No job lands in our bottom band. Here is what the ranking does show.

Every job page on this site answers one question about one job. The rankings put all 427 answers in a single list, and a few things only show up there, because they are about the shape of the whole list rather than any job in it.

All of it comes from the same place as every score here: the International Labour Organization’s ratings of 3,265 everyday tasks for how much of each today’s AI tools could do. A job’s score is 100 minus the average across its tasks. Nothing below is a forecast.

The bottom band is empty

We sort scores into five bands of twenty points. The lowest, “Being rebuilt around AI”, is for jobs under 20: jobs whose tasks average above 80 out of 100 for what AI could do.

Not one of the 427 jobs is in it. The most exposed job the ILO rated, data entry clerks, averages 70 across its tasks and scores 30. To get below 20 a job would need nearly every one of its tasks rated very highly, and no job in the study is that uniform. Even the most routine desk work has some tasks the ratings score low.

The bands are our own round numbers, chosen for reading, and the site says so wherever they appear. An empty band is still worth knowing about: it means the words “being rebuilt around AI” describe no job on the list today.

One group sits apart from all the others

The ILO sorts jobs into nine broad groups. Eight of them average between 61 and 85. The ninth, clerical support workers, averages 47.

The gap is wider than it sounds. The best-scoring clerical jobs, stock clerks and filing clerks, reach 63. That is below the average of seven of the other eight groups. Of the ten lowest-scoring jobs on the whole list, seven are clerks: data entry, typists, statistical and insurance clerks, bookkeeping clerks, payroll, personnel, and the ILO’s catch-all for clerical work it doesn’t name. The other three are contact centre salespeople, financial analysts and securities dealers.

If there is one place where the task ratings point the same way, job after job, it is here.

The top is a twelve-way tie

Twelve jobs share the top score of 91. They are bricklayers, vehicle cleaners, water and firewood collectors, hunters and trappers, crop farm labourers, building structure cleaners, garbage and recycling collectors, civil engineering labourers, building construction labourers, forestry labourers, sweepers, and the ILO’s group for building frame trades it doesn’t name separately.

Their tasks average about 9 out of 100 for what AI could do. At that level the ratings have nothing left to separate them with: there is simply very little in any of these jobs that a language model is good at.

That tie is the least certain number on the list

Every one of those twelve jobs carries the robotics caveat, and that is not a coincidence. The ratings measure AI software, which is why physical work scores so well. They say nothing about robots, and for a bricklayer or a bin collector robots are the bigger question. Across the whole list, 181 jobs carry the same caveat.

So the top of the ranking is where the score is most confident about the question it asks and least useful about the one people usually mean. For two of the twelve, building structure cleaners and sweepers, there isn’t even a close match in the US job data, so the site can’t say how hands-on the work is. Their pages say that too.

We once built a score for how far robots have reached into physical work. It was a guess, so we took it out. How it works explains why the gap is left empty rather than filled.

A group name tells you very little

The group averages hide very wide spreads. Technicians and associate professionals run from 37, for securities and finance dealers, to 88, for religious associate professionals. Professionals run from 38, for financial analysts, to 87, for creative and performing artists.

“Professional” is a description of training and status, not of how exposed the work is. The same is true of every group bar clerical support. If you want to know about a job, look at the job.

What the whole list adds up to

Two numbers, held together. The typical job holds up at 73, and 329 of the 427 are in the two top bands, “strongly” or “mostly holds up”. Most work, measured task by task, is mostly out of reach of today’s AI tools.

And a caution in the same breath: much of that top end is physical work, which is exactly where the score is blindest. The ranking is a careful answer to a narrow question. Find your job on it, and then read its tasks, because that is where the answer actually is.