The task, not the job

"Will AI take radiologists' jobs" has been argued for a decade with almost no numbers attached. The argument stays unresolved partly because it is the wrong unit: a radiologist does perhaps forty distinct tasks, and reading a scan is one of them.

A job is a bundle

Automation does not act on job titles. It acts on tasks — repeatable units with an input, an output, and a person currently doing them.

A bundle can lose several tasks and still exist, sometimes with the remaining work getting more valuable rather than less. It can also lose one task and collapse, if that task was the only one anybody was paying for. Which of those happens is an empirical question about a specific bundle, not a general fact about technology.

Decompose a job you actually know

Pick a job you have done or watched closely. List every distinct task in it — aim for at least twelve. Be specific: "answers the phone" is not a task, "takes a booking over the phone and enters it into the calendar" is.

For each one, note roughly how many times a week it happens and how long it takes. Estimates are fine. The point is the shape of the distribution, and the shape will surprise you.

Which task is the better candidate?

Two tasks in the same job. Which is the stronger automation candidate, before any technology is chosen?

What share of the hours sit in the top three tasks?

Add up the weekly hours across your whole list, then the hours in the three largest tasks. Record the second as a percentage of the first.

This single number does more work than most commentary on the subject: it tells you whether the job is a few big repetitive blocks or a long tail of small varied ones, and those two shapes have completely different futures.

Explain the unit error

Explain to somebody why "will this job be automated" is a question that cannot be answered as asked.