The September 2026 Census working paper finds that firms with more AI-exposed occupational profiles are more likely to report AI adoption. The association is positive but modest, varies across measures, firm sizes, and sectors, and is not evidence of causation. It validates exposure as a broad signal for where adoption may be worth investigating; it does not show which workers use AI, how deeply tasks have changed, or whether jobs will be lost.
What did the September paper test—and what counts as exposure here?
The September 2026 Census working paper asks whether firms with more AI-exposed occupational employment profiles are more likely to report adopting AI. It directly compares an occupation-based exposure measure with firms’ reported adoption, testing whether exposure scores line up with a business signal. It does not establish what any particular employee does each day.
In “AI Exposure and Adoption Among U.S. Firms,” researchers combine 14 occupational exposure measures from prior literature with occupational employment shares from the Bureau of Labor Statistics’ Occupational Employment and Wage Statistics program. This weighting creates a firm-level profile: occupations that make up more of a firm’s workforce contribute more to its measure. The researchers compare these profiles with direct firm adoption measures from the Census Bureau’s Business Trends and Outlook Survey. The Census record reports a positive, statistically significant association, but says exposure explains only a modest share of adoption variation. This is a working paper, and its abstract does not establish a causal effect.
Here, occupational exposure means an estimate of how applicable AI capabilities may be to work associated with an occupation. It is not observed AI use. A firm’s reported adoption does not show which workers use AI, how regularly it enters a workflow, or whether responsibilities have changed. Neither measure establishes displacement. The chain has distinct steps: estimated applicability, a workforce-weighted firm profile, and reported business adoption. Each adds information, but none gives an individual’s task mix or predicts job loss. For a worker, the paper makes exposure more useful as a broad signal for where adoption may be worth investigating, while leaving the daily-work question open.
How was a firm-level exposure score assembled, and what does that aggregation hide?
A firm-level exposure score is assembled in two steps: researchers start with estimates of how exposed different occupations are to AI, then combine those estimates with the occupations represented in each firm’s workforce. The September 2026 Census working paper says it links 14 such firm-level measures, built from occupational exposure estimates in prior literature and occupational employment shares from the Bureau of Labor Statistics’ Occupational Employment and Wage Statistics program, to firms’ direct adoption reports in the Business Trends and Outlook Survey (BTOS). The resulting score summarizes the occupational composition of a firm; it does not record the tasks a particular employee performs or calculate the share of that employee’s work that a system can automate. That distinction is easy to miss because the ingredients sound close to daily work. An occupation label is a useful organizing category, but people with the same title can spend their time differently. One analyst might prepare recurring summaries from standardized records; another might spend more time reconciling exceptions, explaining results to colleagues, or checking information that does not fit the template. A firm-level average cannot preserve those differences. Nor does an occupational share reveal whether a task happens in a process with usable data, approved software, privacy controls, reliable outputs, and someone able to check the result. These are examples of information the aggregated score does not contain, not findings the paper separately measures. The measure is therefore a broad location signal. It lets researchers ask whether firms whose occupational mix scores as more exposed also report adopting AI. In this design, the adoption outcome comes from BTOS rather than being inferred from the exposure estimate itself. The paper reports a positive association, while also calling exposure an incomplete signal. That is a meaningful bridge between a model of which kinds of work may be technically affected and a survey asking firms about use. It still leaves a gap between the firm’s reported adoption and what happens in any one employee’s workflow. The survey context matters too. In “Large Firms With at Least 20 Employees Biggest AI Users,” the Census Bureau describes BTOS as a biweekly view of business AI implementation and says the core question changed in November 2025. It shifted from asking about AI used “in producing goods or services” to asking about AI used “in any business function,” a broader frame that can include task support such as drafting. The same story notes that BTOS also asks about expected use and, in its supplement, use across business functions and operational changes. Those details help explain what a firm report can mean; they do not turn the occupational exposure score into an inventory of implementation depth. For a worker, the practical limit is straightforward: a firm’s score cannot tell you which parts of your own week are changing. Treat it as a reason to look more closely at an industry or employer, then map your recurring tasks and ask where tools are discussed, tested, or routinely used, and what review remains. Your task map answers a different question from the firm-level measure. Neither, on its own, establishes that a role will disappear.
