Mercor maps AI benchmarks to the economy and finds 90% of US work hours uncovered

Mercor published a Benchmark Map of the US economy that tries to answer a question the AI industry mostly avoids: how much of what Americans actually do for a living is even being measured? The answer is uncomfortable. Of 923 occupations in the Department of Labor’s ONET taxonomy, only 208 have a single task covered by at least five benchmark prompts. Of 18,797 ONET tasks, 924 clear that bar. Mercor’s headline estimate is that 90% of US working hours go to tasks with little or no benchmark coverage.

That number will get quoted a lot. It deserves to be, with caveats attached, because the methodology is honest about its own limits and the pattern underneath the number is the more interesting finding.

What Mercor actually built

The map starts from ONET, which splits the workforce into 923 occupations and 18,797 tasks. Mercor selected 29 benchmarks, filtered for public availability (at least 25 public tasks), agentic framing (the model has to do the task, not advise on it), economic realism, economic diversity, and a check against saturation by frontier models. An LLM judge tags each benchmark prompt with the ONET tasks it corresponds to, validated against human expert tags on Mercor’s own APEX benchmarks. A task counts as covered if at least five prompts map to it. Working-time weights come from Hatgis-Kessell et al.

The result is a map with deep, narrow coverage. Computer and Math occupations: 29 of 36 covered. Business and Financial: 33 of 48. Then the drop. Healthcare Practitioners: 12 of 89. Science: 15 of 60. Education and Library: 9 of 62. Six of the 22 job families have no occupations meeting the coverage threshold at all.

The gap is not random, and that is the point

Mercor is careful to say the skew is partly rational. Tasks tagged as AI-exposed in Eloundou et al. (2023) are about 6x more likely to appear in benchmarks than unexposed tasks. That tracks: benchmarks measure where progress is plausible, and progress is most plausible on exposed work. Within occupations, the tasks O*NET experts rate most important are 6–10x more likely to be covered than the least important. Also defensible.

But the wage-bill finding is where the map earns its keep. Even after controlling for AI exposure, tasks from occupations with larger annual wage bills are significantly more likely to be covered. Mercor’s example is sharp: benchmarks cover AI-exposed lawyers’ tasks heavily but barely touch the O*NET tasks of similarly AI-exposed legal secretaries or court clerks. The long tail of middle-income knowledge work falls through.

The benchmarks are not wrong. They are aimed at the biggest, most exposed wage bills, and the rest of the economy is dark.

Read that as a market signal rather than a scandal. Benchmark builders have limited resources, and pointing them at software engineering and financial analysis is where the returns look clearest. Mercor says as much: the allocation is “reasonable” given the resources devoted to high-quality, economically relevant benchmarks so far. The problem is what happens if it stays that way.

What the 90% number can and cannot carry

Mercor flags its own biases, and they matter. The 29-benchmark list is not a representative sample of all unsaturated, public, agentic economic benchmarks, and any incomplete list biases coverage estimates downward. The team says it cannot bound that bias. Capping benchmarks per job family to force breadth biases down coverage estimates in the most-benchmarked families, software engineering chief among them. Excluding saturated benchmarks risks understating coverage on tasks LLMs find easy. Mercor argues the alternative, including saturated benchmarks, would overstate real-world capability more, since benchmarks often overstate deployment performance. The results are labeled provisional.

So treat 90% as directionally right and numerically soft. The direction is what matters. If a frontier lab wants to claim its model is good at “knowledge work,” the map says we mostly cannot check. We can check coding, some financial analysis, a slice of legal work, and a few dozen other pockets. The rest is inference from adjacent evidence.

Why this lands now

The timing is not accidental. Enterprises are being asked to buy agentic software on the strength of benchmark scores, and policymakers are being asked to forecast AI’s labor impact on the same evidence. Mercor names both audiences, plus workers thinking about careers. A benchmark suite that covers 208 of 923 occupations serves none of them well at the margin that matters.

There is a second-order problem Mercor raises and does not fully solve. Even where coverage exists, averaging model performance across occupation-relevant prompts overweights frequently benchmarked tasks. O*NET offers a way to reweight by task importance and time share, but the underlying coverage is still thin. You cannot reweight your way out of a missing task.

What builders should take from it

The obvious move is more benchmarks in the uncovered families: healthcare practitioners, office and administrative support, sales, the trades. Mercor says it will expand the list and update the map. That is the right instinct, and it is also where the incentive problem bites. Benchmarks for gas plant operation and truck driving are expensive to build, hard to validate, and unlikely to produce a leaderboard screenshot anyone shares.

The less obvious move is to stop treating benchmark coverage as a proxy for economic relevance. A model that scores well on the 29 benchmarks in this index has demonstrated something real. It has not demonstrated that it can do the work of the 715 occupations with no covered task. The gap between those two claims is where a lot of enterprise procurement and policy forecasting currently lives.

Watch whether the next round of benchmark funding follows the wage-bill gradient or breaks it. Mercor’s map gives the field a way to check.