Understanding AI's Labor Market Impacts

Sam Manning December 2025 AI’S LABOR MARKET IMPACTS Opportunities for the Department of Labor’s AI Workforce Research Hub UNDERSTANDING Table of ContentsIntroduction ...................................................................... 1 Background ....................................................................... 3 Part 1: Establish Data-Sharing Partnerships with Leading AI Developers to Gather Data on AI Usage .................................... 5 Part 2: Establish Data-Sharing Partnerships with Payroll and Hiring Platforms to Gather Data on Employment, Wages, and Hiring .......... 10 Part 3: Enhance Federal Data Collection to Measure AI Usage ........ 15

1. Add an AI Usage Supplement to the Current Population Survey ................................................................. 15

2. Retain and Expand Software and AI Expenditure Items in the Annual Integrated Economic Survey ........................ 16

3. Sustain the Business Trends and Outlook Survey AI Supplement ........................................................... 16 Part 4: Build Analysis and Forecasting Capacity through a Voluntary Expert Committee .................................................. 18 Conclusion ........................................................................ 20 There is deep uncertainty about how significant AI’s workforce effects will be, how quickly they will emerge, and which groups they will affect most. Some economists expect only modest effects on employment in the near term, while several leading AI developers1 and other economists2 warn of the possibility of large-scale labor market impacts within one to five years. This uncertainty creates a dual risk: policymakers may be unprepared for labor-market disruption, and talent shortages could slow AI adoption and sap U.S. growth. America’s AI Action Plan recognized this and directed the Department of Labor to establish an AI Workforce Research Hub “to lead a sustained Federal effort to evaluate the impact of AI on the labor market and the experience of the American worker.”3 The pending launch of the Hub offers an opportunity to strengthen how policymakers monitor and understand AI’s workforce impacts. This paper proposes several options for the AI Workforce Research Hub to collect data and enable frequent analyses so that policymakers can be agile in planning for, and responding to, the workforce impact of advanced AI. The Hub could consider the following four-part approach to enhance policymaker visibility into the labor market impacts of advanced AI. 1 Jim VandeHei and Mike Allen, “Behind the Curtain: A White Collar Bloodbath,” Axios, May 28, 2025, https://www. axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic. 2 Lauren Foster and Anton Korinek, “The ‘AI Takeoff’ Is Coming for Office Jobs — and the Future of Work,” UVA Darden School of Business, October 21, 2025, https://ideas.darden.virginia.edu/the-ai-takeoffand-future-of-work. 3 Office of Science and Technology Policy, Office of the President of the United States, Winning the Race: America’s AI Action Plan (2025), www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-ActionPlan.pdf.

Introduction

1. ESTABLISH DATA-SHARING PARTNERSHIPS WITH LEADING AI DEVELOPERS TO GATHER DATA ON AI USAGE. Work with leading AI developers such as OpenAI, Anthropic, Microsoft, xAI, Meta, and Google to establish agreements and standards for sharing anonymized AI usage statistics that can be linked to occupation-level employment, earnings, and hiring statistics.

2. ESTABLISH DATA-SHARING PARTNERSHIPS WITH PAYROLL AND HIRING PLATFORMS TO GATHER DATA ON EMPLOYMENT, WAGES, AND HIRING. Work with firms such as ADP, LinkedIn, Revelio Labs, and Indeed to establish data-sharing agreements for data on employment, wages, hiring trends, and skill demands across the economy, so that these data can be linked to AI usage statistics and made available to researchers.

3. ENHANCE FEDERAL DATA COLLECTION TO MEASURE AI USAGE. Support the Bureau of Labor Statistics and U.S. Census Bureau to add an AI usage supplement to the Current Population Survey (CPS) and AI expenditure items to the Annual Integrated Economic Survey (AIES), while continuing the AI Supplement to the Business Trends and Outlook Survey (BTOS).