Sources: AI Exposure and Adoption Among U.S. Firms; Large Firms With at Least 20 Employees Biggest AI Users
What does the positive but modest association actually say?
The September 2026 Census working paper finds a positive association between firms’ occupational exposure to AI and their reported adoption. Across 14 standardized exposure measures, a one standard deviation higher firm exposure was associated with a 4–11 percentage-point higher adoption probability. With year and subsector fixed effects, the range fell to 1–8 percentage points. The paper calls the relationship statistically significant but modest in explanatory power. These are working-paper results, not a final official statistical series. [“AI Exposure and Adoption Among U.S. Firms”](https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-61.html) measures firm-reported adoption, not worker use or job loss.
A standard deviation describes a position in the distribution of the paper’s firm-level exposure measure; it does not mean a fixed number of exposed tasks or a personal score. Percentage points compare probabilities on a 0-to-100 scale: a rise from 20% to 25% is five percentage points. The ranges summarize results across alternative measures and specifications. They are not a single estimate with a confidence band, nor odds that a worker in an exposed occupation will lose a job. The outcome is firm adoption, not an individual’s use, task automation, or displacement.
The smaller estimates after adding year and subsector controls show that timing and industry composition account for part of the raw relationship. The association remains positive in the controlled specifications reported in the abstract, but this does not establish that exposure causes adoption. The analysis links firms’ occupational composition to their survey reports; other firm and workplace features may shape both. Controls narrow the interpretation, but do not make the comparison an experiment or identify a causal effect.
A skeptical reading is that modest explanatory power makes exposure useless. That goes too far. A broad measure can indicate where adoption is more common on average while still predicting little about any particular firm. Variation across measures, firm sizes, and sectors reinforces that limit. Yet a screening signal can help prioritize what to inspect without explaining every decision. Direct observation can then distinguish tools that are discussed, piloted, embedded in a workflow, or used informally by workers.
The practical reading is narrow: occupational exposure contains information about firm adoption, while leaving substantial variation unexplained. Use it to decide where investigation may be worthwhile, not to forecast a specific employer’s adoption or a person’s job security. For a worker, the next question is whether recurring tasks, review duties, access rules, or responsibility for errors are changing locally. The paper cannot answer that. It shows exposure is informative at firm level and leaves a task-level check necessary in practice.
Why does the relationship vary by measure, firm size and sector?
The Census working paper reports that the exposure–adoption relationship differs across its 14 measures, firm sizes and sectors. There is no single gradient to apply to every business: higher occupational exposure is associated with greater reported adoption on average, but the strength of that association depends on how exposure is defined and which firms are compared. The abstract establishes heterogeneity; it does not identify why the differences arise. Treating the spread as one universal effect would erase this qualification.
The measurement choice matters because the 14 measures are alternative estimates of how occupations may be affected by AI, aggregated into a firm profile using its occupational employment mix. They can encode different judgments about which work activities AI can apply to, so two measures need not assign the same exposure level to a firm. This sensitivity does not make them forecasts for an employee. The paper’s outcome is firm-reported adoption, not observed task change or a worker’s likelihood of displacement.
Firm size and sector add context. Census’s May 2026 BTOS account, “Large Firms With at Least 20 Employees Biggest AI Users,” reports adoption differences by size and industry in data collected from December 2025 through May 3, 2026. Firms with at least 250 employees reported higher use than smaller firms; Information and Finance and Insurance were above the national rate, while Retail Trade was below it. Census also says BTOS broadened its question in November 2025 from AI used in producing goods or services to AI used in any business function. These descriptive patterns illustrate why adoption context matters, but do not explain the September paper’s variation.
Resources, data availability, workflow fit, privacy controls and capacity to train staff or check outputs are plausible reasons adoption could differ. They are hypotheses, not findings established by the abstract or BTOS account. A firm might have technically compatible tasks but no approved tool or worthwhile process; another might adopt AI narrowly without changing most employees’ responsibilities. The sources do not show which mechanism explains a particular company’s choices.
