90% of US working hours sit on tasks that AI benchmarks barely measure

Mercor published a Benchmark Map of the US economy this week. It maps 29 AI benchmarks against O*NET, the US Department of Labor's catalogue of work. The main finding is that 90% of American working hours go on tasks with little or no coverage from existing benchmarks. Only 208 of 923 occupations have a single task with at least five benchmark prompts. Of 18,797 tasks, 924 are covered. Six of 22 job families have no covered occupation at all. Mercor puts the result plainly: it "makes it very hard to know how good AI models are at the work Americans actually do."

The measurement follows the money

Coverage follows a pattern. Mercor finds that tasks rated as exposed to AI, using the Eloundou et al. (2023) tags, are about six times more likely to appear on benchmarks than unexposed tasks. Even after controlling for exposure, tasks from occupations with larger wage bills are overrepresented. O*NET's domain experts rate how important each task is to its occupation. The tasks rated most important are 6 to 10 times more likely to be benchmarked than the least important.

The legal example shows the pattern clearly. The benchmarks in the index cover a lot of lawyers' tasks. They barely touch the tasks of legal secretaries or court clerks, which Mercor describes as similarly exposed to AI. In Mercor's words, "the long tail of smaller, middle-income knowledge work jobs appears undercovered."

Mercor calls the results provisional. Any incomplete list of benchmarks will understate coverage, and the company says it is hard to tell by how much. It left out saturated benchmarks partly because, as it writes, benchmarks "often overstate LLM performance in real-world deployments."

What the gap means for your role

Flip the finding around. Across most of the economy, any confident claim that AI can now do a given job rests on little or no direct measurement. That goes for the vendor deck, the board memo and the restructuring plan. For most roles, the 29 benchmarks Mercor mapped offer little or no direct evidence to back them. Mercor finds the benchmarks are full of tasks in software engineering and financial analysis.

This matters differently depending on where you sit. Lawyers, analysts and engineers work in the measured zone. Their tasks are scored all the time, and the most important tasks are scored most. In those roles, your value sits in the work around the benchmarked tasks. That means deciding what the draft should say, handling the client conversation and owning the outcome.

In the long tail, the benchmarks Mercor mapped barely touch your work. The person who measures it holds the evidence that everyone else in the room is guessing at.

AI does the routine work. You do the thinking. Sorting your own role into those two piles is now a job you have to do yourself.

Your Next Move

1. Benchmark your own work. Take five tasks from last week and give each one to the AI tool your organisation already approves. Note which outputs you could use without corrections, which needed your judgement and which failed. The full method, from listing a week's work to scoring it, is in our step-by-step method for assessing how exposed your role is.
2. Find your occupation on the map. Mercor's Benchmark Map shades each of the 923 O*NET occupations by coverage. If your work is heavily covered, shift your time towards the judgement and accountability around the scored tasks. If it is barely covered, your own test results are the only evidence anyone has. Write them down.
3. Ask for the evidence in the next restructuring conversation. When someone proposes automating part of your team's work, ask which evaluation supports the plan. If there is none, offer to run the test on your team's real tasks and present the results yourself.