Most of the scores used to say what AI can do come from a small set of jobs. When you lay the full American job list, 923 occupations, over 29 of the best public AI tests, about 90% of working hours fall on tasks with little or no coverage. Only 208 occupations have even one task that five or more test questions touch. The 29 tests are a sample of what exists, so the true share is probably a little below 90%. It is still nowhere near most of the work.
The tests go where the money and the attention are
People who build tests spend their effort where AI looks strongest and where the most pay is at stake. Tasks that AI is expected to handle are about six times as likely to be tested as other tasks. Among those, jobs with a large total wage bill get tested first. That is why there are deep tests for programmers and financial analysts and almost none for truck drivers or gas plant operators. With limited effort, it is a sensible order.
But it leaves out the long tail of smaller, middle-income office jobs: the legal secretary, the court clerk, the person who runs a clinic's front desk. A manager is quite likely to ask whether AI could take over a desk like that, and it is exactly where the public evidence is thinnest. Work with a physical part is further away still, because tests are run on a screen, and most jobs include some physical work. A warehouse shift manager is even further from a score than an office manager.
What the missing scores do to claims and forecasts
A vendor can truthfully say its tool does well on accounting tasks, and the claim can still say nothing about your accounts-payable desk, with its own supplier formats and approval rules. The score belongs to the tasks that were tested.
Forecasts have the same problem. Any prediction of job losses, such as the September forecast of 12% unemployment by 2030, needs a view of what AI can do in each kind of work. For most work that view is an assumption, and it can be wrong in either direction. Some jobs may be safer than feared. Others may already be within reach of the tools, with nobody having checked.
Build the test your work is missing
You do not have to wait for someone to measure your job. A practice manager at a twelve-person law firm wants to know whether AI can handle the routine parts of a legal secretary's week: formatting a court filing to the local rules, chasing missing documents, drafting hearing reminders. She picks 20 real tasks from last month, removes the client names, and gives each one to the assistant she already uses. Then she marks every answer as right, fixable or wrong.
At about ten minutes of checking per answer, that is an afternoon. At the end she knows what no public score could tell her: how often the tool is right on her firm's forms and her court's rules, and which kinds of task fail. She also decides what a wrong answer costs. A wrong date on a hearing notice is expensive and an awkward sentence in a letter is not, so she lets the tool draft letters and keeps a person on every deadline.
The gaps in public testing will be filled once someone has a reason to fill them, and legal secretaries will get scores eventually. Until then, twenty of your own examples are better evidence anyway, because they are your forms, your customers and your rules.