Break every role into tasks, then score each one
Asking which jobs AI will replace is the wrong question. A job is a bundle of tasks that were grouped together because a person was the only one who could do them. AI does not take jobs. It takes some of the tasks inside them, and which ones depends on the job.
The exercise
Pick the function that involves the most coordination, the most people or the most cost. List every role in it. For each role, list the tasks it actually performs, not the ones in the job description. A function usually produces several dozen distinct tasks.
Sort each task into one of four kinds: judgment (deciding when the rules do not settle it), pattern (recognising something seen before), coordination (moving information or work between people), and creation (producing something new).
Then score each task from 1 to 5 for how much of it a system could do today. 5: a system can do it now. 4: a system can do it with a person checking. 3: worth a trial. 2: a system can help, a person decides. 1: stays with a person.
What the scores tell you
Tasks scored 4 and 5 can move first. Tasks scored 3 are the next trials. Tasks scored 1 and 2 stay with people, and are often the parts of the job people value most.
Most teams discover the same thing: a large share of the tasks are coordination. Chasing approvals, copying figures between systems, preparing reports for meetings, asking who has the latest version. That is where the first gains are, and it is work almost nobody will miss.
Tasks, not headcount
The matrix also shows how each role changes. A role that loses its coordination tasks keeps its judgment tasks, and can take on more of them. Plan that change for each person before any task moves. The matrix is a map of the work, not a list of people to let go.