What the St. Louis Fed’s 2026 AI Adoption Evidence Shows

What changed in the 2026 evidence?

The important change in the St. Louis Fed’s 2026 evidence is finer measurement, not a sudden event in September. The new release links workers’ reported use of generative AI to both their occupation and specific work activities, giving researchers two views: the share of workers in an occupation who use AI for their jobs, and the share of people who perform a listed task who say AI regularly helps them with it. That is more informative about reported use within work than a broad question about whether someone uses AI at work. It still does not predict what will happen to any particular job. The window matters. The indexes draw on nearly 14,000 U.S. workers ages 18 to 64 surveyed in four quarterly waves from August 2025 through May 2026. The St. Louis Fed’s working paper, “What Work Does Generative AI Do?,” is dated August 25, 2026; the Fed’s summary appeared September 1. Those dates mark publication of an analysis of earlier survey responses, not a measured surge in adoption that occurred when the article appeared. The data describe a period and population, and the authors say future waves will update the indexes. Readers should treat them as a snapshot of reported assistance, not a live reading of their employer or a personal forecast. The survey sequence explains what is new. The Real-Time Population Survey first elicits a respondent’s occupation. It then presents the ten detailed activities O*NET rates as most important for that occupation and asks which the worker performs. Among workers who use AI at work, it asks which of those activities AI regularly helps them complete. The result connects a person, a task they say they do, and reported AI assistance. It is not a capability test of what software could do, nor a count of platform conversations. This differs from the St. Louis Fed’s 2024 survey analysis, “The Rapid Adoption of Generative AI,” which examined broad work activities and estimated the share of work hours involving AI assistance. That earlier work helps show how the Fed was tracking use, but its activity menu and work-hour framing are not interchangeable with the 2026 occupation-and-task indexes. A change in instrument can sharpen the question without proving a directly comparable change in the underlying rate. The defensible verdict is therefore specific: the 2026 release makes reported task use easier to locate and compare, while leaving quality, time saved, employer deployment, and employment effects unanswered. A worker can use it to identify a task worth examining locally, then check whether assistance is permitted and how the result is verified before deciding what to do.

Sources: What Work Does Generative AI Do?; What Work Does Generative AI Do?; The Rapid Adoption of Generative AI

What exactly does the task index count?

The task index estimates a specific conditional share: among surveyed workers who say they perform a listed activity, what share report that generative AI regularly helps them with it? The occupation index answers a different question: among surveyed workers in an occupation, what share report using AI for their job? Neither figure says how much of a task AI completes, how much time it saves, whether its output is correct, or whether it changes the worker’s job. “What Work Does Generative AI Do?,” from the St. Louis Fed, describes reported adoption measures built from the Real-Time Population Survey, not performance tests or employer deployment records. The sequence creates the task denominator. Respondents identify their occupation, indicate which listed activities they perform, and, if they use AI at work, report which tasks AI regularly helps them complete. A task’s rate concerns people who say they do that task, not every worker in the country and not every AI conversation tagged with a similar label. A high rate means reported assistance is common among surveyed performers in this measure; it does not mean AI does most of the task or that most workers perform it. The list comes from O*NET, the U.S. Department of Labor-sponsored occupational information system. The Fed says O*NET contains roughly 2,000 detailed work activities, too many to show each respondent. Instead, the survey displays the ten activities rated most important for that respondent’s occupation and asks which ones they perform. This standardized structure makes questions more comparable across workers in the same occupation. But O*NET is a framework for describing work, not a complete inventory of every duty in a particular workplace. A locally important responsibility, a new activity, or a lower-ranked task may not appear in the ten-item list. The index cannot establish that omitted work is absent from jobs or untouched by AI. “Regularly helps” is a self-report about use. The index does not directly measure assistance intensity, independent task completion, time allocation, output quality, or verification burden. A worker may use AI for one part of an activity and check or revise the result; the rate cannot distinguish that pattern from more extensive assistance. Nor does it show whether a tool was approved, integrated into a formal workflow, or used informally. Those require local evidence. “Measuring what work generative AI does: Survey evidence versus chat logs,” the authors’ CEPR explanation, notes that occupation or task categories with fewer than 20 observations are excluded from the reported comparison. This avoids presenting especially thin groups as stable estimates, but means some detailed categories cannot be compared. A national estimate may also differ from a particular employer, location, team, or current workflow. When you encounter a task value, ask who is counted, what they say they do, and what behavior qualifies as use. Then compare the listed activity with your actual responsibilities. Treat the number as a prompt to investigate a task, not a measure of your personal exposure or a verdict on your role.