For a worker, national and sector patterns can guide where to look, but cannot tell whether a tool is available in a team, used in recurring tasks, or changing responsibilities. Ask what is happening in the workflow: is AI only discussed, being tested, or used routinely, and who verifies its output? That local evidence matters more to a career decision than assuming a sector average describes an employer or an occupational score describes every worker.
Sources: AI Exposure and Adoption Among U.S. Firms; Large Firms With at Least 20 Employees Biggest AI Users
Why can broad exposure and patchy task use both be true?
A role can include tasks AI could assist with without AI assisting every worker or task. The St. Louis Fed analysis “What Work Does Generative AI Do?” measures reported use by occupation and task, showing broad reach but patchy use. These are different points in a chain, not conflicting findings.
The analysis draws on nearly 14,000 workers in four quarterly waves of the Real-Time Population Survey, from August 2025 to May 2026. Respondents identified their occupation, reported which of its ten most important O*NET tasks they performed, then indicated which tasks AI regularly helped them complete. This yields separate measures of workers using AI in an occupation and task performers using it for that task.
In more than 80% of occupations, at least one in five workers reported job-related AI use. Yet fewer than 3% of listed tasks had adoption above 50%, and none exceeded 70%. Workers in the same occupation may use AI for different parts of the job, so high overall use need not mean one task is routinely handed over by everyone. These are survey reports, not a direct observation of every workplace workflow.
That distinction limits what occupational exposure can tell an individual. Exposure estimates describe technical applicability; adoption records reported use. The Fed analysis finds exposure scores explain roughly half of variation in adoption in some occupation- and task-level analyses, and are poor predictors of whether a particular worker adopts AI. People doing similar tasks may use it differently.
The study reports lower-than-predicted use for work involving sensitive records, citing privacy rules and error costs, and higher-than-predicted use in some hands-on occupations where information gathering or planning can be assisted. The authors also suggest latitude over work methods matters. Access, permission, trust and verification effort are further plausible influences to check locally, not universal findings established by this survey.
Use broad exposure to decide where to look, then inspect your own recurring tasks. Mark two or three as absent, discussed, tested or routinely supported by AI; note who verifies the output and whether responsibilities or quality standards have changed. The worker survey adds a closer lens than firm adoption, but it covers self-reports and selected O*NET tasks, not every job or job-loss outcomes. Local observation remains necessary. This is a task-level check, not a prediction about whether your role will remain.
Sources: What Work Does Generative AI Do?
What does a firm adoption report say about implementation depth?
A firm adoption report answers a narrower question than “Are employees using AI in their work?” The April 2026 Census working paper, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks,” separates three layers: whether a firm uses AI, deploys it across functions, and workers use it on tasks. Its nationally representative Business Trends and Outlook Survey supplement covered November 2025 through January 2026. In that period, 18% of firms reported use in a business function; weighted by employment, 32%. Separately, worker task use was reported by 23% of firms, or 41% employment-weighted. These are distinct measures, not competing estimates of one rate. Adoption can be uneven inside a business. Among firms using AI in a business function, 57% used it in three or fewer functions. Sales and marketing, strategy and business development, and IT were most common. Writing, document analysis, and information search led generative AI task uses, while 65% of firms limited use to three or fewer tasks. A business can answer “yes” while use remains concentrated. That answer does not tell an employee whether their tasks changed, how often a tool is used, or whether it is part of a controlled workflow. The paper describes top-down and bottom-up routes into use. Worker task use sometimes appears without formal firm adoption, while formal adoption sometimes appears without workers reporting task use. Procurement or policy can precede routine use; an individual may experiment before the organization records adoption. These pathways do not prove a pilot will persist or alter a role. “Available,” “tested,” and “embedded in a recurring process” are separate evidence states. Labor outcomes require another distinction. The abstract reports that 66% of users relied on AI solely to augment tasks, and 2% of firms reported AI-related employment decreases. Functional breadth and operational investment were positively associated with employment decreases, while worker-task integration had no significant link to headcount reduction after those factors were considered. These results do not show task use caused job cuts or measure an individual’s risk. A use report cannot stand in for implementation depth or employment impact. At work, ask what stage the evidence describes: informal experimentation, a limited pilot, routine use in named tasks, or process redesign tied to operational investment. Then ask who checks outputs and whether responsibility, standards, or staffing decisions changed. The September study’s firm adoption outcome helps compare broad uptake with occupational exposure; this April comparison clarifies what an adopter label can conceal. Inspect the workflow before drawing a career conclusion.
