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

Mercor released the map on October 15 as part of its Labor Index. It tags every prompt in 29 public agentic benchmarks with matching tasks from O*NET, the Labor Department's catalog of 923 occupations and 18,797 tasks. An LLM judge does the tagging, checked against human expert tags on Mercor's own APEX benchmarks. A task counts as covered only with at least five matching prompts, which 924 tasks meet. Six of 22 job families have no covered occupation. Hours per task come from separate time-use estimates, which produces the 90% figure. Coverage skews toward work that is already exposed to AI and well paid. AI-exposed tasks are about six times more likely to appear than unexposed ones, and larger wage bills predict more representation. Lawyers are well covered, while legal secretaries and court clerks barely register. Mercor calls the results provisional. Its cap on benchmarks per job family may understate coverage in heavily tested fields such as software engineering. For founders, benchmark scores are a weak proxy for readiness outside a few domains like coding and law. Customers in uncovered fields will need task-specific evals, which means building them in-house or buying them. Mercor sells expert data and evaluations, so the gap it documents is also its market, and the method deserves that scrutiny.