4. BUILD ANALYSIS AND FORECASTING CAPACITY THROUGH A VOLUNTARY EXPERT COMMITTEE. Engage leading economists to analyze combined datasets and produce regular forecasts on AI’s workforce impacts. The first three components focus on assembling the underlying data by bringing together information on AI usage, employment, wages, and hiring into a secure and consistent home within the Hub. With improved data on both AI diffusion and key economic indicators, researchers could examine questions that are currently difficult to answer, such as how adoption in particular occupations relates to wages, hiring demand, or job turnover for different categories of workers in those occupations over time. The fourth component would then task a small group of independent experts to analyze these linked datasets and produce regular reports for the Department of Labor, other relevant agencies, congressional committees, and the public. This last component would ensure that emerging workforce effects are monitored and interpreted consistently over time. Background Given how widely expert views diverge and how rapidly some expect transformative changes to occur, reducing uncertainty about AI’s labor market impacts would be a particularly valuable goal for the Hub. Better visibility into near-term labor market effects could support smarter, more agile policy design across areas such as education, workforce training, and social safety net reform while helping identify where AI is generating new growth opportunities. By generating empirical insights into what labor market effects are unfolding now and what may lie ahead, the Hub can play a pivotal role in shaping effective policy responses that accelerate the economic benefits of AI, boosting worker resilience to AI’s disruptive effects, and dampening unjustified fears. A key question for the Department of Labor is how best to deliver on the Hub’s mission of evaluating the impact of AI on the labor market. The Hub is not the only potential home for this work, and different parts of the effort could, in principle, be carried out by different institutions. The Hub’s central role would be to coordinate access to AI-usage data, payroll data, hiring data, and economic data from key government surveys so that these sources can be linked and made securely available to researchers. Academic institutions are well positioned to lead much of the empirical analysis once such data are accessible. Non-governmental organizations such as the Frontier Model Forum could, in theory, help facilitate parts of the data-sharing process with AI developers. Institutions such as the National Bureau of Economic Research could assist with access to private-sector economic data, and international bodies such as the World Bank could support comparative work that links AI usage and economic data across countries. For the purposes of informing U.S. policy, however, each of these institutions would benefit from close partnership with agencies such as the Department of Labor, the Census Bureau, and the Federal Reserve in order to integrate private-sector data with core economic statistics. Housing the initiative within the Department of Labor would offer a more consistent and coordinated approach than relying on multiple independent organizations to manage different parts of the effort. The Hub therefore provides a potentially promising venue for assembling the data infrastructure while enabling external researchers to generate the analysis needed to guide policy. While much of the necessary data already exists in some fashion, it is either incomplete or inaccessible to researchers and relevant government agencies. For example, AI developers (such as Anthropic and OpenAI), private payroll services (such as ADP), and job-listing sites (such as LinkedIn) already have data on AI usage, employment, and skill demands that are more timely, comprehensive, and informative than government-collected data on its own. By developing partnerships with the private sector, making simple changes to ongoing government data collection, and ensuring that the Hub can work with independent experts to conduct recurring analysis, the Hub can turn fragmented data into policy-relevant insights while avoiding major new spending on government data collection. Part 1: Establish Data-Sharing Partnerships with Leading AI Developers to Gather Data on AI Usage To understand how AI is affecting the workforce, one must know where and how AI is being used by workers and firms across the economy. Currently, the best sources of data on AI usage are the leading AI developers themselves. These companies—for example, OpenAI, Anthropic, and Google—deploy products such as ChatGPT, Claude, and Gemini, and serve APIs that other developers build software and automation processes on top of. One option for the AI Workforce Research Hub would be to coordinate with all leading AI developers to establish secure data-sharing arrangements that provide insights into how AI is being used by workers and businesses in the United States. The specific data collected would need to balance two aims: first, to enable access to information that is genuinely useful beyond what existing data collection efforts can reveal; and second, to do so in ways that respect user privacy and companies’ legal obligations to their customers. The most appropriate structure for such data partnerships would need to be developed collaboratively between the Hub and the major AI developers. A natural starting point might be an industry-wide analogue to Anthropic’s Economic Index, which reports anonymized breakdowns of the worker tasks for which Claude is used. If other AI companies adopted a comparable reporting format and shared data on a recurring basis, this could enable consistent, high-frequency tracking of how frontier AI systems are used by individuals. Linking such usage data to economic indicators (such as employment, hiring, and job posting data) could, in turn, help clarify where AI is creating demand for new skills and where it may be contributing to faster job turnover or slower hiring. 