arrows identify questions for interpretation: what is the denominator, where is use located, how deep is it, and which complements determine its value?
Breadth separates tentative use from organizational capability
The extensive margin—whether a firm uses AI at all—does not reveal the intensive margin of integration. The Census supplement asks where AI is used across 15 business functions and in which worker tasks. Those questions show that current adoption is usually narrow. Among firms using AI in at least one function, Sales and Marketing was the most common location, reported by 52% of functional users. Strategy and Business Development followed at 45%, Information Technology at 41%, and Research and Development at 40%. These shares identify where adoption is concentrated among users; they do not mean that 52% of all firms use AI in Sales and Marketing.
Breadth falls quickly as the number of functions increases. About 57% of functional users reported AI in only one to three functions, and about 1% reported use across all 15. The same pattern appears in worker tasks. Writing and editing was reported by 85% of firms with task-level generative-AI use, information search by 50%, and other tasks such as document analysis and research appeared less often. Nearly 65% of firms with worker-task use limited GenAI to one to three task types. Initial adoption is therefore more often a focused intervention than a comprehensive organizational overhaul.
Narrow use is not necessarily unimportant. A single application in sales, customer communication, coding, or information search can create value if it addresses a costly bottleneck. But its competitive implications differ from those of a system that connects product development, operations, finance, data management, and customer service. Broad integration allows learning from one function to support another, spreads fixed investments, and gives management more opportunities to redesign how the firm works. It may also create more dependencies and make the system harder to replace.
The distribution of functional use is varied rather than uniform. A weighted latent-class analysis identified five descriptive profiles. Minimalist adopters, the largest group, had low use probabilities across functions and represented about 37% of functional users. Marketing specialists, about 31%, concentrated on sales, outreach, customer service, or public-facing communication. Technical strategists, about 12%, showed stronger use in research and development, information technology, strategy, and sales. Institutional-administrative integrators, about 15%, used AI across managerial, financial, human-resources, communications, and legal functions. Comprehensive adopters, only about 4% of functional users, showed high probabilities across most of the functional spectrum.
These profiles are useful because they connect breadth to organizational purpose. Marketing specialization may be a rational way for a small firm to improve outreach without rebuilding its entire operating model. Technical strategy use may signal a firm whose growth depends on product or process innovation. Administrative integration may reflect the need to coordinate a complex enterprise. Comprehensive use suggests deeper organizational capacity, but it is rare. The classes are not permanent types of business, and the statistical method does not prove that firms pass through them in a fixed order. They are a way to see that “adoption” contains several strategies.
The relationship between functions and worker tasks is also imperfect. Firms with broad functional use tend to report broader task use, but the two dimensions are not identical. AI can operate in a function as background infrastructure while workers perform largely manual tasks. Workers can use a tool in a task that has not been formalized as a firm-wide capability. Operational investment is another separate dimension: a firm may use AI in several places without changing its data practices, workflows, or staffing, or it may invest in infrastructure before task use becomes widespread.
This distinction is central to competitive capacity. A narrow use case can produce a local productivity gain, but a broad capability may be harder for a rival to reproduce because it combines software with data, routines, trained staff, and managerial knowledge. That does not mean every broad adopter will gain market power. A broad system may be expensive, poorly implemented, or quickly imitated. It means that breadth changes the mechanism by which AI can affect competition, moving the question from access to organizational execution.
The Portuguese evidence on GenAI exposure adds a second qualification. The OECD estimates a firm's exposure by combining the GenAI exposure scores of the occupations in its workforce, weighted by the employment share of each occupation. This is a measure of potential exposure, not a record that the firm has adopted GenAI or realized a productivity gain. It is also agnostic about whether exposure will result in augmentation or automation.
Firms in the highest exposure quintile were smaller in employment and somewhat younger than firms in the lowest exposure quintile. That pattern suggests a possible opening: GenAI may be particularly relevant to younger or smaller businesses whose work is rich in tasks that the technology can assist. The exposure distribution was not simply a story of the largest incumbents capturing every opportunity. Some mid-sized firms may be positioned to use the technology to challenge established providers.
But the same high-exposure firms were more productive, had a larger share of tertiary-educated workers, and showed higher markups. Firms with higher exposure were also more likely to report AI use. These relationships do not establish that GenAI caused the productivity or markup differences; the occupational composition may be reflecting the type of firm that was already more capable. They do show why potential access is not the same as effective use. A firm with programmers, analysts, data, and the ability to reorganize work may extract more value from a general-purpose model than a firm facing the same model but lacking those complements.
The pattern in turnover, value added, capital, and market share was mixed, with middle exposure groups often showing the highest values rather than the most exposed firms dominating every measure. That inverted-U pattern leaves room for contestability. High exposure is not yet exclusively an incumbent characteristic, and a mid-sized firm may use GenAI to improve its position. At the same time, higher productivity and skills among exposed firms mean that the benefits could widen differences if less capable firms cannot acquire the necessary complements.
The resulting judgment is deliberately conditional. Prevalence tells us how many firms have crossed the adoption threshold. Exposure tells us which workforces contain tasks that may be affected. Functional and task breadth tell us how widely use has spread inside an adopter. Capability tells us whether the firm can convert that use into a reliable and scalable advantage. These measures point in different directions because they observe different stages of diffusion.
The next question is what happens after a firm moves beyond a narrow use case. If AI mainly assists workers, its immediate effect may be local and reversible. If it is coupled to workflow redesign, capital replacement, and new organizational routines, it can become a structural capability. That transition is where the distribution of adoption meets the distribution of competitive power: not at the moment a firm first says “yes,” but as it learns, invests, and changes how it operates.