Mercor Says 90% of US Work Hours Have No Real AI Benchmark

Mercor published a map on Oct. 15 that lays 29 AI benchmarks over the US Department of Labor's list of jobs, and concluded that "90% of American working hours are spent on tasks with little or no coverage from existing benchmarks." The page, titled "A benchmark map," calls the figure provisional.

The map is part of Mercor's research section, built by its APEX research team. It lists no individual authors and carries no publication date beyond a 2026 copyright line. The date of Oct. 15 comes from Runtime Wire, which wrote the release up the same day.

What the map counts

Mercor took 923 occupations from ONET, the Labor Department's occupational database, grouped into 22 job families. Those occupations carry 18,797 tasks. An LLM judge tagged each benchmark prompt with the ONET task it resembles, and a task counts as covered once at least five prompts map to it.

Under that rule, Mercor's page says 208 of the 923 occupations have at least one covered task. Across all occupations, 924 of the 18,797 tasks are covered.

Measure	Mercor's count
Occupations	923
Tasks	18,797
Benchmarks indexed	29
Occupations with a covered task	208
Tasks covered	924

The 90 percent figure weights tasks by time. Mercor took working-time shares from a paper by Hatgis-Kessell and co-authors, so a rare task counts for less than one that fills many hours of the US workweek.

1. Occupations with a covered task (208 of 923)22.5 %
2. Tasks covered (924 of 18,797)4.9 %

Source: Mercor, A benchmark map, accessed 2026-10-17; ratios calculated by The Terminal

Where coverage clusters

Mercor says coverage follows importance and exposure. Tasks tagged as exposed to AI, using the 2023 classification by Eloundou and co-authors, are about six times as likely to be covered as unexposed ones. The most important tasks within an occupation are six to 10 times as likely to be covered as the least important.

The page runs a case study on investment banking, law, accounting and software engineering. It compares scores weighted by O*NET against simple averages, and says only about a third of the relevant tasks had enough APEX coverage for the weighted version.

What Mercor says could be wrong

The page lists its own limits. The 29-benchmark list is incomplete, which biases coverage downward. Capping benchmarks per job family may understate fields with many tests, such as software engineering. Choosing benchmarks that are not yet saturated may understate tasks models already handle.

Two more caveats cut the other way. Benchmarks mostly measure automation rather than human augmentation. About 63 percent of jobs, Mercor says, involve physical-world components that current benchmarks do not capture.

The LLM tagging is itself a model judgment. Mercor said it checked the tags against human expert labels on its own APEX benchmarks, but the page does not give an agreement rate for the rest of the 29.

Mercor's own APEX-1 and APEX-Agents benchmarks are part of the index, so the company is both a mapmaker and one of the things on the map. Runtime Wire's write-up quotes no outside critic.

The full list of 29 benchmarks sits on a separate tab of the page, and Mercor calls the map "our first step toward building a comprehensive, continuously updated resource." The company has not said when the next update will appear.

Earlier on The Terminal: Microsoft's MAI base model and Nvidia's Nemotron 3.5 Bolt.