Most American work has barely been tested on AI, a Mercor map finds

Nine of every ten hours Americans spend at work go to tasks that public AI tests barely touch, according to a Benchmark Map published by Mercor on its research site. The number is easy to misread. It says nothing about what AI cannot do. It says how little of the question anyone has tried to answer.

Mercor is a company that connects human experts with AI labs, including for building evaluations and training data. That gives it a commercial interest in the idea that the industry needs more and better tests of real work. The finding can be convenient for the publisher and still be true, so the method matters more than the motive.

The map starts from ONET, the US Department of Labor’s catalogue of 923 occupations and 18,797 tasks. Mercor then chose 29 benchmarks: public ones with at least 25 open tasks, built around agents doing work rather than answering questions, realistic about what enterprises ask for, and not yet mastered by frontier models. An AI model acting as judge tagged every benchmark prompt with the ONET tasks it matches. Mercor checked that tagging against human expert tags, but only on its own APEX benchmarks, a subset of the 29.

The hours come from a second model. Mercor combined its coverage estimates with a published LLM-derived estimate, from Hatgis-Kessell and colleagues, of how much working time each task takes up. So a model sits at both ends of the 90% figure, and the page does not say how much error that adds.

“Little or no coverage” is a single cutoff, not a graded score. A task counts as covered only if the judge matched it to at least five prompts. A task with four prompts is uncovered, and so is one with none. The page does not split the 90% into “nothing exists” and “something thin exists,” which are different claims. On the occupation view, only 208 of 923 occupations have even one task past the cutoff, and six of 22 job families have none.

Mercor lists its own caveats. Any incomplete list pushes measured coverage down, so the gap is probably overstated, though the company says that bias is hard to bound. Leaving out benchmarks that models already ace may understate coverage of easier tasks. Capping benchmarks per job family understates software engineering. Several selection criteria, such as “economic realism,” involve judgment calls. Mercor calls the results provisional.

The covered slice follows exposure and money. Tasks rated as exposed to AI by Eloundou and colleagues in 2023 are about six times as likely to appear in benchmarks as unexposed ones. Software engineering and financial analysis are dense; truck driving and gas plant operation are sparse. Larger wage bills attract more tests even after adjusting for exposure, and the tasks O*NET rates most important are six to ten times likelier to be covered than the least important. One example is legal: AI-exposed lawyer tasks are heavily tested, while those of legal secretaries and court clerks barely are.

Nearly every public argument about AI and jobs runs through benchmarks, whether a lab is boasting or a skeptic is deflating. If a task has under five tests behind it, a confident claim in either direction rests on almost nothing. A slide saying a model is ready for court clerks and a post saying it is hopeless at them would both lean on the same thin evidence.

Mercor also tried to turn benchmark scores into per-job scores for four occupations, weighting each task equally. The weighted and plain averages came out close. But fewer than a third of the relevant tasks had enough prompts to be included at all. Physical work is a further blind spot: Mercor cites an estimate that only 37% of US jobs could be done entirely from home.

The page is public, with a clickable map, a benchmark index and a methodology tab. Mercor says it will update results as it widens the list, but the page carries no date or version number, so the update schedule cannot be checked from outside.

The practical test for the next ranking of models by profession: ask how many separate tested tasks sit under that job. On this map, fewer than five counts as none.

Based on Mercor’s Benchmark Map research page (mercor.com/research/labor-index/benchmark-map), which carries no publication date; accessed 19 October 2026.