Sources: What Work Does Generative AI Do?; Measuring what work generative AI does: Survey evidence versus chat logs; O*NET OnLine Help: About O*NET

What does ‘widespread but shallow’ mean?

“Widespread” describes how many occupation and task categories have reported use; “shallow” describes how far use has penetrated within most categories. The distinction matters because a broad footprint can coexist with most workers not using AI for a particular activity. In What Work Does Generative AI Do?, the St. Louis Fed reports that at least one in five workers uses AI on the job in more than 80% of occupations, and more than 40% of listed tasks have adoption above 20%. Those are category distributions, not a claim that four-fifths of workers use AI, or that AI handles one-fifth of every job. The denominator changes with each statement. The same release shows why broad reach should not be mistaken for saturation. Only 40% of occupations have adoption above 50%; for tasks, fewer than 3% exceed 50%, and none exceed 70%. So reported use appears across a wide range of occupational categories, while use within most individual listed tasks remains below half of the workers who perform them. The working paper, What Work Does Generative AI Do?, summarizes this pattern as adoption across many occupations and tasks, but fewer than half of workers adopting within most of them. Diffusion across categories can be substantial even when the typical task is not routinely assisted by most of its performers. An occupation’s rate is the share of surveyed workers in that occupation who report using AI for their jobs. A task’s rate is the share of people who perform that listed activity and report AI use for it. The second denominator is task performers, not everyone employed, and neither rate says how much time the tool takes over. If many people in a high-adoption occupation use AI, they may be applying it to different parts of their work. The Fed release notes that its highest-adoption occupations are not necessarily driven by a single task everyone delegates. A high occupation rate therefore does not mean one common workflow has been transformed from end to end. The distributions count categories, not the share of all workers, hours, output, or economic value represented by those categories. “More than 40% of tasks,” for example, means that more than 40% of task categories have an adoption rate above the stated threshold. It does not mean AI is used for 40% of tasks in a typical person’s week. Likewise, “more than 80% of occupations” is the proportion of occupation categories crossing the threshold, not the proportion of the U.S. workforce using AI. These are summaries of rates within categories; a worker-weighted total would require a different calculation. The practical conclusion is that reported assistance has spread across categories, with substantial variation among people and tasks within them. The evidence does not establish that a particular occupation has undergone broad workflow redesign, that a named worker is adopting AI, or that a job is likely to disappear. A national average can prompt a closer look at recurring duties, but cannot replace local information about which steps are assisted, how consistently, and under what review. Broad diffusion makes task-level questions relevant in many kinds of work; shallow penetration leaves room for differences in access, choice, and workflow.

Sources: What Work Does Generative AI Do?; What Work Does Generative AI Do?

What do task examples reveal that job labels hide?

Task examples make broad job labels less misleading because they show where reported assistance sits inside a bundle of responsibilities. In the St. Louis Fed’s “What Work Does Generative AI Do?,” the most-assisted listed activities are reading documents to gather technical information (61.3%), preparing research reports (60.7%), and analyzing data to identify trends (57.5%). At the other end, the authors report no AI use for tasks including driving trucks, presenting menus to customers, assisting practitioners with medical procedures, collecting biological specimens from patients, and preparing treatment areas. These rates describe workers who perform a task and report using AI for it; they are not estimates of how much of the task a system can complete. A role rarely consists of one kind of work. A nurse may perform physical duties alongside information gathering, documentation, and planning. The examples do not establish that every nursing workflow is alike or that the occupation is protected; the unit is the listed task, not the person’s whole job. The same point appears in occupations commonly described as manual or interpersonal. The St. Louis Fed reports adoption rates of 75.7% for computer and office machine repairers, 70.9% for special education teachers, and 49.0% for laundry and dry-cleaning workers, each roughly twice what common exposure measures predicted. The authors suggest these jobs also involve information gathering, planning, and communication that AI can assist. The figures do not mean AI performs repair, teaching, or garment care autonomously. They show why a title can conceal tasks workers report doing with assistance. Consider a technical report: the work may include locating source documents, extracting details, drafting a summary, checking figures, interpreting implications, and signing off. High reported use for reading technical documents and preparing research reports makes those steps worth examining. But the survey does not say whether a tool had the right documents, whether its summary was accurate, how much review it needed, or who remained accountable. Assistance at an early step does not establish that later judgment or approval has been automated. Likewise, “no reported use” for a task describes respondents in this survey frame; it does not prove technical impossibility or permanent non-use. A task may be subject to local restrictions or simply not be a useful place to apply available tools. It does not identify every reason for a zero or establish time saved, better output, or staffing change. A worker can apply the evidence by mapping one recurring responsibility: note its inputs, repeatable information step, checks, interaction, and final decision or sign-off. Then ask which step is permitted and verifiable in the actual workplace. This is a practical interpretation of the task-level design, not a separate Fed finding. The evidence supports a limited conclusion: reported assistance appears in information-heavy tasks across varied occupations, including some roles with visible physical or relational work, while some hands-on tasks show little or no reported use. Task-level analysis is therefore more informative than a single label, but it keeps capability, adoption, quality, and employment consequences distinct. Use the examples to inspect one responsibility in context, not to predict that an occupation will shrink, grow, or remain unchanged.