4 Ruth Appel et al. Anthropic Economic Index Report: Uneven Geographic and Enterprise AI Adoption, Anthropic (2025), https://www.anthropic.com/research/anthropic-economic-index-september-2025-report. At present, policymakers and researchers have access to some AI usage data, but existing sources provide only a partial and sometimes inconsistent picture. Available information comes from academic surveys,5 firm-level surveys conducted by the Census Bureau,6 data aggregation efforts such as WildChat,7 and publications by three of the leading AI developers: OpenAI,8 Anthropic,9 and Microsoft. Together, these sources suggest that AI adoption varies widely across industries. For instance, approximately 18 percent of businesses in the Information sector report using AI, compared with roughly 1.4 percent in Construction and Agriculture. The research that Anthropic and OpenAI released indicates that generative models are used most intensively for writing-related tasks, while API usage tends to align more closely with automation than with collaborative use between humans and AI. Other research also finds that AI is diffusing faster than earlier general-purpose technologies such as the internet or personal computers. Despite the data that are available to date, important uncertainties remain about how intensively AI is being used across the economy. For instance, a Census Bureau survey in late 2024 found that 6.6 percent of non-farm businesses reported using AI,13 while a separate nationally representative individual-level survey found that 23 percent of employed respondents had used AI for work during the previous week. While such surveys can provide binary measures of whether firms or workers report using AI and can also sometimes capture the kinds of tasks AI is being applied to, they offer limited 5 See, e.g., Alexander Bick et al., “The Rapid Adoption of Generative AI,” NBER Working Paper No. 32966 (September 2024), https://www.nber.org/papers/w32966; Jonathan Hartley et al., “The Labor Market Effects of Generative Artificial Intelligence,” SSRN, December 18, 2024, https://doi.org/10.2139/ssrn.5136877. 6 Kathryn Bonney et al., “Tracking Firm Use of AI in Real Time: A Snapshot from the Business Trends and Outlook Survey,” NBER Working Paper No. 32319 (April 2024), https://www.nber.org/papers/w32319. 7 Wenting Zhao et al., “WildChat: 1M ChatGPT Interaction Logs in the Wild,” arXiv (May 2, 2024), https://doi. org/10.48550/arXiv.2405.01470. 8 Aaron Chatterji et al., “How People Use ChatGPT,” NBER Working Paper No. 34255 (September 2025), https:// www.nber.org/papers/w34255. 9 Appel et al., Anthropic Economic Index Report. 10 Kiran Tomlinson et al., “Working with AI: Measuring the Applicability of Generative AI to Occupations,” arXiv (last revised October 17, 2025), https://doi.org/10.48550/arXiv.2507.07935. 11 Bonney et al., “Tracking Firm Use of AI in Real Time.” 12 Bick et al., AI Diffusion Report: Where AI Is Most Used, Developed, and Built, Microsoft (2025), https://www. microsoft.com/en-us/research/wp-content/uploads/2025/10/Microsoft-AI-DiffusionReport.pdf. 13 Bonney et al., “Tracking Firm Use of AI in Real Time.” 14 Bick et al., “The Rapid Adoption of Generative AI.” visibility into the intensity of use—such as how much compute, API activity, or token volume is being devoted to different applications across industries, firm sizes, or geographies. API providers and enterprise software providers hold the most comprehensive data on these usage patterns. Understanding this dimension of AI adoption is particularly important if, as some economists have modeled, compute may become a more direct substitute for human labor as AI capabilities advance. In that case, measuring how intensively AI systems are used for different purposes could provide early evidence of where automation pressures are emerging most strongly. So far, Anthropic is the only frontier AI developer that consistently releases usage data for research purposes. Its anonymized statistics, published through the Economic Index, represent an important early contribution. However, relying on any single company’s data to understand AI’s effects across the American workforce will provide an incomplete and misleading view. The available evidence already indicates substantial variation in how different frontier models are used: around one-third of Claude’s work-related interactions involve coding, compared with only 4 percent of ChatGPT messages. Similar discrepancies appear when comparing “AI exposure” measures with real-world usage data. These measures attempt to identify which worker tasks are most likely to be affected by LLMs—for example, because they can help a worker complete the task more efficiently. These estimates correlate strongly (90 percent) with Microsoft Copilot usage17 and with some national survey results on AI adoption (67 percent)18 but less so with Anthropic Economic Index data. These inconsistencies suggest that a broader, multi-company data-sharing effort would provide a more accurate picture of AI’s labor-market footprint. 15 Pascual Restrepo, “We Won’t Be Missed: Work and Growth in the AGI World,” NBER Working Paper No. 34423 (October 2025), https://www.nber.org/papers/w34423. 16 Tyna Eloundou et al., “GPTs Are GPTs: Labor Market Impact Potential of LLMs,” Science vol. 384, no. 6702 (June 20, 2024): p. 1306–1308, https://doi.org/10.1126/science.adj0998. 17 Tomlinson et al., “Working with AI.” 18 Bick et al., “The Rapid Adoption of Generative AI.” 19 Kunal Handa et al., “Which Economic Tasks Are Performed with AI? Evidence from Millions of Claude Conversations,” arXiv (March 7, 2025), https://doi.org/10.48550/arXiv.2503.04761. To support evidence-based policymaking, the Hub could aim to facilitate access to data that are comprehensive (i.e., from all major frontier AI developers), shared on a recurring basis (e.g., quarterly), standardized across providers, and linkable in aggregate to employment and wage information at the occupation level, as well as to economic statistics broken down by industry and firm size. While today’s mix of survey data and voluntary disclosures provides a starting point, coordinated partnerships with AI developers— building on initiatives such as Anthropic’s Economic Index—would offer a more systematic foundation for understanding how advanced AI is diffusing across the U.S. economy. There are several open questions about how usage data-sharing agreements could be structured in practice. One uncertainty concerns the degree of standardization that would be feasible across companies, including how frequently data should be shared, how usage should be sampled from developer user bases, and what methodological approach is used to map prompts or API calls onto work tasks or other occupational taxonomies. Another open question involves choosing a common way to classify usage, such as by mapping usage data to O*NET Intermediate Work Activities (as Microsoft and OpenAI have done in their publications) or to O*NET task statements (as Anthropic has done).20 Aligning on a shared approach can ensure that usage data from different models can be compared in a consistent way. A further design choice relates to methods for categorizing usage—for example, distinguishing between automation-oriented and augmentation-oriented behavior based on the nature of the input and output data from a given human-AI interaction. A separate uncertainty concerns what data AI