Hands-on and licensed jobs: what the data shows and what it does not

What the data actually shows

The low scores for manual work are real, and they come from a consistent pattern rather than a rounding artefact. The verb families that make up hands-on and supervisory work sit at the bottom of the range set out in what AI is good and bad at: every activity beginning "collect" or "direct" scores zero.¹

But the occupations built from those activities are not all in the basement, and licensed professional work in particular is not. Pharmacists are in the top quintile of all 785 occupations. Elementary teachers and lawyers are above the median. Licensed professional work does not cluster at the bottom of this measure.²

Those zeros also carry a specific meaning that is easy to misread. A zero here is a frequency floor, not a test result: the activity fell below the study's coverage threshold, not below some standard of what AI can do. Reading your score sets out the cut-off.

The thing this score does not measure at all

Generative AI is one technology. Physical automation is another, and it is the one with an actual employment record attached.

Acemoglu and Restrepo's study of industrial robots in United States labour markets estimated that one additional robot per thousand workers cut the local employment-to-population ratio by 0.39 percentage points and local wages by 0.77%.³ Their local estimates imply roughly six fewer jobs per robot in the affected commuting zone relative to others. Adding assumptions about how commuting zones trade with each other, their model puts the national figure at 0.2 percentage points and 0.42%, or about 3.3 workers per robot, which they report as roughly 400,000 jobs across 1990 to 2007.³ They also found no offsetting employment gains in other local occupations.

The measure covers only industrial robots as defined by the International Federation of Robotics, a narrow slice of automation, with 38% of the installed stock in car manufacturing. The same paper puts the employment effect of Chinese import competition at roughly 2.5 times larger over the same period. And the total is measured in hundreds of thousands of jobs, across 1990 to 2007, not tens of millions.³

The honest summary for anyone in a trade: the number on your occupation page is not measuring the technology most likely to affect you, and the technology that has affected manual work did leave a real mark. Those authors go further and argue that occupation-susceptibility scores of the kind this whole field produces are "not informative about the equilibrium impact of automation", because they ignore how the rest of the economy responds.³

What a licence actually does

In 2024, 21.6% of employed people in the United States held an occupational licence, and 24.0% held a licence or a certification.⁴

The concentration is what matters. Licence coverage runs to 71.3% across healthcare practitioner and technical occupations, 62.2% across legal occupations and 49.2% across education, training and library occupations, against 18.0% in installation, maintenance and repair, 15.5% in construction and extraction, and 7.0% in computer and mathematical occupations.⁴

Licensing works through two different legal instruments, and the difference is the point. A title act reserves a job title, so an unlicensed person may do the work but not use the name. A practice act makes the work itself unlawful without a licence: in California, for example, practising medicine without a licence is a criminal offence carrying a fine of up to $10,000 and possible imprisonment.⁵ Scope-of-practice rules then operate at the level of individual tasks rather than whole jobs, which is why a nurse practitioner may prescribe independently in one state and only under a collaborative or supervisory agreement with another health provider in another.⁶

So a licence is a genuine barrier, and it is a legal one. It holds exactly as long as legislatures and regulators want it to hold, and it is unusually exposed to lobbying in both directions. Anyone treating a licence as permanent protection is making a political forecast, not a technological one.

What this means for you

If your occupation page shows a low score, it is telling you something true and narrow: over the nine months of 2024 this study covers, the consumer assistant it measured was not being used for the core of your work. That is worth knowing, and it is not a guarantee about the next decade, a statement about robotics, or a promise that your scope of practice will not be rewritten. If it shows a mid or high score, as it does for pharmacists, teachers and lawyers, then the overlap is real and the licence is not what is protecting the tasks.

One correction worth making, because it is a common assumption: the overlap in clinical and licensed work is not in the paperwork. Explaining medical information to patients or family members scores 0.83 and advising patients or clients on medical issues 0.66, while the administrative work people expect to be automated first, scheduling appointments, preparing medical or legal documents, recording information about legal matters, all score exactly zero in this dataset.¹

The overlap this dataset can see sits in the explaining rather than the filing, which is not the same as saying the filing is safe. The study's authors note that their data represents only one slice of the AI market, and they name the gap directly: programmers work inside AI-enhanced development environments, and legal, medical and financial work requires compliant AI tools.⁷ A zero here means the work was rare in this channel, not that it is out of reach.

Your occupation page lists which activities apply to you, and what to do with a score covers the practical side.