Census AI Study: Firm Adoption vs Worker Task Use

What did the Census study measure, and why are its rates different?

A firm’s adoption answer and a worker’s task-use answer measure different layers of diffusion. The U.S. Census Bureau’s working paper, published in April 2026, analyzes a nationally representative supplement to the Business Trends and Outlook Survey (BTOS), covering November 2025 through January 2026. It separately measures overall firm use, deployment across business functions, and workers’ use in work-related tasks. The authors report that 18% of firms used AI in a business function; when weighted by employment, that share becomes 32%. For worker-task use, the corresponding figures are 23% of firms and 41% employment-weighted. The denominator changes the question: firm-weighted rates describe how common use is among businesses, while employment-weighted rates give more weight to firms with more workers. Neither percentage means that the same share of individual workers has had a job task automated.

There is also no simple one-way path from a central adoption decision to every employee’s behavior. Census reports that workers sometimes use AI for tasks without their firm reporting formal business-function adoption, while some firms report business adoption without worker-task use. A worker may use an approved or informal tool to draft or summarize material; a firm may deploy AI in a back-office function that does not enter most employees’ daily tasks. This is why “the company adopted AI” and “my work uses AI” should not be used interchangeably.

The paper describes adoption as concentrated rather than uniform. Among firms using AI, 57% reported use in three or fewer business functions. Worker-task use was also narrow: 65% of firms limited it to three or fewer tasks. Sales and marketing, strategy and business development, and IT were leading business functions; writing, document analysis, and information search led generative-AI task use. These categories show where respondents reported use, not how much time was saved or whether outputs met workplace standards.

Pew Research Center offers a useful second lens, but not a competing estimate of the same quantity. In its September 2025 survey of 5,010 U.S. workers, 21% said at least some of their work was done with AI, and 65% said they used it little or not at all. Pew asked workers about their own experience; BTOS asks businesses about adoption and worker task use. Different respondents, wording, and reference periods mean the 21% should not be placed beside the Census 23% as if one validates or overturns the other. An earlier Census BTOS paper used earlier questions and a different observation period; its adoption estimates should likewise not be stitched to the 2026 supplement as a clean trend line without accounting for those changes.

The practical reading is modest but useful: firm adoption is an organizational signal; task use is closer to the work a person can inspect. Both remain survey reports. They do not directly observe every tool, prompt, output, quality check, degree of use, or production workflow.

Sources: The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks; About 1 in 5 U.S. workers now use AI in their job, up since last year; Tracking Firm Use of AI in Real Time: A Snapshot from the Business Trends and Outlook Survey

What can worker task use tell me that firm adoption cannot?

Task-use evidence points toward activities changing, but it cannot by itself tell you whether AI performs them reliably, whether a manager has redesigned the workflow, or whether a position will be removed. The Census paper says writing, document analysis, and information search are leading generative-AI uses. It also reports that 66% of users relied on AI solely to augment tasks in the measured period. That is evidence about reported use and stated role in a task, not proof that augmentation will always remain the outcome or that a task is unaffected by automation.

A task can be exposed to a tool’s capabilities without being used at work. Use can occur without formal company adoption. Adoption can occur without a specific worker using AI. And task use is not the same as displacement: headcount depends on how an organization combines tools with staffing, demand, process design, costs, and accountability. These are separate steps in a chain, not interchangeable labels. In the Census analysis, AI-related employment decreases were reported by 2% of firms. Broader functional integration and operational investment were positively associated with employment decreases, while worker-task integration alone showed no significant link to headcount reduction after accounting for those factors. Those are associations in this survey analysis, not causal estimates or a forecast. The low observed share also does not guarantee that future effects will be small.

Pew’s February 2025 chatbot report helps make the distinction concrete at task level. Among workers who had used chatbots at work, common reported uses included research or information search, editing written content, and drafting reports or documents. Yet among nonusers, 36% named lack of a job-related use as a major reason for not using chatbots. This suggests that tool availability and task fit vary. The survey is about chatbots and worker reports, not all AI systems or formal employer deployment, so it cannot establish the prevalence of all AI use.

