The most useful AI-and-jobs data this week is not a layoff count. It is a map of what nobody is measuring. Mercor, a hiring platform that builds AI benchmarks for real work tasks, has published a Benchmark Map of the US economy that asks a deceptively simple question: of all the hours Americans spend working, how many are covered by the AI benchmarks that decide which jobs get automated? The answer is about 10 percent.
🔍 THE BOTTOM LINE
Mercor’s index spans 923 O*NET occupations and 18,797 tasks, matched against 29 benchmark collections. Only 208 occupations — roughly one in five — have even a handful of benchmark prompts testing their real tasks, and just 90 percent of American working hours fall into tasks with “little or no coverage” from existing benchmarks. Coverage is heavily skewed toward the highest-paying jobs: computer and mathematical occupations lead (29 of 36 covered), business and finance roles are close behind (33 of 48), while healthcare practitioners (12 of 89), education (9 of 62) and science (15 of 60) lag far behind. The map does not say those jobs are safe. It says nobody has built the measurement to know — and that itself shifts how career risk should be read.
What the map actually measured
The Benchmark Map treats every O*NET occupation as a tile and shades it by how many of its tasks appear in the 924 benchmarks in Mercor’s index. A single benchmark prompt can map to multiple tasks, so the measure is deliberately generous — and even on that generous reading, whole job families go dark. Of the 22 families, six have no occupation with even five benchmark prompts covering its tasks.
The numbers by family: Computer & Math, 29 of 36 occupations covered; Business & Financial, 33 of 48; Legal, 5 of 7; Office & Admin, 21 of 51; Management, 21 of 56; Architecture & Engineering, 19 of 56; Sales, 6 of 22; Arts & Media, 13 of 40; Healthcare Practitioners, 12 of 89; Education & Library, 9 of 62; Science, 15 of 60.
The distribution is not random. Mercor ran a logistic regression of coverage probability against each occupation’s annual wage bill and its AI-exposure tag — the binary classification from Eloundou et al. (2023), the OpenAI-affiliated paper that produced the widely cited “80 percent of workers” exposure estimate. Coverage probability climbs steadily with the wage bill: the occupations with the biggest payrolls are the ones the benchmark community has mapped first. The healthcare worker, the teacher and the scientific technician are not absent because AI cannot touch their work. They are absent because their tasks have not been worth benchmarking yet.
Why the gap cuts both ways
For workers in well-covered fields — coding, financial analysis, law, administration — the map confirms what hiring data already suggested: these are the tasks AI can do and be evaluated at, which is exactly why so much of the displacement evidence concentrates there. Goldman Sachs economists found call-centre employment 39 percent below trend in the US with software and consulting close behind, and the Hiring-not-firing research graded tasks the same way Mercor does — by what they require of a human — and found the adjustment lands in hiring, not layoffs.
For everyone else, the honest reading is more cautious than optimistic. An uncovered task is not an automated-proof task; it is an unmeasured one. The Stanford and King’s College work on the senior-junior hiring split found effects showing up in fields nobody had benchmarked in advance. What the gap does tell workers is that the visible, quantified risk is where the measurement investment went — and that the biggest benchmark players are hiring domain experts to close exactly this gap, which has turned into one of the stranger job categories of 2026, from Mercor’s own expert bench to the dedicated AI-trainer market.
The practical signal for career planning follows from the economics of benchmarking. If capability measurement concentrates first on the tasks that pay the most, then the fields where “AI is coming for you” is a tested claim are also the fields where employers have the tools, dashboards and procurement paperwork ready to act on it. In the unmeasured 90 percent, adoption is slower, messier and more dependent on local trust — not because AI is worse there, but because nobody has yet packaged it as a product.
The New Zealand read
New Zealand’s employment mix amplifies the point. SEEKNZ data shows agriculture, healthcare, education and construction are among the country’s largest occupational groupings — all sitting in Mercor’s sparsely measured families, with Healthcare Practitioners at 12 of 89 covered and Education & Library at 9 of 62. The same SEEK reports that keep showing AI skills breaking records — the most recent showed mentions in 4.2 percent of job ads, near double Australia’s rate — are measuring demand for AI skills, not exposure of existing work. A Wellington policy analyst’s tasks sit in a family with 33 of 48 occupations covered; a Tauranga builder’s sit somewhere with no prompts at all. The first group can see its score. The second cannot — yet.
That asymmetry is worth holding onto for the next jobs panic, in either direction. When a benchmark says AI now passes the bar exam, it is saying something about a measured field. When a report says trades and care work are “safe”, the same epistemic caveat applies — those claims are running ahead of any measurement.
What job-watchers should watch next
Mercor describes the map as the first step toward a continuously updated resource. If the coverage frontier moves the way the wage-bill regression predicts, the next families to gain benchmarks will be the mid-wage, high-exposure ones — office admin and management rather than brain surgery. Each new benchmark suite effectively announces where vendors think a market exists. For workers, the map is a crude early-warning system: your job’s family getting benchmarked is the labour-market equivalent of a flood gauge going in upstream.
FAQ
What share of US working hours are covered by AI benchmarks? Mercor’s Benchmark Map finds roughly 90 percent of US working hours are spent on tasks with little or no coverage from existing benchmarks — only about 208 of 923 occupations have tasks with at least five benchmark prompts.
Which jobs are best covered by AI benchmarks? Computer and mathematical occupations (29 of 36 covered) and business and financial roles (33 of 48) lead. Healthcare practitioners (12 of 89), education (9 of 62) and the sciences (15 of 60) are among the least covered.
Does being unbenchmarked mean a job is safe from AI? No. It means the tasks have not been systematically measured. The gap is a statement about measurement, not capability — but it does mean employers have fewer ready-made tools for automating those roles today.
Why does benchmark coverage follow wages? Mercor’s regression ties coverage probability to each occupation’s annual wage bill: the benchmark community has historically invested in measuring the biggest labour markets first, which are also the highest-paying ones.
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
