Nine occupations, 201 rated tasks and 75 job titles. The profession everyone else comes to for advice about AI and work has not, so far, been told very much about itself.
Exposure means published research judged that an AI could halve the time a task takes, at the same quality. It does not mean the task stops needing a person, and it is not a prediction about any job.
Read the solid block as work an AI can halve today, the pale block as work that would need software building first, and the outline as work it cannot speed up at all. In the average job needing little preparation, that outline covers 76 per cent of the role. Here it never gets above 38.
Hold a degree is the share of people actually doing the job who have a bachelor's or higher, from O*NET's survey of job holders. It is a description of who is there, not a requirement. Twenty-seven per cent of HR Specialists do not have one, and neither do sixty-five per cent of HR Assistants. Samples are small, between 17 and 41 people per occupation, so read these as roughly rather than exactly.
One bar per occupational group, measuring the share of work an AI cannot speed up at all. Short bar means little protected work. The People profession is 3rd of 23 from the exposed end. Only computing and business roles have less protected work.
Share of an occupation group's listed tasks rated as impossible to halve. Construction sits at 90 per cent and Production at 86. The People profession has less protected work than all but two groups in the economy.
This chart measures one thing: whether a language model can halve the time a task takes. It is not a ranking of which careers are safe, and it should not be used to choose one.
Construction sits at the protected end because a model cannot lay bricks. That says nothing about wages, demand, physical toll, or the entirely separate question of what robotics does to the same work. A job can be untouched by this chart and still be a poor bet, and the reverse.
One square is one task. Rows are different lengths because occupations list different numbers of tasks, and none of them is stretched to fill the width.
That length is the point. A percentage drawn from twelve tasks is a much rougher measure than one drawn from twenty-eight, and reading it as a bare number hides that completely.
Training and Development Managers is the shortest row here. Its 8 per cent of protected work is a single square out of twelve. Labor Relations Specialists, at the other end, has 28.
Hover any square to read the task. Ordered by band within each row so the rows can be compared; O*NET gives the tasks no meaningful order of its own.
Training and Development Managers have 8 per cent of their listed tasks beyond an AI's reach. Eighty-three per cent sits in the middle band, waiting on software somebody has not built yet. That is the thinnest margin of any People occupation, and it belongs to a role most people would assume was safe because it has direct reports.
Seniority is doing nothing here. Compensation and Benefits Specialists, a non-management role, have 38 per cent of their work beyond reach, nearly five times as much protection.
Caveat worth keeping in view: that occupation has only 12 rated tasks, the smallest sample of the nine. Treat the direction as real and the precise figure as soft.
Look along the middle band. For seven of the nine occupations it is the largest of the three, and for Training and Development Managers it covers 84 per cent of the role.
That band does not mean the work is being done by an AI. It means a researcher judged that purpose-built software could halve it, and that the software may or may not exist. So the honest reading of this chapter is not that most People work is automated. It is that most People work sits behind a tool somebody would have to build first, and whether that happens is a decision about investment rather than a fact about capability.
That is also the softest of the three bands. "It is easy to imagine additional software" is a judgement call, and two competent raters disagree about where it starts more than they disagree about anything else in the rubric.
Every task in this profession rated beyond an AI's reach has a person on the other side of it. That is not a coincidence, it is the rubric: it says outright that work requiring a high degree of human interaction is classified as having no exposure.
Read the two columns together. On the left, every entry is a person in a room. On the right, every entry is a document, a record or an explanation. If you want to know which half of your week is exposed, that is the test, and you do not need this report to apply it.
Every figure here measures whether a task could be finished in half the time. A task that takes half as long still needs somebody doing it, and nobody surveyed a single employer about what they intend to do.
This must never be used to choose who to make redundant. An exposure figure is not an objective, fair and consistently applied selection criterion, and it would not survive a tribunal being asked about it.
55 per cent of this profession's work sits in the band that needs software building first. Read as a warning that is the largest share of the three and the most uncomfortable. Read as a list it is something else entirely.
Nothing in that band is waiting on a breakthrough. It is waiting on somebody building the tool. And the reason nobody has is a gap in who knows what: the people who understand these tasks well enough to specify them cannot usually build, and the people who can build have never run a grievance, an induction or a pay review.
A People professional who learns to build sits in that gap on their own. Not competing with engineers, and not competing with peers who have learned to write better prompts. Building the thing that was on the list, for the function they already understand.
That is also the honest answer to "how do I stay valuable". Not learning AI in the abstract, which everyone is doing. Picking one item off a 55 per cent backlog, scoping it properly, building it, and being able to show what it saved.
There is no data here on what People teams actually do with AI. A dataset exists that appears to measure it, and an earlier draft of this chapter used it. It has been removed. Its two relevant columns are not defined in any documentation we could find, and the figures Anthropic have published cannot be reproduced from the file under any aggregation we tried. Until somebody can say precisely what the number counts, it does not belong on a page like this one.
So everything above is about what published research judged to be possible in 2023. None of it is evidence about what anyone is doing. That is a real limitation and it is the next thing to fix.
The exposure ratings themselves cover all 201 tasks in these nine occupations, with no gaps.
Every question is about something observable: whether the purpose is written down, where ideas arrive, who decides, whether anything was measured before the change. It gives you a score out of 12, the band it falls in, your biggest gap and what to address first. There is a competency scorecard for your AI Champions in the same place.
Free. It asks for your name and a work email, because the result is emailed to you as a record.
It is a check on your own programme, not a survey. Nothing you put in appears anywhere on this site.
Exposure ratings from Eloundou, Manning, Mishkin and Rock,
GPTs are GPTs: Labor market impact potential of LLMs, Science 384,
1306–1308, 2024, used under the MIT licence. Observed use from the
Anthropic Economic Index, used under CC BY.
This page includes information from the O*NET Database by the U.S.
Department of Labor, Employment and Training Administration (USDOL/ETA).
Used under the CC BY 4.0 license. O*NET® is a trademark of USDOL/ETA.
People Team AI has modified all or some of this information. USDOL/ETA has
not approved, endorsed, or tested these modifications.