Sources: The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks
How much do survey wording and reported productivity change the interpretation?
Survey wording changes what “adoption” includes, while a worker’s estimate of time saved answers another question. Neither substitutes for the September paper’s firm-level association. The Census Bureau’s May account of the Business Trends and Outlook Survey (BTOS) says its core question changed in November 2025: instead of asking whether businesses used AI “in producing goods or services,” it asks whether they use it “in any business function.” The broader wording can include support work such as drafting emails, even when AI is outside production of a firm’s goods or services. The account summarizes biweekly business survey data collected from December 2025 to May 2026. It helps define the reported outcome; it does not show that the September paper’s association is causal. BTOS asks whether a business used AI in the past two weeks and whether it expects use in the next six months. Its later supplement asks about use across 15 business functions and operational changes such as training, workflow adjustments and technology investment. These distinguish current reported use, expected use and signs of implementation. A firm expecting adoption has not necessarily embedded AI in routine work. A firm reporting use does not establish which employees use it, how often, or whether responsibilities changed. The Census Bureau’s August account of March 2026 Household Trends and Outlook Pulse Survey (HTOPS) data asks workers about 11 listed job tasks. It reports that 56% of U.S. workers had used AI for at least one task. Recent users were asked how many additional hours they would have needed to finish work without AI: a quarter reported less than an hour saved, 31% reported one to two hours, 10% no time saved, and 4% said AI required extra time. These are recalled estimates of a counterfactual, not timed experiments measuring productivity. The Census account says updated weights revised estimates but did not change statistical-test conclusions. HTOPS is a worker task-use lens, unlike BTOS’s business-level adoption measure. Read the surveys as complementary. BTOS asks what businesses report doing; HTOPS asks workers about listed tasks and perceived time effects. Before applying an adoption headline to your workplace, check who answered, what counted as use, and the measurement period. Reported time saved may prompt closer task review, but it does not establish net output, quality, firm-wide adoption or staffing change.
Sources: Large Firms With at Least 20 Employees Biggest AI Users; AI Use at Work
Does adoption evidence settle the question of labor demand?
No. The September Census working paper asks whether firms with more AI-exposed occupational profiles are more likely to report adopting AI. That links workforce composition to reported business uptake; it does not measure job postings, layoffs, or the risk to a particular employee. It cannot be turned into a displacement estimate. The Dallas Fed’s “Job postings show early signs of AI automation impact” analyzes millions of Lightcast online job ads using an occupational exposure measure based on observed Claude use. It compares postings for more- and less-exposed occupations within the same industries, In Texas, postings for more-exposed positions fell 5 percent relative to less-exposed positions by the end of 2023 and about 8 percent by the first quarter of 2025. The U.S. occupation-industry comparison produced similar results. These are relative changes in advertised openings, not jobs eliminated. The headline aggregate estimates concern Texas; the U.S. occupation-industry comparison is a separate check. Lightcast covers jobs commonly advertised online, and the Dallas Fed notes that farming, construction, building maintenance and personal-service openings are underrepresented. The exposure measures also differ: the Dallas Fed maps tasks to observed use of a particular tool, while the Census paper aggregates literature-based occupational estimates using firms’ occupational employment shares. Together, the studies do not establish a path from adoption to labor demand. A firm may report adoption before its workflow changes, while postings can move for reasons beyond AI. Adoption could later alter task allocation or hiring without an immediate layoff. The Dallas Fed describes early changes in online demand; it does not show that every exposed firm reduced hiring or establish why each posting changed. Its outcome and unit differ from the Census measure. For a worker, watch local evidence separately: changes in responsibilities or quality standards, new capabilities requested in relevant job ads, and shifts in hiring patterns. These signals prompt review; none gives an individual job-loss probability. Census exposure helps identify where adoption may be more common, while posting evidence offers another view of labor demand. Neither settles what is happening in one person’s job.
