Most U.S. Work Hours Fall Outside AI Benchmarks
90% of American working hours go to tasks with little or no coverage in existing AI benchmarks. That is the central finding of Mercor's Benchmark Map research, published Oct. 16, 2026, as 'A benchmark map of the US economy' Benchmark Map.
The study uses the U.S. Department of Labor's ONET taxonomy as its reference. ONET, the government's catalog of jobs and work activities, divides the workforce into 923 occupations and lists 18,797 tasks across them.
To measure coverage, Mercor gathered 29 benchmarks chosen for quality and economic relevance. An LLM judge, an AI system used to match text at scale, mapped prompts from those benchmarks to O*NET tasks. A task counted as covered only if it linked to at least five prompts.
By that test, 208 of 923 occupations are represented in existing benchmarks Mercor. The rest fall below the five-prompt threshold.
Coverage is also uneven. Benchmarks cover many tasks for AI-exposed lawyers but cover few O*NET tasks for similarly AI-exposed legal secretaries or court clerks.
Mercor places the map alongside its AI Productivity Index (APEX), a family of benchmarks that track how well AI models and agents handle economically valuable tasks APEX. In March 2026, the company added APEX-SWE, a benchmark built with Cognition Introducing APEX-SWE.
The practical implication here is about scope rather than rigor. Current tests can be hard within their own tasks and still leave most working hours unmeasured. That leaves a gap between reported scores and daily operations. A model can move up a leaderboard without facing the routine drafting, filing, scheduling and record-keeping that keep an office running. Like a report card that grades one subject in detail and leaves the rest blank, the scores give a partial picture. Procurement teams and builders that use benchmark scores as a stand-in for job readiness will overestimate readiness for uncovered roles and underestimate it where tests cluster. Hours-weighting shifts the question from whether a model can answer isolated prompts to whether test time matches where people spend time. It does not require a test for every task, but it does require a list of what is missing.
In my view, a census like this is useful because missing work becomes explicit and can be addressed. An automated judge will misclassify edge cases, so practitioners should treat the map as directional rather than exact. Directional is enough to act on. Once uncovered tasks are listed, they can be prioritized for new tests and for product work that fits real workflows. I saw this with my own children, who picked up new tools not because the demos impressed them but because the tools fit into homework and part-time jobs without friction. The broader context here is positive. Better maps should lead to better benchmarks, and better benchmarks should lead to systems that are useful across more of the economy, not only its most visible jobs.
