AI benchmarks miss 90% of US work hours
Original: 90% of US working hours are spent on tasks not covered well by AI benchmarks
Why This Matters
Without broad benchmark coverage, capability claims about AI in real work remain largely untestable.
Mercor Research finds only 208 of 923 US occupations have tasks covered by 5+ AI benchmark prompts, leaving 90% of American working hours with little or no AI evaluation coverage, per its new Benchmark Map.
Mercor's Benchmark Map, built on the US Department of Labor's O*NET taxonomy, indexed 29 benchmarks against 18,797 tasks across 923 occupations. The result: only 208 occupations have any meaningful benchmark coverage (defined as at least 5 prompts), and 6 of 22 job families have none at all. That gap means roughly 90% of American working hours involve tasks that existing AI benchmarks barely touch. The skew isn't random. Benchmark coverage tracks occupation wage bills—high-earning, high-AI-exposure roles like lawyers get disproportionate attention, while similarly AI-exposed but lower-wage roles like legal secretaries and court clerks are largely ignored. The same pattern holds across middle-income knowledge work broadly. Computer and math roles lead coverage (29 of 36 occupations), while healthcare practitioners (12 of 89) and education and library roles (9 of 62) trail far behind. Mercor frames this Benchmark Map as a first step toward a continuously updated index of AI capabilities across the economy, with explicit goals to surface and help fill measurement gaps.