Sources: Job postings show early signs of AI automation impact; AI Exposure and Adoption Among U.S. Firms
What is the fair verdict on the study’s contribution?
The fair verdict is that “AI Exposure and Adoption Among U.S. Firms” makes occupational exposure a useful screening signal, but not a substitute for what a worker does. The September 2026 Census working paper compares firms’ occupational exposure profiles with their survey-reported adoption. The association is modest and varies across measures, firm sizes, and sectors. Exposure tracks some uptake differences; it does not show causation or predict an individual’s job loss. Modest does not mean meaningless. Exposure can help identify where adoption is more common without explaining most differences between firms. Occupational shares summarize a workforce, though people with the same occupation may do different tasks. A firm adoption response also says little about how deeply a tool enters a workflow. The result validates one link in a chain, not the path from capability to changed responsibilities or employment. The comparison studies show what remains missing. “What Work Does Generative AI Do?” finds use across many occupations but less often across any particular task; its unit is workers and tasks, not firm exposure. “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks” separates firm adoption, function deployment, and worker-task use, showing that these layers can diverge. Together, they make the September study a bridge between exposure estimates and observed uptake, not a full account of implementation. A stronger predictive case would require repeated, comparable links from exposure to task use and implementation depth. The September abstract does not establish those links, causal direction, or employment effects. Exposure models help researchers choose where to look. For an individual decision, evidence about actual tasks, verification, accountability, and durable workplace changes matters more.
Sources: AI Exposure and Adoption Among U.S. Firms; What Work Does Generative AI Do?; The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks
What should a worker check next?
Choose two or three recurring tasks and record whether AI is absent, discussed, tested, or used routinely. Note what still needs a person to verify or approve, what information is restricted, and whether expectations for speed, quality, or accountability have shifted. This answers a different question from the September paper, “AI Exposure and Adoption Among U.S. Firms.” That working paper links firms’ occupational exposure profiles to reported adoption; it does not show which tasks in your role have changed or how deeply a tool has entered a workflow. A demonstration or informal trial is a reason to observe, not by itself evidence that the job has been redesigned. Look for repeated use and durable changes in responsibilities before committing time or money to a move. The free task checker can organize this review into change-pressure signals; it does not estimate your probability of losing a job. Exposure helps choose where to look. Observed changes in tasks, checks, and accountability should guide what you investigate next.
Questions readers ask
Does the September Census study estimate an individual worker’s chance of losing a job?
No. It links firms’ occupational exposure profiles with firm-reported AI adoption. It does not estimate individual displacement probability.
Why can occupational exposure be high when task-level AI use is limited?
Exposure estimates technical applicability, while task-use reports measure observed or self-reported use. Access, workflow fit, permissions, verification, and other conditions can affect whether a capability is used.
What is a reasonable next step for a worker?
Map two or three recurring tasks and note whether AI is absent, discussed, tested, or used routinely. Record what still requires human review and whether responsibilities or standards have changed.
Sources and notes
- AI Exposure and Adoption Among U.S. Firms
Supports the design and reported results of the September 2026 working paper linking firm occupational exposure profiles to firm-reported adoption.
- Large Firms With at Least 20 Employees Biggest AI Users
Supports the BTOS question wording change and descriptive context on reported business AI adoption by size and sector.
- What Work Does Generative AI Do?
Supports worker- and task-level survey comparisons, including broad occupational use alongside patchier use across listed tasks.
- The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks
Supports distinctions between firm adoption, business-function deployment, and worker-task use in the November 2025 to January 2026 survey period.
- AI Use at Work
Supports March 2026 HTOPS worker task-use reporting and respondents’ estimates of time saved or added.
- Job postings show early signs of AI automation impact
Supports a separate job-posting labor-demand signal and its geographic, exposure-measure, and online-posting coverage limits.
Apply this to your own work
See the whole job market at once.
Explore which occupations AI may reshape, then turn the signal into a practical response.
Explore the job map