Sources: What Work Does Generative AI Do?

Why does exposure not tell you who is using AI?

An exposure score is a clue about where a capability might apply; an adoption measure records whether people say they are using it. Those are different questions. The St. Louis Fed’s 2026 evidence compares potential applicability with reported assistance across occupations and listed work activities. Exposure helps explain some adoption differences, but does not reliably identify which particular worker uses AI. People doing similar work can report different patterns. The index is more useful for choosing what to inspect than for predicting what an individual will do. That gap is not surprising, but the survey does not tell us which explanation governs every case. A task may look suitable in an exposure framework and still be unavailable because an employer has not approved a tool, information is private, or results would be costly to check. A worker may lack discretion to change a workflow, or find that review takes longer than doing the task directly. These are plausible local conditions, not causes established by the occupation-level comparison. “What Work Does Generative AI Do?” reports the pattern between exposure and adoption; it does not measure these possible barriers for each worker. The distinction also runs the other way. A manual occupation can include information tasks that workers report using AI to support, even when the occupation’s overall exposure signal is modest. A role bundles planning, information lookup, writing, communication, physical work, and responsibility for decisions. A broad exposure score can surface a candidate task, while task-level adoption evidence can reveal reported use within work that the title alone might obscure. Neither measure says that the whole role is automated or that assistance is reliable. The Fed article reports that workers with at least six months of prior AI use tend to report assistance across more tasks. That association is consistent with learning through experience or wider experimentation. It does not show prior use caused broader adoption: people with access, permission, suitable work, or confidence may both use AI longer and across more tasks. Nor does wider use establish better results, time saved, or a career benefit. The observation distinguishes experience from capability; it does not promise that a course or trial will produce the same pattern for everyone. Use the signals in sequence. Exposure can shortlist a recurring task worth examining; reported adoption can show where workers say assistance is occurring. Then check local facts: whether use is allowed, what information can be shared, how to verify output, the cost of error, and who remains accountable. If conditions support a bounded, approved test, consider one. If permission or review ownership is unclear, ask first. If the task is sensitive or checking outweighs the help, make no change for now. A national average cannot settle local questions, and neither exposure nor adoption is a personal job-loss forecast. It helps workers avoid two opposite mistakes: assuming a technically applicable task is already being automated, or overlooking useful assistance because the job title sounds non-digital. Neither conclusion follows from the index alone.

Sources: What Work Does Generative AI Do?; What Work Does Generative AI Do?

Why can chat logs tell a different story?

