The common framing of this subject says that AI will come for the bottom of the labour market first, the way automation is assumed to have done before. This release can test that, because every occupation in the index carries a standard SOC code, and so does the Bureau of Labor Statistics wage file.
Matching the two gives 767 occupations covering 143.6 million US jobs, about 93% of national employment. What the match shows is the opposite of the common framing.
The bottom of the pay scale is the least exposed part of it
Splitting employment into five equal bands by median annual pay, then taking the employment weighted mean exposed share inside each band:
| Pay band | Median annual pay | US employment | Exposed share of task load |
|---|---|---|---|
| Lowest fifth | $30,890 to $37,390 | 30.0M | 23.9% |
| Second fifth | $37,450 to $46,420 | 27.5M | 32.6% |
| Middle fifth | $46,460 to $62,890 | 29.0M | 28.8% |
| Fourth fifth | $62,960 to $91,240 | 28.5M | 39.2% |
| Highest fifth | $91,640 and above | 28.7M | 37.4% |
The lowest paid fifth of American employment has the lowest exposed share of any band, by a margin of nine points over the fourth fifth. Thirty million people at the bottom of the wage distribution hold the work this index rates as hardest for a current system to produce.
Across all 767 occupations the rank correlation between median pay and exposed share is +0.34. Against the logarithm of pay, which is the form wage data usually takes, the straight correlation is +0.23 for exposed share and minus 0.36 for untouched share. The better paid the occupation, the less of its work sits beyond current capability.
Why the middle dips
The middle fifth breaks the pattern, and the reason is worth stating rather than smoothing away. That band contains the two extremes of the index at once. It holds bookkeeping and auditing clerks at 59.4% exposed and secretaries at 53.8%. It also holds 2.1 million heavy truck drivers at 25.1% and 1.1 million construction labourers at 5.1%. The band average is a collision between the most exposed clerical work in the economy and some of the least exposed physical work, not a description of a typical job inside it.
This is the general problem with any single exposure figure for a group. It is also why this index publishes the task level underneath every occupation score.
Two different walls
The reason low paid work scores low is not that AI finds it difficult in some abstract sense. It is embodiment. Dishwashers score 0.0% exposed. Food preparation workers score 6.7%. Maids and housekeeping cleaners score 5.4%. Every task in those occupations requires a body moving objects in a room, and the index rates capability against what a generally available system can produce, not against what a robot might one day do.
At the top of the distribution the wall is different. The 35 occupations that fall in the assisted band employ 5.1 million people at an employment weighted mean median wage of $110,098, against $74,572 for exposed occupations and $52,359 for untouched ones. Assisted is the most expensive band in the index. That is not an accident of the scoring. Assisted means capability is high and friction is high with it, and the frictions that produce it are accountability, presence and verification cost, which is a fair description of what a professional licence buys.
The correlation between assisted share and log pay is +0.55, the strongest relationship between any part of this index and any outside variable we have tested.
Lawyers sit at 22.0% exposed on a median wage of $159,670. Registered nurses sit at 18.6% on $97,550. Neither number reflects a machine that cannot draft a brief or read a chart. Both reflect a person who has to sign.
What this does and does not support
It supports one claim: exposure in this index is not concentrated on low paid work, and any argument built on the assumption that it is should be checked against the numbers above before it is made.
It does not support the claim that high paid workers are at greater risk. Exposure is not displacement. A high exposed share means a machine can produce the work product with little structural obstacle in the way. Whether anyone stops paying a person to do it depends on firm economics, liability, procurement cycles, union agreements and inertia, none of which this index measures and none of which follow from a task score.
The honest summary is narrower than either headline. Capability has climbed the wage ladder faster than anyone expected it to, the top rungs are held by friction rather than by difficulty, and the bottom rungs are held by the fact that a model has no hands.
Pay and employment figures are now published on every occupation page in this index and are sortable in the full rankings table, so every number above can be checked one occupation at a time.
Wage and employment figures: BLS Occupational Employment and Wage Statistics, national cross-industry, May 2025. Exposure figures: release v2026.Q3 of this index. Occupations are matched on six digit SOC code; where ONET publishes several detail codes under one SOC, their scores are averaged before matching.*