AI Labor Market Tracker: July 2026

Employment and headcount. Revelio Labs measures employment using hundreds of millions of public professional profiles. Individual employment histories are linked to employers, occupations, industries, locations, education, and skills to construct a longitudinal workforce panel. Throughout this report, employment is measured as the weighted number of workers employed by an organization or within a labor market in a given month.

Two adjustments are central to interpreting these figures. First, raw online profiles do not represent a random sample of the workforce: white-collar occupations and workers in large cities are more likely to appear than blue-collar occupations or workers in smaller places. To correct this, every observation is assigned a sampling weight that scales it up to the true population, calibrated against government labor statistics by occupation and country. A software engineer with a 90% chance of appearing online counts for roughly 1.1 workers; a nurse with a 25% chance counts for four. Reported employment therefore reflects population estimates, not raw profile counts.

Second, people are slow to update their profiles when they change or lose jobs, which causes the most recent months to understate hiring and separations. We correct for this reporting lag with a nowcasting model that estimates the flows that will eventually be revealed once profiles catch up, drawing on how past under-reporting resolved, seasonal patterns, and company-specific events such as layoffs. This matters especially at the end of the sample, where the AI signal is strongest and the raw data is least complete.

Job postings. Job postings come from Revelio's unified, deduplicated postings data, which integrates listings from employer career pages, job boards, and staffing firms. Because a single opening can appear across many channels in different formats, near-duplicates are collapsed into a single posting. Each posting is mapped into a common taxonomy of employers, occupations, seniority levels, skills, and work activities. Postings measure hiring demand rather than existing employment, and so serve as an early indicator of changes in labor demand. U.S. job postings exclude regional-aggregator-only listings.

Employee sentiment. Employee sentiment is measured from public employee reviews, with each review's text split into positive and negative components and mapped to occupation, seniority, and geography. Sentiment measures reflect employees' reported workplace experiences rather than employer communications.

Layoff notices. Layoff activity is measured using Worker Adjustment and Retraining Notification (WARN) notices filed with U.S. state labor agencies. WARN notices provide advance notification of qualifying mass layoffs and plant closures, and therefore capture announced layoffs rather than realized employment changes.

Unless otherwise noted, all analyses draw on data through June 2026. Job postings begin in January 2021; workforce, sentiment, and layoff histories extend further back.

AI exposure. Revelio Labs measures AI exposure at the activity level rather than the occupation level. Using our proprietary activity taxonomy, each work activity is assigned an AI exposure score between 0 and 1 by a large language model that evaluates how much of the activity current generative AI can perform. Occupation-level exposure is the aggregate of the exposure of the activities that occupation performs. For the difference-in-differences results, occupations are sorted into exposure quintiles fixed on pre-ChatGPT postings, before November 2022, to avoid endogeneity.

Because this measure is proprietary, we benchmark it against two established external measures — GPTs-are-GPTs (Eloundou et al. 2023) and Felten et al. (2023) — which correlate strongly with ours at the occupation level. As a robustness check, we re-estimate the main headcount results using the GPTs-are-GPTs measure in place of our own and find the findings qualitatively unchanged.

Each external measure is aggregated up to the four-digit occupation level used throughout this report, weighting by posting volume where a broader occupation spans several more detailed ones. Figure A1.1 reports the pairwise correlation between all exposure measures at this level; because the comparison is symmetric, only the lower half of the grid is shown. These are occupation-level correlations, not agreement at the level of an individual job posting, worker, or firm.

AI adoption. AI adoption is measured following the same method as Hosseini Maasoum and Lichtinger (2025), which identifies adopting firms from observed employer behavior rather than stated intentions, so that adoption is measured consistently across companies and over time.

AI integrator adoption. Firm-level adoption is measured from job postings for roles that build or operate generative-AI systems. Candidate postings are first identified using terms associated with generative AI and LLM implementation, including retrieval-augmented generation or RAG, embeddings and vector databases, agents, LangChain, LlamaIndex, fine-tuning, model serving, evaluation, guardrails, and API integration, as well as relevant vendor and product names. A classifier then distinguishes postings for employees who build or operate these systems from postings for employees who merely use tools such as ChatGPT or Copilot. Only the integrator classification counts toward adoption.

Adoption timing. A firm's adoption month is the month of its earliest posting classified as an integrator role. The measure captures a revealed hiring signal associated with AI implementation; it does not establish that a system was already deployed successfully or identify the amount invested in AI.