A platform’s share of conversations assigned to a task is not the share of workers who perform that task and use AI for it. Chat logs start with conversations; the St. Louis Fed’s Real-Time Population Survey starts with workers. It links respondents’ occupations to listed activities, asks which they perform, and asks AI users where it regularly helps. In “Measuring what work generative AI does: Survey evidence versus chat logs,” the researchers explain that platform records generally lack the user’s occupation and people who do not use the platform. A conversation count cannot by itself show how common adoption is among colleagues doing the same task. The denominator problem can make a familiar activity look dominant. Many people may edit text, but only some use AI for editing. A platform share can be high because editing is common, adoption among editors is high, or both. The survey separates task prevalence from adoption among task performers; the authors also construct a comparable share of AI-assisted worker-task pairs. That is more informative than placing an unadjusted chat percentage beside a worker adoption rate. Each statistic can be accurate for its own unit while answering a different question. Classification adds another source of disagreement. A classifier sees a conversation’s wording, not necessarily its purpose in a particular occupation. The CEPR column gives an example: a professor asks for help analyzing data trends. In the professor’s O*NET profile, this belongs to researching topics in the professor’s area of expertise. A classifier without occupational context may assign it to “analyze data to identify trends,” an activity associated with data scientists and analysts. The conversation is real, but its inferred task and occupation can be wrong. The comparison shows this problem is measurable. Across 332 intermediate work activities, correlations between survey task shares and shares from three platform datasets were 0.34 for Anthropic, 0.10 for Microsoft, and 0.11 for OpenAI. The chat sources also agreed modestly with one another, with correlations from 0.08 to 0.38. Grouping activities into nine broader categories raised agreement with Anthropic and OpenAI to about 0.6. This suggests classification detail contributes to disagreement, though it does not erase it. These correlations compare distributions across task categories; they do not measure accuracy for an individual worker or an employment effect. Chat logs still have a useful role: they show platform interactions at scale. Their limitation is context. The researchers conclude that logs can misattribute activity across occupations, while worker-linked survey data can include non-adopters and connect reported use to occupation and task. The survey also has limits, including its selected task list and inability to resolve small occupational groups. They are complements. When estimates differ, compare the population, unit counted, denominator, task taxonomy, and time window. A chat share describes conversations in a platform sample; a worker-task adoption rate describes reported use among people who say they perform a listed activity. Differences may reflect classification, sampling, or real use. Neither number settles whether your own workflow is suitable or approved. Ask a narrower question about your task, then check access, review responsibilities, and what each measure observed.

Sources: Measuring what work generative AI does: Survey evidence versus chat logs; What Work Does Generative AI Do?; What Work Does Generative AI Do?

How should the Fed index be compared with other adoption numbers?

Compare measures by asking what is counted, who is counted, and what the result can establish. The 2026 St. Louis Fed index asks: among people performing a listed activity, what share report regular AI assistance? It links workers to occupations and O*NET activities. This locates reported use within work, but does not measure time saved or organizational rollout. Three lenses answer different questions. Capability or exposure estimates ask where AI might apply, based on task characteristics or capability evaluations. They help shortlist tasks for investigation; they do not show that workers use AI there. “What Work Does Generative AI Do?” reports exposure scores explain roughly half of adoption variation across occupations and tasks, while documenting gaps between predicted exposure and reported use. That is a relationship between measures, not a conversion into adoption or employment effects. Platform chat analyses count conversations classified into activities. They can describe platform activity, but usually omit non-users and the user’s occupation. “Measuring what work generative AI does: Survey evidence versus chat logs” explains that a task’s share of chats combines how common the task is with AI use among its performers. Across 332 intermediate activities, the authors report low correlations between survey-based and chat-based shares, with stronger agreement after grouping activities broadly. Classification and granularity matter; this does not establish labor-market effects. The worker-linked RPS index connects reported behavior to workers and tasks. The 2026 release pools nearly 14,000 U.S. adults ages 18–64 across four quarterly waves from August 2025 through May 2026. Its frame supports occupation comparisons, while selected tasks and self-report limit what it describes. It is the clearest measure here for asking where workers say assistance occurs, not whether assistance is reliable or consequential. Firm surveys use another unit. “Monitoring AI Adoption in the US Economy” compares business, individual, and executive surveys. For late 2025, it reports about 18% of firms adopting AI in the Census survey, about 41% of individuals reporting work-related generative-AI use in the RPS, and an employment-weighted estimate that 78% of the labor force worked at firms reporting AI adoption. These are not estimates of one quantity: the note points to unit, sample distribution, weighting, wording, and respondents’ information. The 78% figure describes workers’ connection to firms reporting adoption, not personal use by 78% of workers. Earlier RPS results provide history, not a clean before-and-after. “The Rapid Adoption of Generative AI” asked recent AI users to select from ten broad activities. Its August 2024 figures describe activities among AI users; the 2026 O*NET measure asks about occupation-linked activities and regular assistance among task performers. Different prompts, taxonomies, dates, and denominators prevent subtracting the figures to claim implementation grew by a measured amount. Match each source to the decision: exposure for potential applicability, chat data for platform activity, worker-task reports for reported assistance, and firm surveys for business adoption or access. The 2026 index improves description of reported use at worker-task level; it does not show how much implementation deepened since the earlier broad survey. Treat the measures as complementary, then investigate the task and workplace before changing a workflow or career plan.

