Which jobs are safe from AI? The dividing line is not the industry — it is whether the work is augmented or automated

The usual answer to this question is a list of industries, which turns out to be the wrong unit.

The Stanford Digital Economy Lab's analysis of ADP payroll data — covering millions of US workers — found employment falling for early-career workers in the most AI-exposed occupations, by roughly 16 per cent in relative terms for those aged 22 to 25. In the same dataset it found something that gets much less attention: **employment growing in occupations where AI is used to augment workers rather than automate their tasks.**

That is the dividing line. Not sector, not collar colour, not whether the work involves a computer. Whether the technology substitutes for the task or assists someone doing it.

The distinction is practical rather than philosophical. A task is a substitution candidate when it has a well-defined input, a well-defined output, and a correct answer that can be checked. Routing a support ticket. Transcribing a call. Producing a first draft from a template. These are the tasks where a model can be dropped in and measured.

Work resists substitution when the difficulty is in deciding what should be done rather than doing it: where the inputs are ambiguous, where being wrong is expensive, where accountability has to attach to a person, or where the job's actual content is physical presence and judgment in an unpredictable environment. In those cases the technology tends to attach to the worker rather than replace them, and the payroll data shows employment rising rather than falling.

Two cautions on how to use this.

First, the safe unit is the task, not the job title. Most jobs are bundles of tasks and the bundle is rarely uniform — a role can have several highly substitutable components and a core that is not substitutable at all. What changes for those roles is composition, not existence. Asking whether your job is safe produces a worse answer than asking which parts of your week are the checkable, well-defined ones.

Second, the measured effect so far is concentrated by age rather than spread evenly. The clearest signal in the data is in early-career hiring, which suggests that the immediate risk sits with the entry-level version of an occupation rather than the occupation itself. Junior roles are often deliberately composed of exactly the well-defined, checkable tasks that substitute most easily — which is a problem for the pipeline into the senior roles that are not at risk.

What this evidence cannot tell you is where the line sits five years out. The augment-versus-automate boundary is a description of how the technology is being deployed now, not a fixed property of the work. Tasks move across it as capability changes, and nothing in the payroll record forecasts which ones move next.