Figure 3.5 uses a rolling panel of eligible U.S. hiring firms. The panel is rebuilt for each release using the preceding 48 complete months. A firm enters the panel if it records at least 20 new U.S. positions and has matched U.S. job postings during that window. Firms in the top 1% of postings-to-new-position ratios are excluded to limit the influence of unusually noisy posting matches. The denominator is fixed within a release, although firms may enter or leave when the panel is rebuilt for a later release. The cumulative share therefore measures integrator adoption within this eligible-firm panel, not among all U.S. firms.

AI workers. AI workers are employees in AI-related occupations, identified using Revelio's occupation taxonomy together with a keyword-based classification of AI-related job titles.

On the supply side, we track AI-related education, both degree and non-degree credentials. We measure the share of workers entering each broad field of study for degree credentials over time. We also assess non-degree certification trends related to AI, and their composition. Both are computed from the dated education and credential records on worker profiles. Undergraduate field shares are drawn from bachelor's-degree records at a fixed set of schools, classified into broad fields of study; graduation year is shifted back four years to approximate when a student started their degree, and each field is indexed to its estimated 2022 starting-year share. Because the school sample is fixed, these trends describe that sample rather than all U.S. college graduates.

To measure the impact of AI on employment, we compare occupations in the top and bottom AI-exposure quintiles before and after the release of ChatGPT in November 2022. The reference month we use is October 2022. This design follows the event-study approach of Brynjolfsson, Chandar and Chen (2025), applied both to job postings and headcount. Job postings are used as a proxy for hiring intentions, while headcount shows the realized hiring. We report log-point estimates as percentage changes.

Headcount and postings. For headcount and postings outcomes, we estimate the event study with a two-way fixed effects regression of the log outcome on event-time indicators interacted with treatment status. Occupation fixed effects absorb time-invariant differences across occupations and month fixed effects absorb common macroeconomic and seasonal shocks. Standard errors are clustered at the occupation level. A 12-month centered rolling average is applied before estimation to reduce noise. For postings by seniority, the specification is run separately for junior and senior postings, excluding entry-level and manager roles.

AI vs. non-AI role employment. Headcount is indexed to November 2022 for AI-designated roles versus all other roles, tracking their divergence since ChatGPT's launch.

Overall adoption headcount gap. Firms are marked as adopters once they post an integrator-level AI job — see adoption timing above. Each month, total headcount is compared between adopters and non-adopters after removing that month's average level, and the series shows the change in this adopter/non-adopter gap relative to October 2022, expressed as a percentage. Confidence intervals combine firm-level variance in the current month with variance in the October 2022 reference month. The specification absorbs common monthly shocks but does not control for industry, so it describes the change in an adjusted cross-sectional gap rather than a causal, within-firm estimate.

Adoption headcount gap by seniority. The same adopter/non-adopter comparison, estimated separately for junior and senior employees each month. The series is indexed to October 2022 and shown as a percentage change in the adjusted gap, with confidence intervals built from firm-level variance in the current and reference months.

WARN layoffs. Because mass layoffs are sparse events at the company level, the fixed effects specification is unreliable. We instead use a bootstrap difference-in-differences: layoff rates are aggregated to quintile-month totals normalized by the number of companies, converted to log rates, normalized to the reference period, and differenced between the top and bottom exposure quintiles. Company exposure is averaged over the ten months preceding ChatGPT's launch, and confidence intervals use 200 bootstrap iterations that resample full company time series to preserve serial correlation.

The wage result estimates how advertised salaries in job postings move with AI exposure, holding the effective size of each occupation fixed at its true employment share. Salaries are winsorized at the top and bottom 1% and log-transformed so effects read as percentages, and exposure is standardized so that one unit always means one standard deviation. Each posting is weighted by its occupation's overall headcount relative to how many postings that occupation generated that month, so that occupations which advertise heavily do not dominate the estimate. Fitting this month by month yields the headcount-weighted salary premium per standard-deviation increase in AI exposure, using the full continuous exposure score rather than a high/low split.

To measure how AI is changing work, we track how the mix of activities performed in a job is changing by looking at new job postings, and how the mix of occupations in the economy is also changing. Alongside the fastest-growing and shrinking activities, we compute a dissimilarity index: for each month, it sums the absolute differences between the current share of each occupation or activity and its share in a baseline period, scaled to run from zero, meaning no change, to one hundred, meaning complete turnover.