Sources: Monitoring AI Adoption in the US Economy; The Rapid Adoption of Generative AI; What Work Does Generative AI Do?; Measuring what work generative AI does: Survey evidence versus chat logs

What outcome claims remain outside this evidence?

The St. Louis Fed’s 2026 indexes show workers reporting that generative AI regularly helps with listed tasks. They do not show that a system completes those tasks independently, improves the work, saves time after checking, or changes staffing. The Fed presents the indexes as an input for future study of productivity, wages, and employment. Reported use is evidence of behavior in the survey’s U.S. worker and task frame, not proof of employer-wide deployment or role redesign. Several steps separate reported assistance from a labor-market outcome. A tool must be capable of handling a real workflow step; a worker or employer must have permission and access; the result must hold up after verification; and someone must decide how the task is allocated. Even then, effects on hiring or employment depend on what the organization does with any capacity released, what other work remains, and how demand changes. The St. Louis survey does not measure that chain: it does not establish output quality, net productivity, task time, staffing decisions, wages, or an individual’s displacement risk. Reading an adoption percentage as a forecast skips evidence at each link. “Do Job Postings Show Early Labor-Market Effects of AI?” combines an occupational exposure measure with U.S. Lightcast postings from company pages and job boards. Its event study compares posting trends in more- and less-exposed occupations around ChatGPT’s late-2022 release. The authors find a relative decline in postings for more-exposed occupations, but the divergence began before that release, showed no clear additional break afterward, and stabilized after 2023. They say the pattern is difficult to attribute directly to AI adoption. This complicates a simple exposure-equals-displacement story; it neither confirms nor refutes the St. Louis task-use measure. Postings indicate labor demand, exposure estimates potential applicability, and neither directly counts individual job losses. Stronger conclusions need research connecting these stages. Repeated waves with comparable survey questions could show whether reported assistance changes. Measures of task time, output quality, and human review could clarify whether assistance improves the full process. Local implementation evidence could establish which tools are approved and how responsibilities change. Separate analysis of hiring, wages, job redesign, and worker outcomes would be needed to assess labor effects while accounting for other causes. No single indicator substitutes for this chain. For a worker, investigate one recurring task and its local conditions. Ask whether an approved tool can assist a specific step, who checks the output, and who remains accountable. A national rate can suggest where to look, but it cannot settle those questions for your workplace or predict your job. By itself, it is no reason to quit, buy training, or assume your role is safe. Until comparable task, workflow, and labor-outcome evidence is available, use the index to guide proportionate inquiry and skill practice, not as an employment verdict.

Sources: What Work Does Generative AI Do?; Do Job Postings Show Early Labor-Market Effects of AI?

What is a proportionate next action?

Choose one recurring task, then separate its steps: gathering material, producing a first draft or summary, checking facts and numbers, interpreting the result, and approving the final work. In What Work Does Generative AI Do?, the St. Louis Fed’s task index can help flag activities worth examining because it records workers’ reports of regular assistance with listed tasks. It cannot tell you whether your employer permits a tool, whether confidential material may be entered, or whether checking an output would take longer than doing the task yourself. Those are local workflow questions, not conclusions supplied by a national adoption rate. If an approved tool is available and the task is low consequence, try it on a bounded, reversible step. Keep the usual review in place and compare the result with your normal process before expanding use. For sensitive or consequential work, ask who owns approval, what information can be used, and who verifies the result. If those conditions are unclear, document the steps and wait for guidance rather than improvising. The durable move is to frame the task, check sources and outputs, apply domain judgment, and know when to escalate. The free task checker can organize change-pressure signals; it does not estimate a validated probability of displacement. The practical question is: Is assistance approved here, and who verifies it?

Sources: What Work Does Generative AI Do?; O*NET OnLine Help: About O*NET

Questions readers ask