What jobs will AI replace? The question is wrong, and that is why the answers are bad

Every few months a new list appears of the jobs artificial intelligence is about to eliminate, complete with percentages. The lists disagree with each other, they disagree with the previous year's lists, and following any of them would have led you somewhere unhelpful.

The problem is not that the forecasters are careless. It is that the question is built wrong.

Jobs are bundles, and bundles do not vanish at once

A job is not one activity. It is a collection of tasks held together by an employment arrangement, and those tasks differ enormously in how automatable they are.

Take any role and list what actually fills the week. Some of it is routine, high volume, and easy to check. Some of it is judgement in situations with no clean precedent. Some of it is coordination, persuasion, and knowing who to ask. Some of it is being accountable, which is to say being the person who carries it if the decision was wrong.

Automation does not arrive at the level of the job. It arrives at the level of those tasks, and it takes the routine and checkable ones first. The role does not disappear. It changes composition.

This is why job level predictions age so badly. They ask whether a bundle will vanish, when what actually happens is that the bundle gets rearranged.

Which tasks are genuinely exposed

The pattern is consistent, and it follows directly from what these systems do well.

Exposed tasks are high volume, follow a stable pattern, produce output whose correctness can be checked quickly, and carry a low cost when they are wrong. First drafts of routine documents. Standard replies. Extracting fields from forms that arrive in a predictable shape. Producing variants of something that already exists. Summarising material for someone who knows enough to spot a bad summary.

Notice that these are the same tasks that reward using the tools well. That is not a coincidence: exposure and usefulness are the same property seen from two directions.

What stays difficult

Work whose value comes from physical presence in an unpredictable environment. Work where somebody must personally carry the consequences of a decision, because accountability cannot be delegated to a system that cannot be held responsible. Work that depends on knowing a specific context that was never written down anywhere. And work where the hard part is deciding what should be done rather than producing it.

There is also a category people underrate: work where the value is that a person did it. Being listened to by a human being is not a slower version of being processed. In some settings it is the entire service.

The part that is not about technology

Here is where most coverage stops, and it is the part that decides what actually happens to people.

When a portion of a role gets absorbed, something has to fill the gap, and what fills it is a management decision rather than a technical one. The same capability can produce two completely different outcomes.

In one, the routine part shrinks and the role grows towards the harder work: more judgement, more context, more of what the person was actually good at. The job gets better and the person is more valuable.

In the other, the routine part shrinks and the role is redefined around supervising output at volume. The judgement work does not expand to fill the space, because judgement is slow and hard to measure. What expands is throughput.

Which of the two happens is not determined by the technology. It is determined by whether the organisation treats the freed capacity as room for better work or as room for more work. Technology forecasts cannot tell you that, which is a large part of why they keep being wrong about outcomes even when they are roughly right about capabilities.

What to do with this if it is your job

Do the task inventory on your own week. Write down what actually filled it, and mark each item honestly: routine and checkable, or judgement and context dependent.

If most of the week is in the first column, that is worth knowing early rather than late, and the useful response is to move deliberately towards the second, which usually means getting closer to the decisions rather than to the output.

If most of it is in the second, the realistic near term change is that the first column shrinks and you spend more time on the part that was already the hard part. That is a real change, and it is a different thing from replacement.

Why there is no number here

You will have noticed this piece contains no percentage of jobs at risk. That is deliberate. Every such figure comes from a model with assumptions inside it about which tasks count, how fast adoption happens, and how employers respond, and the assumptions do the work while the number gets the attention.

A figure whose basis is not published is not evidence, and repeating it does not make it one.