Payroll Estimates Were Revised Down. BLS's Workforce Is 20% Smaller Than in 2024.

The number that moved markets on Aug. 7 was negative 23,000. That was the Bureau of Labor Statistics' first estimate of the change in U.S. payroll employment in July. The numbers that changed the recent past were larger: May and June together had 103,000 fewer jobs than BLS had reported one month earlier.

Large downward revisions invite a simple suspicion. If the first estimate was wrong, someone must have failed, or worse, manipulated it. The actual process is less dramatic and more important. BLS publishes an early estimate from a voluntary employer survey, keeps collecting reports, revises the estimate twice, and later benchmarks the entire series to far more complete unemployment-insurance records. Revision is not a correction imposed from outside the system. Revision is part of the system.

But the routine nature of revisions does not settle the question of whether the system has enough capacity. Public personnel records show that BLS had 1,865 covered civilian employees in June 2026, down 456, or 19.6%, from September 2024. Its largest professional groups all became smaller. Employer response rates remain far below their level a decade ago. The agency's own recruiting slowed sharply after 2024.

None of those facts proves that staff losses caused the May or June revisions. A June 2026 Government Accountability Office review found the opposite of an immediate collapse: BLS met its accuracy goals for payroll revisions from fiscal 2020 through 2025, and most of the experts GAO interviewed considered the report reliable enough for policy and business decisions.

The evidence supports a narrower conclusion. The July update was a normal, consequential revision. At the same time, the statistical institution responsible for producing it has less workforce capacity and faces a harder data-collection environment. That is a reason to watch the machinery, not to invent a motive.

What changed in the July report

The July 2026 Employment Situation contains two different surveys. The Current Employment Statistics survey, commonly called the establishment or payroll survey, produced the negative 23,000 jobs estimate. The Current Population Survey, or household survey, produced the 4.1% unemployment rate. Census Bureau interviewers collect the household data for BLS, while BLS collects the employer data.

The payroll revisions in the July release affected May and June:

May's revision history is longer than the one-month comparison in that table. Its first estimate was positive 172,000, its second was positive 129,000, and its third was positive 63,000. From first release to third, May moved down by 109,000. June had reached only its second estimate when the July report was issued, so its final monthly estimate was not yet available.

That distinction prevents two common counting errors. The 103,000 figure is the combined change from the estimates published immediately before the July report, not the combined change from the first estimate ever published for each month. It also mixes May's third release with June's second. A fair historical comparison has to hold the revision stage constant. That is why the long-run chart below compares first estimates with third estimates and treats 2026 as incomplete.

The negative 23,000 estimate for July was also preliminary. BLS described payroll employment as having "changed little," language that reflects the uncertainty around a first survey estimate. GAO reported that the margin of error around an initial monthly payroll change is about 130,000 jobs at the 90% confidence level. That does not make the estimate useless. It means one month's point estimate should not be read as a census.

Revisions are how an early estimate becomes more complete

The establishment survey asks about 119,000 businesses and government agencies, representing roughly 622,000 worksites, for payroll employment, hours, and earnings. Employers report voluntarily. BLS begins collecting after the pay period that includes the 12th of the month, then closes the first estimate shortly before the Jobs Report is released.

That timetable is the tradeoff. Policymakers and the public receive a national employment estimate only weeks after the reference period, but not every employer has answered by then. BLS continues collecting the missing reports and publishes a second estimate the following month and a third estimate one month after that. Seasonal factors are recalculated as well.

GAO found that in 2024, the reports available to BLS covered about 60% of sampled employer employment at the first release and about 90% by the third. Late reports are not a surprise discovered after publication. They are the main reason the second and third estimates exist.

An annual benchmark is a different operation. The monthly estimates are based on a sample covering about 26% of U.S. employment. Each year, BLS compares them with unemployment-insurance tax records that cover about 97%. It replaces the March survey level with that administrative count and adjusts surrounding months. The benchmark can revise a much longer period than the two monthly updates.

The net birth-death model is another separate piece. New businesses do not immediately appear on the survey frame, while a business that closes often stops responding before its status is known. BLS estimates those openings and closures. In February 2026, the agency introduced a model change intended to make that estimate more responsive to current conditions and reduce future benchmark error.

These steps do not guarantee that the first estimate will be close to the final one. They explain why the final one can be better. A system that never revised an incomplete early estimate would be less transparent, not more.

The BLS workforce is smaller than at any point in this series

FederalHiringData analyzed public Office of Personnel Management employment files for agency subelement DLLS, the Bureau of Labor Statistics. September snapshots provide a consistent annual series from 2000 through 2025; June 2026 is the latest monthly snapshot available.

The count was 2,586 in September 2000, 2,390 in 2014, and 2,321 in 2024. It fell to 2,165 in September 2025 and 1,865 by June 2026. The latest count is 28% below the 2000 level and 19.6% below September 2024.

These are covered civilian employees, not full-time-equivalent positions and not a program budget. BLS also relies on Census staff to collect the household survey. The OPM files do not reveal how many BLS employees work specifically on the national payroll survey, seasonal adjustment, benchmarking, release review, or other labor-market programs. The agency also produces inflation, wage, productivity, workplace-injury, occupational, and employment-projection data.

