90% of US work happens on tasks AI benchmarks can't measure
Mercor's new Benchmark Map shows most American jobs are left out of AI performance testing.

- Mercor studied 923 U.S. occupations
- They matched tasks to AI benchmark questions
- Only 208 jobs had enough benchmark coverage
- They found 90% of work is untouched
A map of what AI can and cannot do
Mercor released a new research report called the Benchmark Map, which tracks how well existing AI benchmarks cover real American jobs. The findings are striking: about 90% of working hours in the United States are spent on tasks that standard AI benchmarks barely touch 1.
To build the map, Mercor studied 923 occupations listed by the US Department of Labor's O*NET system, which describes all major types of work in the country. Across those jobs, O*NET lists 18,797 individual tasks — things like "prepare, draft, and review legal documents" or "advise clients concerning business transactions." (examples from the source) Mercor then checked how many of those tasks appear in a curated list of 29 AI benchmarks known for quality and real-world relevance.
Only 208 of the 923 occupations had at least five benchmark prompts that could be linked to their tasks. Six out of 22 major job families had no occupations with that level of coverage at all.
The bias is obvious
The gaps are not random. The data show that AI benchmark makers tend to focus on large, high-paying industries first. Jobs in fields like software engineering and finance dominate popular benchmarks, while equally important middle-income roles get far less attention.
Take the legal field, for example. Benchmarks heavily cover tasks done by lawyers, but they barely test the work of legal secretaries or court clerks — even though those roles face similar levels of AI exposure 1. The same pattern repeats across other sectors: the "long tail" of smaller and mid-size knowledge work jobs is largely invisible to AI evaluators.
Why this matters
The problem goes beyond missing data. If benchmarks don't cover what most Americans actually do at work, it becomes very difficult to know whether current AI models are truly ready for real-world use. An AI model might score well on coding challenges but struggle with the everyday tasks that make up the bulk of the economy.
Mercor said it plans to keep expanding its benchmark coverage so researchers can better understand where AI stands and where it falls short. For now, the Benchmark Map serves as a reminder that today's AI progress reports may not tell the full story about how well machines can handle the American workforce.
Corrections and updates
- Corrected on 12 November 2026. An earlier version cited example tasks not found in the source; the replacement correctly uses verbatim O*NET task descriptions for lawyers. For item 2, both versions conveyed the same key finding supported by the source.
Key facts
- 90% of US working hours are spent on tasks with little or no benchmark coverage
- Only 208 of 923 US occupations have at least 5 benchmark prompts linked to their tasks
- Six of 22 major job families have no occupations with 5+ benchmark coverage
- Benchmarks overrepresent high-wage industries like software and finance
- Mercor published the Benchmark Map to track AI capability coverage across the economy
Sources
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
How it worksHow Meta boosts engineering productivity by 3× using agents to propagate ad model techniquesHerald of AI
How Meta boosts engineering productivity by 3× using agents to propagate ad model techniques
Meta's new AI agent system adapts small improvements across thousands of ad models automatically, saving millions in computing costs.
How it worksA large AI model learned to fix its own brain without being retrainedHerald of AI
A large AI model learned to fix its own brain without being retrained
Researchers found a 27-billion-parameter model can edit its own internal settings to become more accurate — no human teacher needed.
Apps and productsJava bug found by AI could let hackers take over bank and hospital computersHerald of AI
Java bug found by AI could let hackers take over bank and hospital computers
A flaw in the widely used Java programming language lets someone gain full control of vulnerable machines.