Mercor Labor Index: 90% of US Work Lacks AI Benchmarks

While AI benchmarks have proliferated rapidly across research labs, how well do existing evaluations measure economically valuable work across the real labor force? On October 16, 2026, Mercor released the US Benchmark Map under its Labor Index initiative, systematically mapping 29 public, unsaturated, agentic AI benchmarks onto the US Department of Labor's O*NET occupational taxonomy.

Only 208 of 923 Occupations and 924 of 18,797 Tasks Covered

The O*NET framework partitions the American workforce into 22 job families, 923 distinct occupations, and 18,797 granular occupational tasks. Mercor researchers analyzed how many tasks are evaluated by at least 5 benchmark prompts across the 29-benchmark index.

The empirical findings expose a stark blind spot: only 208 of the 923 occupations (and just 924 of the 18,797 tasks) meet the 5-prompt coverage threshold, while 6 of the 22 job families contain zero covered occupations. Weighted by Bureau of Labor Statistics (BLS) employment hours, 90% of American working hours are spent on tasks with little or no representation in current AI benchmarks.

Wage-Bill Concentration: Why Lawyers Are Benchmarked and Clerks Are Ignored

Running a task-level logistic regression controlling for binary AI-exposure classifications (Eloundou et al.) against log occupational wage bills, the study demonstrates that AI benchmark creators disproportionately target occupations with massive aggregate wage bills:

“90% of American working hours are spent on tasks with little or no coverage from existing benchmarks, making it remarkably difficult to measure how capable frontier models actually are at the long tail of middle-income knowledge work.”