developers actually capture and are contractually permitted to share. In some cases, usage routed through cloud providers such as AWS, Azure, or Google Cloud never reaches the model developer at all, and cloud providers themselves may not always retain user prompts or logs. For enterprise 20 O*NET structures work content at multiple levels, including Intermediate Work Activities (IWAs) and more granular task statements. See “The O*NET Content Model,” O*NET Resource Center, accessed December 4, 2025, https://www.onetcenter.org/content.html. 21 For example, see the data protection details for Amazon’s Bedrock cloud service. “Data Protection,” Amazon Web Services, accessed December 4, 2025, https://docs.aws.amazon.com/bedrock/latest/userguide/ data-protection.html. customers, contractual terms may also restrict what can be analyzed or retained. As a result, the observable usage available for economic research may represent only a subset of total activity, and it may not be clear how representative that subset is. Clarifying these constraints—including how much usage is visible, analyzable, and shareable— would be important for interpreting any resulting statistics and for understanding how much unmeasured usage may remain. Questions also remain about the appropriate level of aggregation for any statistics that are shared, and what additional demographic, geographic, or firm statistics could be provided alongside usage data while still adequately protecting user privacy. Finally, documentation and reproducibility standards would need to be clarified so that researchers and Department of Labor officials can see how usage statistics were constructed and can confirm that results are comparable across companies. These outstanding questions would need to be resolved collaboratively between the Hub and AI developers. Clarifying them early in the process could help establish a transparent and durable framework for usage-data reporting. Part 2: Establish Data-Sharing Partnerships with Payroll and Hiring Platforms to Gather Data on Employment, Wages, and Hiring Usage data are only useful for assessing AI’s effects on the workforce if they can be linked to economic indicators. Yet government data on employment and wages have meaningful limitations, preventing policymakers from obtaining a current picture of how AI is changing job opportunities across occupations and for workers at different career stages. However, private-sector data can fill key gaps at a much lower cost to taxpayers than expanding government data collection. Therefore, a second opportunity for the AI Workforce Research Hub would be to form data-sharing agreements with large-scale payroll providers and hiring platforms that capture changes in employment, wages, and hiring patterns with greater precision than government sources alone. Combined with data-sharing arrangements with leading AI developers, these partnerships would position the Hub as a central clearinghouse for the economic and usage data needed to understand AI’s workforce impact at the occupation and firm level, and for subgroups of workers within occupations. The Hub could then make this linked, anonymized data publicly available, while also providing access to a trusted set of independent researchers that compose an expert analysis body (see Part 4), enabling timely analysis that can inform policymaking. Publicly available labor data on employment, wages, and hiring are valuable but have important limitations for studying AI’s short-term and fine-grained effects. The Current Population Survey (CPS), for instance, surveys around 60,000 households each month, yielding roughly 45,000 to 50,000 employed individuals. This sample size is sufficient for tracking broad national trends,22 but it becomes far less informative when research22 Martha Gimbel et al., “Evaluating the Impact of AI on the Labor Market: Current State of Affairs,” Budget Lab, Yale University (October 2025), https://budgetlab.yale.edu/research/ ers try to analyze smaller groups, such as recent college graduates within specific occupations or in AI-exposed industries. In practice, only a handful of respondents fall into these sub-categories (e.g., entry-level software engineers, or customer service representatives aged 54+) in a given month, which makes it difficult to detect meaningful changes to groups of interest over time. Because the CPS repeatedly surveys the same individuals and selects participants from geographically clustered areas, its effective sample size is even smaller. The result is that analyses of AI’s labor market impact based on CPS data often become noisy or inconclusive, especially when focusing on smaller or emerging workforce segments. In response to these data gaps, one option would be to expand government surveying and data collection of key economic indicators at the individual level. A recent letter signed by dozens of economists and AI experts calls for an expansion of the CPS’s sample size to address this challenge. Doing so would be commendable, and likely a worthy use of public resources in the long run. However, the data required to understand these patterns already largely exist in the private sector, collected continuously through payroll systems, hiring platforms, and enterprise software used by millions of workers and firms. Some data providers, such as Revelio Labs, combine data from several private-sector sources and via web scraping. Leveraging these existing data sources through public-private partnerships could be both a faster option and one that is far less reliant on taxpayer funding to implement, compared to an overhaul of government surveys. For context, the CPS is already undergoing a multi-year modernization process aimed at improving response rates. The project is expected to take roughly five years to complete, a timeline that will be far too slow to inform the policy decisions that need to be made over the next one or two years as AI integration across the U.S. economy accelerates. evaluating-impact-ai-labor-market-current-state-affairs. 