A New Research Map Shows 90% of US Work Hours Have No AI Benchmark Coverage. That Gap Is a Task-Audit Service Waiting to Be Sold.

Everyone citing an AI benchmark is quoting a number that probably has nothing to do with your buyer's actual job.

Mercor just mapped the gap, and it is bigger than most people assume.

What the research found

Mercor took the US Department of Labor's O*NET taxonomy — 923 occupations, 18,797 associated tasks — and checked how many of those tasks are actually represented in 29 widely used AI benchmarks.

The result: "90% of American working hours are spent on tasks with little or no benchmark coverage."

More specifically: "only 208 of 923 occupations have a task with at least 5 benchmark prompts, and 6 of 22 job families have no such occupations." Across the whole taxonomy, just 924 of 18,797 tasks were covered at all.

And the coverage that does exist isn't evenly spread. Tasks rated as AI-exposed are roughly 6 times more likely to be benchmarked than tasks that aren't. Even controlling for that, tasks in occupations with larger wage bills are overrepresented — while middle-income knowledge work stays undercovered.

The business idea

Every vendor pitch about "AI can do your job" is backed by benchmark numbers that, per this research, almost certainly never tested your buyer's actual tasks.

That is not a reason to distrust AI. It is a reason to sell verification.

Offer a Task Coverage Audit:

• Pull the real task list for a role or team (O*NET gives you a ready-made structure, or build one from the client's own process docs)
• Run each task against the AI tools the client is considering, or already paying for
• Score every task as covered, partial, or uncovered — with a short note on why
• Hand back a prioritized list: which tasks to automate now, which to revisit next quarter

This turns an abstract "is AI ready for us?" question into a concrete, task-by-task answer the client can act on immediately.

Why middle-income knowledge work is the wedge

Mercor's data points straight at who's underserved: occupations with big wage bills get benchmarked (because AI labs and vendors chase those buyers), but ordinary office, admin, ops, and mid-tier analyst roles don't.

That means the people most likely to ask "will AI actually help my team?" have the least public evidence to go on — and the most reason to pay someone to find out.

How to package the offer

1. Discovery audit — a fixed-price engagement mapping 15-30 real tasks for one team or role

2. Pilot automation — pick the 3-5 clearly "covered" tasks and build the first working agent or workflow

3. Quarterly re-scan retainer — re-test the uncovered and partial tasks every quarter as new models ship, since coverage is a moving target, not a fixed answer

Bottom line

The industry has been selling AI capability off benchmarks that, by Mercor's own count, miss 90% of actual working hours. Anyone willing to test against the real task list instead of the leaderboard has a service buyers can't get anywhere else right now.

Sources:

https://www.mercor.com/research/labor-index/benchmark-map/