The occupational detail still shows where the contraction occurred. Economists, the largest BLS group, fell from 1,185 in September 2024 to 937 in June 2026. Information technology management declined from 378 to 305. Economics assistants went from 332 to 285. Statisticians fell from 79 to 63, and mathematical statisticians from 100 to 85.

Those occupations span more than the arithmetic behind one headline number. Economists design and interpret programs, statisticians work on sampling and estimation, economics assistants support collection and production, and IT staff maintain systems that receive, edit, protect, and publish sensitive data on fixed schedules. The public files do not map a lost position to a delayed task. They do show that the reduction crossed the main professional groups that make a modern statistical agency function, rather than being confined to one administrative category.

The latest workforce also has a succession question. OPM records show 558 BLS employees were age 55 or older in June, 29.9% of the covered workforce. Age does not predict an individual's retirement date, and experience is not interchangeable with job title. It does show that replacing lost expertise may remain a challenge even after broad reductions stop.

Budget documents tell a related story in a different unit. The Department of Labor's fiscal 2026 BLS justification projected 1,995 total FTE supported by budgetary resources, 180 below the fiscal 2025 level. The fiscal 2027 request proposes 1,754 FTE. A request is not an enacted outcome, and FTE cannot be compared directly with OPM headcount. Both series, however, point toward less staffing capacity than BLS had recently.

Fewer employers are responding

Staffing is only one side of survey production. The other is whether households and employers answer.

BLS publishes monthly unit response rates for its establishment surveys. FederalHiringData calculated calendar-year averages for CES at the third release, the point after late reports have had roughly two additional months to arrive. The average was 60.8% in 2016 and 42.2% in 2025. GAO used a fiscal-year comparison and reached nearly the same endpoints: 62% in 2015 and 42% in 2025.

Response rate and collection coverage are related but not identical. BLS's response-rate denominator includes sampled employers that refused or became long-term nonrespondents. Its collection rate focuses on the active sample it reasonably expects to hear from. That is why GAO could report about 90% employer-employment coverage by the third 2024 release while the unit response rate was about 43%.

A falling response rate does not prove the estimates are biased. Bias depends on whether the employers that answer differ systematically from those that do not in a way that changes the result. Statistical weighting and the annual benchmark can reduce some errors. A large number of responses can also preserve useful precision even when the percentage is lower.

Still, response decline has consequences. GAO found that BLS lowered its annual goal for the number of detailed establishment estimates it could publish from 25,450 in fiscal 2020 to 22,900 in fiscal 2025. Smaller samples led the agency to suppress some detailed industry estimates to meet quality and confidentiality standards. BLS missed that usefulness goal in fiscal 2022, 2024, and 2025.

The production burden also continues after a report arrives. BLS has to validate identifiers, compare current and prior payrolls, identify improbable values, resolve reporting-unit changes, protect confidential employer information, calculate weights, produce seasonal adjustments, and prepare thousands of industry estimates for synchronized release. Electronic submission reduces manual collection, but it does not eliminate statistical review. A lower response rate can therefore create two pressures at once: fewer observations for some estimates and more need to understand whether the respondents that did report still represent the intended population.

GAO also found limits in what the public can evaluate. BLS annually studies nonresponse and other sources of error by comparing survey estimates with the administrative benchmark, but it has not released a complete summary of those findings because of methodological and data-quality limitations. GAO recommended publishing an assessment that users can scrutinize. BLS generally agreed.

Recent revisions are larger, but the history is not one-directional

BLS publishes first, second, and third estimates of seasonally adjusted monthly payroll changes back to 1979. FederalHiringData calculated the mean absolute change from first to third release for each year. This asks how far estimates moved, regardless of direction.

The mean absolute revision was 27,700 in 2022, 50,500 in 2023, and 48,400 in 2024. For the 11 available 2025 months, it was 58,000. No initial October 2025 estimate was issued during the government shutdown. For the five 2026 months with completed third estimates, January through May, the mean absolute revision was 54,400.

The recent figures deserve attention, but the series is not a steady deterioration. Pandemic disruption produced much larger averages, about 130,400 in 2020 and 180,500 in 2021. The 2003-to-present BLS summary reports a 51,000 mean absolute first-to-third revision. May 2026's 109,000 downward revision from first to third was large, but not unprecedented in the monthly history.

Direction matters too. The mean first-to-third revision was negative 58,000 across the available 2025 months and negative 14,800 through the five completed 2026 months. In 2021 it was positive 158,700. Those swings are a reason to analyze a sequence of reports rather than treating a single preliminary month as a final verdict.

GAO tested the institution more directly than a visual reading of the chart. BLS's goal is for the third monthly estimate to remain within 0.1 percentage points of the initial employment-growth estimate in at least 10 of 12 months. The agency met that goal in fiscal 2022 through 2025. It also met its five-year benchmark-revision goal through 2025, even as some individual benchmarks were large.

Related research

Put this finding in context

All articles

Data investigations/August 23, 2026

NASA Cut Its Civil-Service Workforce 22%. Now It Wants Core Technical Work Back In-House.

