Start with the work. Then ask what changes.
A better starting point than predicting which jobs disappear.
The wrong unit of analysis?
A job title is a convenient label. It is also a surprisingly poor description of what someone does all day. A role might combine research, judgment, coordination, persuasion, production, and the occasional spreadsheet rescue. AI can affect each of those activities differently.
Follow one task
Imagine a team that uses AI to prepare the first draft of a report. That is an illustrative scenario, not a finding about every team. The draft might arrive faster. But somebody still needs to choose the question, check the evidence, resolve contradictions, and decide what to do next. The useful research question is not simply whether the tool can write. It is where the work moves after the first draft gets cheaper.
Look for what gets added
Automation can remove an activity while creating a different one. Verification, exception handling, and explaining a decision may become more important. Or a process may get simpler overall. Both are possibilities worth investigating. Counting only the minutes saved in the first step can miss time spent downstream.
Three levels to watch
At the individual level, ask what people actually do differently. At the team level, ask who makes decisions and where handoffs happen. At the workplace level, ask what those new patterns require of the environments people use. A change at one level does not automatically prove a change at the next.
The question we want to follow
When an AI tool enters a workflow, which tasks disappear, which expand, and who takes responsibility for the result? That question is concrete enough to study and broad enough to matter to a new graduate, a manager, or someone planning an entire workplace.
A small exercise
Choose a recurring task. Write down its inputs, its output, who reviews it, and how you know it was done well. Introduce AI at one step. Then compare total time, quality, rework, and the experience of the people involved. Keep the original baseline. The goal is to find out what changed, including the parts that did not.