How AI Diffuses Inside a Firm
AI diffusion inside a firm is not a single event. The useful question is how formal adoption, deployment across business functions, and worker-task use connect—or fail to connect—because that alignment determines whether access becomes an operating capability. A firm-level answer tells us that an organization recognizes or reports AI use; it does not by itself tell us where the technology operates, who interacts with it, or whether it is changing how work is done. Understanding those layers lets readers interpret adoption statistics and judge which kind of diffusion could eventually affect competitive strength.
Three layers of use inside one enterprise
The Census Bureau's business survey treats AI diffusion as a structure with three connected layers. The first is firm-level adoption: does the enterprise report using AI? The second is functional deployment: in which business functions, such as sales, finance, information technology, or research and development, is AI used? The third is worker-task use: do employees use AI in particular work-related activities? Each layer has a different unit of observation and a different visibility problem. They should be read together, not collapsed into a single measure.
Firm-level adoption is the broadest and simplest layer. It captures a formal organizational answer about whether AI is in use, but the answer depends on how the respondent understands “use” and the reference period in the question. In the Census supplement, the core firm question concerns recent use, while other questions ask about use over a longer six-month window. A business that used an AI system for one project, a background prediction service, or a recurring process may answer differently depending on which interval is considered. The firm-level response is thus useful for measuring the extensive margin—whether an enterprise has crossed into reported use—but it says little about intensity.
The second layer is deployment across business functions. This reveals whether AI is confined to one part of the enterprise or has been integrated into several. The survey asks about 15 functions, including customer-facing work, strategy, finance, information technology, human resources, research and development, operations, and compliance. The purpose is not to create a ranking of departments. It is to identify the organizational surface across which AI is expected to operate. A firm using AI only in sales and marketing has adopted something important, but its integration problem differs from that of a firm using AI in sales, finance, IT, and product development.
Function-level deployment also supplies an interface between a firm's formal decision and the work performed inside it. Management may authorize an AI service centrally, but the service becomes economically meaningful only when a function incorporates it into a workflow, data process, or decision. Conversely, a function may begin experimenting before the organization has a formal policy or a shared infrastructure. Looking at functions therefore exposes choices about scope and coordination that a yes-or-no firm measure hides.
The third layer is worker-task use. Here the relevant question is whether workers use AI in tasks connected to business activity, not how many workers use it or how frequently. The survey's worker measures are reported by firm respondents rather than collected separately from employees. A positive answer may mean that at least some workers use AI in a relevant task; it does not establish that a particular share of employees uses it. That distinction matters when a firm has thousands of workers but only a small group using an AI tool, or when use is uneven across roles.
Worker-task use can also be indirect. AI may process information, optimize a system, or generate an output that a worker receives without operating the tool themselves. In that case, the firm and function can be deeply reliant on AI while the associated task remains partly manual. A worker may be a recipient of an AI-generated recommendation rather than an active user. Conversely, an employee may use an external writing or analysis tool in a task without that tool being part of the firm's formal infrastructure. The same organization can contain both patterns.
This is why the layers should be understood as interfaces rather than as a ladder in which one automatically leads to the next. Firm adoption can provide authority, budget, procurement, and governance. Functional deployment turns that decision into a local process. Worker use connects the process to particular activities. But information, control, and responsibility do not necessarily travel cleanly across those boundaries. A formal system can be installed without changing a worker's routine; a worker can change a routine without changing the firm's official answer.
A hypothetical customer-service firm illustrates the difference. Suppose management installs an AI system that classifies incoming requests in the background. The firm and its customer-service function may both count as adopters even if representatives never open an AI interface. If representatives independently use a generative tool to draft replies, worker-task use can exist before that practice is formally approved. If management later embeds drafting into the official service workflow, all three layers may align. These are different states of organizational integration, not contradictions about whether AI exists.
The diagram captures possible paths, not a required sequence. The upper path represents a formal implementation that moves from an organizational decision into a function and then into worker activity. The lower path represents worker-led use that may not yet be visible in formal adoption. Background automation can also stop the upper path before worker interaction. The point is to preserve the interfaces: who authorizes use, where the system is integrated, and whether a worker's task actually changes.
Two channels, one measurement problem
The observed overlap in the Census data makes the distinction measurable. During the supplement reference period, formal firm-level AI use was about 18% of firms, while worker use in work-related tasks was about 23%. These rates answer different questions and use different reference frames, so the difference should not be interpreted as a precise estimate of hidden adoption. It does, however, show that the organizational and task layers are not identical.
Excluding “don't know” responses from the Venn comparison, 14.4% of firms reported both firm-level and worker-level AI use. Another 6.8% reported worker use without reported firm-level adoption, while 2.5% reported firm-level adoption without corresponding worker-task use. Among firms with any reported AI presence in those two layers, about 61% showed both, 29% showed worker use alone, and 10% showed firm-level use alone. The employment-weighted pattern points in the same direction, although the shares differ because larger firms are more likely to use AI.