NASA lost more than 4,000 civil servants before directing its centers to restore in-house engineering, operations and science capability.

Data investigations/August 23, 2026

ICE Got 150,000 Applications. Its Workforce Grew by About 9,600 at the Peak.

OPM recorded 10,030 new-hire actions during ICE's surge, but tentative offers were not final hires and covered employment had receded by June.

GAO interviewed 14 former officials, policymakers, researchers, business representatives, and financial-market users. All 14 said the report used accepted statistical practices and was timely enough for decisions. Twelve considered overall accuracy and reliability acceptable. Several also warned that occasional large revisions create uncertainty. The report's strength and its risks can coexist.

Recruiting slowed as the workload stayed public

The FederalHiringData historical USAJOBS archive offers a separate view of BLS replenishment. It contains 268 distinct BLS announcements opened in 2024, 47 in 2025, and 39 through Aug. 13, 2026. The series begins in March 2018 and counts announcements, not vacancies or hires. One announcement can advertise multiple positions, and not every selection becomes an OPM accession.

OPM action files provide a second, nonidentical flow measure. They record 12 BLS accessions and 441 separations in calendar 2025, compared with 176 and 195 in 2024. Through June 2026, they record 68 accessions and 66 separations. Action totals do not reconcile one-for-one with point-in-time headcount because timing, coverage, corrections, transfers, appointments, and reporting rules differ. The direction is nonetheless consistent with the headcount and announcement evidence: 2025 was a year of unusually little entry and substantial exit.

The operational effects are not entirely hypothetical. GAO reported that funding constraints, staffing reductions, and shutdown delays slowed development of an online response option for the household survey, limited field testing, and may delay its planned 2027 launch. That finding establishes an effect on modernization. It does not establish an effect on the May or June payroll estimate, which comes from the establishment survey.

The distinction matters. It would be irresponsible to use a 20% agency headcount decline as a formula for estimating payroll error. It would also be complacent to assume a smaller workforce can indefinitely maintain every survey, quality review, modernization project, release schedule, and public explanation without tradeoffs.

What a credible capacity test would look like

The next test is not whether a revision occurs. It is whether BLS continues to meet transparent quality standards while publishing enough information for outsiders to understand the result.

Four measures would make that test concrete. First, BLS should continue publishing first, second, and third payroll vintages with clear revision summaries. Second, it should publish the nonresponse assessment GAO recommended, including what can and cannot be inferred from lower employer participation. Third, Congress and the public should track program-level staffing and the number of detailed estimates suppressed for quality reasons. Fourth, BLS should report modernization milestones, costs, and delays instead of treating a future online response option as a binary launch date.

A useful capacity scorecard would also keep the measures separate. Headcount would describe people employed at a point in time. FTE would describe labor funded over a fiscal year. Response rates would describe participating sample units. Collection coverage would describe the share of expected employer employment represented by received reports. Monthly revisions would describe the arrival of additional sample information, while benchmark revisions would compare the survey with administrative records. Combining those measures into one verdict would obscure the exact risk each one is designed to reveal.

The preliminary annual benchmark for March 2026 is scheduled for Aug. 28, after this article's data cutoff. That release will provide a much broader comparison between the payroll survey and unemployment-insurance records. It should not be confused with the 103,000 monthly revision in the July report, and it should be evaluated on its own terms.

The public interest is served by resisting two easy stories. One says every revision is evidence of manipulation. The other says revisions are routine, so institutional capacity does not matter. The evidence supports neither. Revisions are the visible mechanism by which a fast survey absorbs more information. A smaller statistical workforce and lower response rates raise the cost of sustaining that mechanism well.

Methodology and limitations

FederalHiringData downloaded the official BLS July 2026 Employment Situation, the BLS total-nonfarm revision table, and the BLS establishment-survey response-rate series on Aug. 24, 2026. Revision calculations use seasonally adjusted changes in total nonfarm payroll employment. The 2025 annual average uses 11 months because BLS did not issue an initial October estimate during the shutdown; the 2026 value uses only January through May, the months with third estimates available.

Workforce calculations use public OPM FedScope and Enterprise Human Resources Integration records for BLS agency subelement DLLS. Annual points are September snapshots through 2025; June 2026 is the latest monthly file available. These counts exclude workers outside the covered OPM population and do not identify staffing assigned to a specific BLS program. Headcount is not FTE.

USAJOBS calculations use the FederalHiringData historical archive and count distinct BLS control numbers by announcement opening year. The BLS series begins in March 2018. Announcement counts are not vacancy counts, applications, selections, or hires. The 2026 total is incomplete through Aug. 13.

This analysis does not estimate a causal relationship between staffing, response rates, and any monthly payroll revision. The strongest available external review, GAO-26-107538, found the Jobs Report generally met users' needs and BLS met its accuracy goals through fiscal 2025, while also documenting response-rate risk, reduced detail, limited public nonresponse analysis, and modernization delays linked to funding and staffing constraints.

Readers looking for current federal openings can browse active jobs. FederalHiringData's statistics and agency directory provide separate views of the current and historical federal hiring archive. More workforce investigations are available in Articles.