90% of US working hours are spent on tasks not covered well by AI

Mercor launched its Benchmark Map of the US economy, evaluating how effectively contemporary artificial intelligence benchmarks measure the tasks performed by American workers. According to mercor.com's report, existing evaluation suites leave ninety percent of American working hours on tasks with little or no benchmark coverage. By auditing twenty-nine agentic benchmarks against federal labor data, researchers found that out of eighteen thousand seven hundred ninety-seven occupational tasks across nine hundred twenty-three occupations, only nine hundred twenty-four tasks are covered by five or more benchmark prompts.

This analysis details the uneven distribution of automated evaluation across the United States labor market, identifying which job families possess rigorous testing coverage and which remain unexamined by model developers. Enterprise software architects, technology executives, machine learning researchers, and workforce strategists will find concrete data regarding where frontier model capabilities have been validated against operational labor. The findings demonstrate how current model evaluations favor high-wage occupations with large aggregate payrolls while neglecting middle-income knowledge workers who perform routine digital tasks across corporate administrative workflows.

90% of US working hours are spent on tasks: what actually changed

The release of the Benchmark Map establishes a systematic audit of AI evaluations mapped against the United States Department of Labor's ONET taxonomy. Across nine hundred twenty-three ONET occupations comprising the domestic labor force, only two hundred eight occupations feature at least one task evaluated by five or more benchmark prompts. Out of eighteen thousand seven hundred ninety-seven discrete tasks cataloged by federal analysts, only nine hundred twenty-four meet that threshold within the index. Furthermore, six of twenty-two major job families contain zero covered occupations.

The findings show that the proliferation of AI testing suites has created an illusion of comprehensive workplace evaluation. While frontier benchmarks test complex academic exercises or software repositories, ninety percent of American working hours involve operational responsibilities that receive zero or negligible evaluation. In constructing the index, Mercor selected twenty-nine benchmarks totaling hundreds of prompts, including public benchmarks and its proprietary APEX suite. The criteria required that each benchmark contain at least twenty-five public tasks, evaluate agentic execution rather than multiple-choice question answering, and reflect realistic enterprise workloads.

90% of US working hours are spent on tasks: how it works

Mercor mapped twenty-nine public and proprietary benchmarks against eighteen thousand seven hundred ninety-seven tasks defined across nine hundred twenty-three O*NET occupations. Following methodology established by Wang and colleagues, researchers deployed a large language model judge to associate each benchmark prompt with corresponding occupational tasks. To ensure evaluation fidelity, the team validated the automated tagging against human expert annotations on Mercor's APEX benchmark subset. A task qualified as covered only when the judge mapped it to at least five distinct benchmark prompts across the index.

To estimate the economic labor represented across these benchmarks, researchers linked their coverage tallies to working-time share estimates developed by Hatgis-Kessell and colleagues. The team paired these allocations with occupation-level wage bills calculated from Bureau of Labor Statistics data, assuming forty-hour work weeks. They executed task-level logistic regressions combining binary AI-exposure tags from research by Eloundou and colleagues with log occupational wage bills, computing delta-method ninety-five percent confidence intervals. Benchmarks were capped per job family during selection to curb the severe overrepresentation of software engineering.

90% of US working hours are spent on tasks: why it matters now

Enterprise teams currently deploy artificial intelligence systems into corporate operations without empirical benchmarks measuring how models perform across ninety percent of domestic working hours. Data demonstrates that evaluations prioritize occupations based on total wage bills rather than operational necessity. Even after controlling for AI exposure tags from Eloundou and colleagues, tasks from occupations with larger wage bills are significantly more likely to receive benchmark coverage. An O*NET task flagged as exposed to AI is six times more likely to appear in benchmarks than an unexposed task.

This benchmark concentration produces significant blind spots for organizations adopting automated tools across administrative workflows. While benchmarks thoroughly evaluate the core responsibilities of high-wage lawyers, including analyzing legal precedents and drafting contracts, they almost completely omit tasks performed by court clerks and legal secretaries who possess similar exposure ratings. The long tail of middle-income knowledge work remains untested by frontier evaluation suites. Consequently, corporate leaders risk mistaking benchmark scores in software engineering or financial analysis for generalized autonomous capability in untested corporate units.

90% of US working hours are spent on tasks: who is affected

Uneven coverage documented across twenty-two O*NET job families reveals stark disparities in how artificial intelligence benchmarks measure different sectors of the workforce. Computer and mathematical occupations show the strongest presence, with twenty-nine out of thirty-six occupations covered, followed by business and financial operations with thirty-three out of forty-eight covered. Community service has ten out of fourteen occupations covered, while legal professions see five of seven covered. Office and administrative support has twenty-one of fifty-one covered, management sees twenty-one of fifty-six, and engineering records nineteen of fifty-six.

Conversely, major sectors of the domestic economy remain almost entirely unmeasured by modern benchmark infrastructure. Arts and media counts thirteen of forty occupations covered, while science has fifteen of sixty covered. Healthcare practitioners record only twelve of eighty-nine occupations covered, education and library occupations feature nine covered roles out of sixty-two, and sales occupations have six of twenty-two covered. Non-office sectors like transportation, truck driving, and gas plant operations lack meaningful task representation, leaving workers across these professions subject to unvalidated enterprise software deployments.

90% of US working hours are spent on tasks: what to watch

The research outlines several methodological boundaries that enterprise evaluators must consider as the Benchmark Map expands over future cycles. Mercor designated its findings as provisional because the analysis relies on an initial cohort of twenty-nine curated benchmarks. Capping benchmarks included per job family intentionally reduced measured coverage in heavily evaluated domains like software engineering to prevent distortion. Furthermore, by deliberately excluding saturated benchmarks where models score near ceiling, the study avoided tests that inflate real-world capabilities, though this criterion risks understating coverage on simpler administrative tasks that models complete reliably.

Going forward, tracking how evaluation developers expand benchmark diversity into the long tail of corporate labor will determine the reliability of autonomous enterprise deployments. Mercor plans to update the index continuously, incorporating additional public and held-out APEX prompts while formalizing subjective selection criteria like economic realism. Observers should monitor whether new agentic benchmarks begin testing administrative support, healthcare documentation, and court administration, or continue gravitating exclusively toward software development. Until public benchmarks evaluate tasks across the remaining ninety percent of working hours, claims regarding enterprise workplace automation will remain disconnected from empirical measurement.