The conditional rates reveal an asymmetry. Among firms reporting formal AI adoption, 80.8% also reported worker-level integration. Among firms reporting worker-task use, only 64.0% reported formal firm-level use. This does not prove that workers normally lead management, or that experimentation eventually becomes policy. The survey is not a longitudinal record of adoption events, and different question wording and respondent knowledge can generate part of the mismatch. It does show why a bottom-up channel must be taken seriously alongside top-down implementation.
A top-down path begins with an organizational decision. A firm may purchase a service, change a workflow, train staff, or deploy AI as infrastructure. Worker use may follow when the system is made visible and relevant to a task. But formal adoption can precede practical integration: a company may have acquired the technology while workers are still learning how to use it, or it may use the system in a background process that has little direct contact with employees. In competition terms, this path can signal investment and a capacity for structural redesign before it creates a visible change in individual tasks.
A bottom-up path begins with a worker, team, or department finding a useful application. The initial tool may be external, informal, or too localized to appear in a firm's formal adoption answer. It can nevertheless reveal a use case, save time, or change a task. If the practice proves valuable, managers may later codify it, connect it to firm data, and extend it to other functions. But that possible progression is an interpretation, not an observed causal sequence in the data. Some experiments will remain isolated; others may stop because of privacy, security, quality, or governance concerns.
The two channels also have different visibility risks. A centralized respondent may know that the firm bought an AI service but not know how often employees use independent tools. The same respondent may know that workers use an application but not classify it as formal AI adoption, particularly if the application is embedded in ordinary software or a department's workflow. The survey's design can therefore be conservative about decentralized use while still giving a useful measure of formal integration. It can also produce a positive correlation between answers when one respondent assumes that firm adoption implies worker use.
The distinction between worker AI use and worker generative-AI use adds another boundary. The survey reports that worker use of AI in business-function tasks was about 23% of firms, while worker use of generative AI in work-related tasks was about 21%. Some firms reported generative-AI use without the broader worker-AI measure, and some reported worker AI without generative AI. The wording helps explain why: one question ties worker use to tasks supporting business functions, while the other uses a broader formulation about work-related tasks. Specialized predictive systems, workflow automation, and generative tools can therefore occupy different parts of the task landscape.
These distinctions change how adoption rates should be used to reason about competition. Firm-level adoption is most informative about formal organizational commitment and the potential for coordinated investment. Functional breadth is more informative about whether AI is becoming part of multiple processes and whether a firm can transfer learning across them. Worker-task use is most informative about operational contact with work, including decentralized assistance that may improve output without a formal redesign. None of the three is a direct measure of market power.
The layer that matters depends on the claim. If the question is whether firms have begun buying or implementing AI, a firm-level rate is relevant. If the question is whether AI is becoming a general organizational capability, functional breadth is more revealing. If the question is whether workers are using AI to augment or alter tasks, worker-task data are the closer measure. A statement about future productivity or competitive advantage needs more than any one layer: it requires evidence about complementary investment, the firm's ability to scale a use case, and the responses of rivals.
There is also a denominator problem. Firm-weighted estimates give every enterprise one vote, so they describe how widespread adoption is among businesses. Employment-weighted estimates give greater weight to firms with more employees, so they describe how much of the workforce is located in adopting firms. A technology can be present in a minority of firms yet reach a large share of employment if adoption is concentrated among large enterprises. The reverse can also occur when many small firms use AI narrowly. Neither denominator is the universally correct one; each answers a different question.
That distinction is directly relevant to concentrated markets. Large firms may show high employment-weighted exposure because they possess the resources to adopt, but high exposure does not tell us whether they have integrated AI deeply or merely purchased a tool. Small firms may be numerous among task-level experimenters, but isolated use does not demonstrate that they can scale a product or challenge an incumbent. The three-layer model prevents both errors: it keeps the unit of adoption visible and asks how use crosses organizational interfaces.
The practical judgment is therefore not to choose between “firm adoption” and “worker adoption.” It is to ask whether the layers reinforce one another, remain decoupled, or reveal different strategies. Alignment across formal adoption, functions, and tasks may indicate that a firm is building a coordinated capability. Worker use without formal adoption may indicate experimentation and a possible source of new productivity, but also weak governance or limited visibility. Formal use without worker integration may indicate implementation lag, background automation, or a deployment whose economic effects have not yet reached daily work.
Once those states can be distinguished, the next question is distribution: which kinds of firms occupy them, which sectors move first, and how often initial use expands from a narrow task or function into a broader capability. The competitive meaning of AI diffusion depends not only on the channels through which it enters a firm, but also on who can sustain and extend it.