23 Letter to Secretary Lori Chavez-DeRemer: We Urge the Department of Labor to Improve and Prioritize the Collection of High Quality and Timely Labor Data Monitoring AI’s Impact, Americans for Responsible Innovation, September 9, 2025, https://ari.us/wp-content/uploads/2025/09/Letter-Leading-Economists-Call-forMore-Data-on-AI-and-Jobs.pdf. 24 “Current Population Survey (CPS) 2023 Modernization Efforts,” Census Bureau, last revised May 13, 2024, https://www.census.gov/programs-surveys/cps/about/modernization.html. Payroll processors such as ADP, Paychex, Rippling, Intuit, and Gusto collectively track tens of millions of American workers and can provide near real-time insight into employment and wage dynamics across occupations and industries. These datasets allow researchers to detect shifts that are invisible in government surveys—for example, subtle changes in hiring, retention, or wage growth within AI-exposed sectors, or for workers with particular education backgrounds. Hiring platform data offer a complementary perspective: they capture employer demand for different skills and roles before these changes appear in official statistics. Job postings and résumé data from platforms such as LinkedIn, Indeed, and Revelio Labs can reveal which skills are rising or declining in value, which occupations are being redefined by AI tools, and how firms are reorganizing work as new technologies diffuse. Recent research illustrates the potential of these datasets. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen found early signs of declining employment in AI-exposed occupations using ADP payroll data; patterns that were not detected using CPS data because of small sample sizes. Seyed Mahdi Hosseini Maasoum and Guy Lichtinger combined LinkedIn’s résumé data with Revelio Labs’s job postings data to track more than 280,000 firms and 62 million workers in the United States—a sample size large enough to allow for analysis of AI’s impact separately for senior and junior workers in different occupations. They found that firms creating “GenAI Integrator” roles (positions aimed at embedding AI into business processes) experienced roughly a 9 percent decline in junior employment, driven by slower hiring rather than layoffs, while senior employment remained stable. Together, these studies show how large-scale private employment and hiring datasets can reveal structural shifts that official government statistics mask. As evidenced 25 Erik Brynjolfsson et al., “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab Working Paper (November 2025), https://digitaleconomy. stanford.edu/wp-content/uploads/2025/08/Canaries_BrynjolfssonChandarChen.pdf. For examples using CPS data, see Bharat Chandar, “Tracking Employment Changes in AI-Exposed Jobs,” SSRN, (August 12, 2025), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5384519; Gimbel et al., “Evaluating the Impact of AI on the Labor Market.” 26 Seyed Mahdi Hosseini Maasoum and Guy Lichtinger, “Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data,” Wharton Research Data Services Research Paper Series (August 31, 2025), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555. by these examples, some individual researchers have already begun building ad hoc partnerships with private-sector data providers. However, forming data-sharing agreements with private companies can be an opaque challenge for many individual researchers who might otherwise benefit from data access. Depending on the provider, there are not always clear pathways or frameworks for establishing such agreements, and even when it is possible, the process of securing ongoing access can be costly and time consuming. As a result, important analyses may depend on who happens to secure access rather than on the quality of research questions being asked. The Hub could reduce these barriers by establishing standing data-sharing agreements with major payroll and hiring platforms. In doing so, it could make aggregated, anonymized data more broadly available—either by licensing public-use versions of the data or by providing secure access to a vetted group of independent researchers. This would enable a wider range of researchers to study AI’s workforce impacts, foster healthy competition among analytical approaches, and ensure that insights are produced regularly and consistently, rather than sporadically. Furthermore, most data-use agreements between academics and private companies restrict access to the immediate research team. This prevents other researchers from reproducing findings or building directly on the underlying data, which limits scrutiny and weakens the evidentiary base available to policymakers. Improving reproducibility and transparency in this domain is particularly important given the scale of forthcoming policy decisions about AI’s workforce effects. House Office of Science and Technology Policy Director Michael Kratsios recently emphasized the need to return to “Gold Standard Science,” highlighting reproducibility and open validation as cornerstones of trustworthy evidence. Establishing shared data access pathways through the Hub would help move research on AI’s labor impacts closer to these principles by enabling more systematic validation, replication, and methodological improvement. 27 Michael Kratsios, “Remarks by Director Kratsios at the National Academy of Sciences,” The White House (May 19, 2025), https://www.whitehouse.gov/briefings-statements/2025/05/remarks-by-directorkratsios-at-the-national-academy-of-sciences/. Despite its potential, there are also important uncertainties about how the Hub could implement this approach. One question is whether the Hub will have sufficient technical capacity and infrastructure to clean, store, standardize, and link large-scale private-sector datasets with AI-usage data at the occupation and industry level. Payroll and job site datasets arrive in varied formats, often with inconsistently coded fields, and linking them reliably to usage data may require significant engineering resources. Another uncertainty concerns the durability of private-sector data access: company policies and data-licensing rules can shift over time, and future access would depend on the long-term stability of partnership agreements. One way to simplify early implementation would be to begin with firms that already aggregate data from multiple sources, such as Revelio Labs or Burning Glass, which could reduce the number of separate agreements needed at the outset as well as the technical data processing burden. Part 3: Enhance Federal Data Collection to Measure AI Usage While partnerships with private-sector data providers could help make critical data analyses available to policymakers, modest updates to existing government surveys would ensure that policymakers retain transparent, long-term, and publicly accessible measures of AI adoption and its economic effects. Three targeted adjustments—adding an AI usage supplement to the CPS, maintaining and expanding AI-related investment coverage in the Annual Integrated Economic Survey (AIES), and continuing the AI Supplement to the Business Trends and Outlook Survey (BTOS)—would strengthen the public statistical foundation for assessing AI’s workforce impact. CPS is run by the Bureau of Labor Statistics, while AIES and BTOS are administered by the Census Bureau. The Hub’s role would primarily involve supporting these agencies in implementing such adjustments, rather than administering changes directly.

