Mercor, a company best known for its AI-driven recruiting platform, has published a mapping exercise comparing how US work hours are distributed across occupations with the tasks actually covered by existing AI benchmarks. The headline finding is that roughly 90% of American working time is spent on activities poorly represented, or entirely absent, from current evaluation suites such as MMLU, HumanEval, and similar references used by labs to track model progress.
The analysis points to a structural imbalance: most benchmarks concentrate on easily quantifiable domains like coding, mathematics, and academic-style problem solving, which together make up only a small slice of the real economy. Meanwhile, activities that occupy the bulk of the workforce, such as team management, sales negotiation, customer service, logistics coordination, and complex administrative work, remain largely untested by existing evaluation frameworks.
This mismatch carries real consequences for how progress claims from model developers should be interpreted. A strong score on a coding or math reasoning benchmark says relatively little about a system's actual ability to assist or replace workers across the majority of existing jobs. Mercor, which operates at the intersection of hiring and skills assessment, has an evident stake in highlighting this gap, since its business model depends on qualifying human workers and, increasingly, on building training data that reflects genuine occupational tasks.
Setting aside the company's commercial angle, the finding echoes a broader critique voiced by several AI researchers: current benchmarks measure what is convenient to measure, not necessarily what matters economically. As debates over labor automation intensify, this kind of mapping argues for more cautious readings of performance announcements and for greater investment in evaluations that better reflect the diversity of real-world occupations.