For a knowledge worker, the useful unit is therefore a small workflow, not an occupation label. Consider a recurring monthly report: collecting figures may be suitable for assisted extraction; checking whether records match, resolving exceptions, explaining an unusual change, and signing off on the result may carry different error costs and responsibility. This is an illustration, not a claim from the study. If a tool drafts a first pass, the next questions are whether the source material is appropriate to share, how errors are detected, who verifies the numbers, and what happens when the output is wrong. These practical boundaries can determine whether the tool removes effort, shifts effort into review, or changes who owns a task.

That is why an occupational exposure label or company adoption headline cannot substitute for inspecting your actual work. List the repeated tasks where information is gathered, transformed, drafted, classified, or checked. Then identify the output’s consequence, the exceptions that require judgment, and the person accountable for the final result. The Census rates help prioritize attention; they do not score an individual’s risk.

Sources: The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks; Workers’ experience with AI chatbots in their jobs

What is a realistic next move if AI is entering my tasks?

Start with a bounded task audit before paying for a course or planning a pivot. For two weeks, note a handful of recurring tasks: what goes in, what comes out, how often the task occurs, which part is assisted or could plausibly be assisted, how much checking is needed, and what knowledge or relationship work remains yours. Include a simple distinction between observed use in your team and theoretical capability. Do not put confidential work material into an unapproved tool to run this experiment; ask about workplace policy or use a permitted environment.

Then compare three levels of response. First, redesign within the current role: test one low-consequence, reversible task with an approved tool, keep a human check, and compare total time and error correction with the existing method. If the result is useful, document the process and the boundary conditions. Second, consider an adjacent move if your task audit shows that your strongest contribution is shifting toward review, domain interpretation, client explanation, or process ownership. Look for overlap with work you already do and verify the actual prerequisites in local job postings or conversations. Third, consider a larger change only when repeated evidence points to a shrinking or undesirable task bundle and the alternative fits your salary floor, geography, training time, health, and family responsibilities.

The learning response should match the intended move. If you want to use AI in your existing field, build practical literacy around task selection, data handling, evaluation, and verification; a small project using a permitted workflow may show more than collecting a generic badge. If you want to build AI-enabled products, add programming, data, and product practice appropriate to the projects you want to make. If your goal is software practice or machine-learning engineering, compare degree, course, certificate, project, apprenticeship, and self-study routes against prerequisites, depth, feedback, signaling, time, and cost. Research engineering or academic work may require deeper mathematics and formal study than workplace tool use. No single credential guarantees readiness or hiring.

Use evidence in proportion to the decision. A worker-task rate can justify reviewing your workflow; it cannot tell you whether to resign. A business-adoption measure can suggest that organizational integration deserves attention; it cannot show that every team or location has adopted tools. The study’s employment associations warrant attention to broader integration and investment, but they do not justify turning the results into personal replacement odds. The evidence is U.S.-specific and based on reported activity in one survey period, so workers elsewhere should treat it as context, then check local conditions.

Verdict: use firm adoption to understand the direction of organizational change and worker-task use to choose what to examine first. Neither is a displacement forecast. Your next move is to map one recurring task, its review burden, and its consequences, then test a permitted low-risk improvement while preserving the experience and judgment that make the result dependable. Escalate to training or a career change only if that local evidence, your goals, and your constraints point there. If you want a structured first pass, the free [task-level change-pressure checker](/ai-job-risk-checker) can help organize tasks; it reports change-pressure signals, not a validated probability of job loss. For comparing a stay-and-redesign path with adjacent and larger-change options around your experience and constraints, the [career roadmap](/career-roadmap) is available as a separate paid option and does not guarantee employment or income.

Sources: The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks; About 1 in 5 U.S. workers now use AI in their job, up since last year; Workers’ experience with AI chatbots in their jobs

Questions readers ask