By tracking activities we aim to compare our results directly to the Yale Budget Lab, which also tracks occupational dissimilarity compared to the pre-ChatGPT baseline. We construct the index on Revelio workforce data and extend it with alternative baselines and a decomposition that separates change in the mix of occupations from change in the activities performed within them.

Between-occupation dissimilarity. Shares are the fraction of total workforce headcount in each occupation each month, compared against three baselines (January 2016, January 2019, and November 2022) with a trailing 12-month moving average.

Within-occupation dissimilarity. Using activity shares within each occupation, we compute how far each occupation's task mix has moved from its baseline, then aggregate across occupations weighted by headcount share. A shift-share decomposition splits the year-over-year change into a within-occupation component, reflecting changing tasks inside roles, and a between-occupation component, reflecting a changing mix of roles.

The matching analysis combines Revelio Labs job-posting and workforce data to examine how many advertised openings correspond to observed external hires. The current firm universe consists of U.S. Russell 3000 companies with matched workforce and posting records. Interview-review measures are analyzed separately and describe candidates' reported experiences during the hiring process.

Job postings per external hire. New job postings are matched to observed external hires at the same company, state, seniority, and role. Hires are counted from the posting month through the following five months. This forward window allows time for an advertised position to produce a hire.

For each posting month, postings and matched external hires are aggregated across firms. The reported series divides total postings over the trailing 12 posting months by total matched external hires associated with those posting cohorts. It is therefore a ratio of trailing totals, not the reciprocal of an average monthly hires-per-posting series.

The measure is not a vacancy-fill rate. A posting can correspond to zero, one, or several matched hires, and some hires or postings may not be matched because the relevant company, location, seniority, or role information is unavailable. Because hires are observed over the posting month and five subsequent months, the chart ends earlier than the other monitor series.

Recruiter ghosting. The ghosting measure uses U.S. interview reviews for interviews that did not result in an offer. Review text is searched for fixed expressions such as "ghosted," "no response," "no reply," "never heard back," "radio silence," and related phrases. The monthly share is smoothed using a trailing 12-month average and indexed to its October 2022 level. A plotted value of 10% means the mention share is 10% above its October 2022 level, not 10 percentage points higher.

Interview complaint topics. The same no-offer interview-review sample is searched for fixed phrases associated with slow processes, poor communication, and automated screening. Each line shows the trailing 12-month average of the monthly share of reviews mentioning that topic. A review can mention more than one topic, so the shares are not mutually exclusive.

Sentiment at adopting firms. Employee-review measures compare business outlook, job security, and senior-leadership sentiment at AI integrator adopters and non-adopters. Firms are grouped by Revelio industry and October 2022 firm size. Size groups contain firms with fewer than 250 employees, 250–999 employees, and at least 1,000 employees. Non-adopter comparisons are drawn from the same industry-size cells as adopters; this is a grouped comparison rather than one-to-one matching.

Within each industry-size cell, the monthly adopter-minus-non-adopter sentiment gap is calculated and then aggregated using the number of adopter firms in that cell as the weight. Each series is expressed relative to its October 2022 gap. The plotted percentages describe standardized changes in the adopter–non-adopter gap, not literal percentage changes in review scores and not a causal effect occurring at each firm's adoption date.

Source: Revelio Labs employee reviews, workforce data, company industry data, and classified job postings.

Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Used for the event-study design behind the anticipatory-effects results in Equilibrium and the exposure-based results in Demand.

Hosseini Maasoum, S., & Lichtinger, D. (2025). Used for the firm-level AI adoption methodology behind the adoption-effects results in Equilibrium and the Adoption timing figure.

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. One of two external AI exposure measures benchmarked against Revelio's own in Figure A1.1.

Felten, E., Raj, M., & Seamans, R. (2023). How will Language Modelers like ChatGPT Affect Occupations and Industries? The second external AI exposure measure benchmarked in Figure A1.1.

Yale Budget Lab. Tracking the Labor Market Impact of AI. Source of the occupational dissimilarity approach extended in the Content & Occupational Churn methodology and Figures A1.4-A1.5.

Simon, L., Kharazian, Z., & Stevens, K. (2026). Joint research with Ramp Economics Lab on AI vendor spending and firm-level adoption, referenced in the adoption-effects discussion in Equilibrium.