1. Add an AI Usage Supplement to the Current Population Survey The CPS remains the nation’s primary household survey for employment and demographic statistics. By adding a short, recurring module on AI use at work, the Bureau of Labor Statistics could capture how workers are adopting AI tools, how frequently they use them, and whether employers provide related training. Because CPS responses are linked to occupation, industry, and demographic data, even a few targeted questions, such as those tested in the Real-Time Population Survey,28 could provide a nationally representative record of how AI use diffuses across the workforce. Implementing a brief annual supplement would be a relatively low-cost way to fill a major gap in understanding how worker experiences and skill demands evolve as AI tools become more integrated into everyday work. 28 Alexander Bick and Adam Blandin, Real-Time Population Survey (RPS) (project homepage), https://sites. google.com/view/covid-rps/home.

2. Retain and Expand Software and AI Expenditure Items in the Annual Integrated Economic Survey At the firm level, the AIES offers an opportunity to systematically track how businesses invest in AI tools and related technologies. As noted in recent public comments to the Census Bureau submitted by Sophia Brown-Heidenreich, the Foundation for American Innovation’s director of artificial intelligence policy, preserving a line item for “software expenditures” and adding a specific subcategory for “AI-related software and services” would help quantify how AI investment fits into firms’ broader spending patterns. These data would complement AI usage measures from developers (Part 1) and workforce outcomes from payroll and job site data (Part 2), allowing researchers to link AI investment, adoption, and employment effects together. Tracking these expenditures would also help policymakers identify whether AI spending is concentrated among certain industries, firm sizes, or regions, which may be essential for anticipating uneven diffusion, productivity effects, and emerging labor needs in high-growth sectors.

3. Sustain the Business Trends and Outlook Survey AI Supplement The BTOS is an experimental biweekly survey that reaches roughly 1.2 million employer businesses. It is currently the most timely and detailed government source of firm-level data on AI adoption in the U.S. Its AI Supplement, first run from September 2023 to February 2024, provided the first estimates of how businesses are using AI, which types of AI systems they employ, and whether those systems are affecting employment or task organization within firms.31 29 Sophia Brown-Heidenreich, “Comments on Revision to the Annual Integrated Economic Survey,” Foundation for American Innovation, October 21, 2025, https://www.thefai.org/posts/comments-on-revision-tothe-annual-integrated-economic-survey. 30 “Business Trends and Outlook Survey (BTOS) Data,” Census Bureau, last revised April 11, 2025, https://www. census.gov/data/experimental-data-products/business-trends-and-outlook-survey. html. 31 See Bonney et al., “Tracking Firm Use of AI in Real Time.” Given its demonstrated value, the Department of Labor could work with the Census Bureau and the Office of Management and Budget to ensure that the AI Supplement is approved to continue on a regular schedule for the next three to five years, after which it could be re-evaluated. Maintaining the supplement would ensure continuity in tracking firm-level AI adoption at high frequency and allow researchers to connect these business trends with workforce outcomes. Together, these three survey improvements would create a coherent, multi-level data infrastructure for understanding AI’s economic effects. The CPS would capture which workers use AI and how, the AIES would measure how firms invest in it, and the BTOS would provide ongoing data on how AI adoption affects business operations and employment. This public data backbone would complement private-sector datasets, which may be richer and more granular but potentially less durably available over time, by ensuring a continuous record of AI diffusion across the economy. Maintaining this foundation could help policymakers and researchers monitor long-term trends and make evidence-based decisions as AI continues to reshape the labor market. Part 4: Build Analysis and Forecasting Capacity through a Voluntary Expert Committee The final component of the AI Workforce Research Hub could be to convene a voluntary, nonpartisan committee of independent economists, AI experts, and labor-market researchers to analyze the linked datasets described above and produce recurring assessments of AI’s workforce impact. This committee would serve as a focused, low-cost mechanism for regular analysis and forecasting, narrower in scope and more focused on delivering analysis than broad advisory bodies such as the Federal Economic Statistics Advisory Committee, but dedicated specifically to understanding how AI adoption is influencing jobs, skills, and economic outcomes for American workers. If necessary, committee members could be vetted through a process similar to Special Sworn Status at the Census Bureau, granting them secure access to anonymized data collected through the Hub’s partnerships and federal surveys. Access to federal survey data could be managed through existing secure infrastructures such as Research Data Centers34 or through existing remote access programs,35 while privately sourced datasets may need to be governed by new data-use agreements and hosted in a comparable secure environment that maintains privacy and confidentiality protections. However, public-use data made available by the Hub might be sufficient for regular reporting of key trends, in which case special access to sensitive microdata might not be necessary. 32 “Federal Economic Statistics Advisory Committee (FESAC),” U.S. Bureau of Economic Analysis, accessed December 4, 2025, https://apps.bea.gov/fesac. 33 “Restricted-Use Data Application Process,” Census Bureau, last revised December 19, 2024, https:// www.census.gov/topics/research/guidance/restricted-use-microdata/standardapplication-process.html. 34 “Federal Statistical Research Data Centers,” Census Bureau, last revised September 19, 2025, https://www. census.gov/about/adrm/fsrdc.html. 35 “Secure Remote Research Environment,” Census Bureau, last revised April 9, 2025, https://www.census.gov/ about/adrm/fsrdc/about/secure-remote-access.html. The value of the expert committee would be that it would be specifically tasked with delivering quarterly reports tracking how AI usage and investment are affecting different workers and sectors. This would ensure consistent reporting on the trends most critical to key national policy decisions. These analyses could include near-term forecasts of job transitions rates, labor shortages, wage dynamics, changing skill demands, and emerging labor-market risks. The committee could also identify data gaps and suggest improvements to government and private-sector data collection efforts, ensuring that the Hub’s analytical capacity continues to evolve as AI advances. And by relying on voluntary participation from qualified experts and adapting existing data-access infrastructure, this approach would minimize administrative overhead and taxpayer costs.

Conclusion

The Department of Labor’s AI Workforce Research Hub represents an opportunity to build a strong empirical foundation for understanding how advanced AI systems are reshaping work. By combining public data collection with data sharing partnerships across the private sector, the Hub could transform fragmented information into a coherent, near-real-time picture of AI’s labor market effects. The four components outlined in this paper—data sharing with AI developers, partnerships with payroll and hiring platforms, targeted updates to existing federal surveys, and a voluntary expert analysis body—offer a practical framework for achieving that goal. Together, they would enable policymakers to anticipate where new skill demands are emerging, where labor market disruption may occur, and which workers or industries are most likely to be affected positively and negatively. If it can effectively fulfill its mandate to sustain a “Federal effort to evaluate the impact of AI on the labor market and the experience of the American worker,” the Hub could help ensure that federal policy is guided by timely, evidence-based insights rather than speculation. In doing so, it could support an approach to AI governance that expands economic opportunity, strengthens national competitiveness, and promotes resilience for American workers. About the AuthorThe Foundation for American Innovation champions the technology, talent, and ideas essential to American prosperity, security, and flourishing. FAI was founded with a unique vision: Advance a more perfect union between technology and the American republic. We are a collection of builders, hackers, policy wonks, and founders; our work is rooted in an optimistic vision for the future, one in which our technologies align to serve human ends–strong institutions, economic prosperity, and robust national security–not the other way around. thefai.org Sam Manning Sam Manning is a non-resident fellow at the Foundation for American Innovation and a senior research fellow at the Centre for the Governance of AI (GovAI). His work focuses on measuring the economic and labor market impacts of frontier AI systems and designing policy options to help ensure that advanced AI can foster widespread economic prosperity.

Acknowledgments

The author would like to thank Rishi Bommasani, Chris Meserole, Michael Lachanski, Markus Anderljung, Bharat Chandar, and Tomás Aguirre for helpful conversations and comments.