The task is not the job.

Picture a job description cut into strips, one task to a strip, and laid on a desk beside the envelope they came in. Draft routine replies to customers. Summarize case notes. Check documents against policy. Approve the exceptions. Call the customer who is upset. Several strips now carry a small mark: a current AI system can do this, or most of it. It is tempting to count the marked strips, divide by the total, and write the answer on the envelope's payroll line. But the envelope holds something the strips do not describe: the service the job delivers, how much of it customers want, and the work that keeps the pieces together. Until those are examined, the payroll line is still a question.

On Saturday, April 25, OpenAI published a framework that makes the same point at the scale of the economy. Instead of asking only which jobs AI could perform, it asks three questions about each occupation: can AI do a meaningful share of its tasks, is a person still central to delivering the work or answering for it, and if AI lowers the cost of the service, will demand grow enough to absorb the gain? Applied to 921 occupations covering about 148 million U.S. jobs, the questions sort work into four groups: jobs at higher automation risk, jobs that will reorganize, jobs that could grow with AI, and jobs with less immediate change. The authors are plain about what the groups are not: "These categories are not predictions that a particular percentage of jobs will disappear."¹

My argument is that technical task exposure is one input to an employment outcome, not a forecast of it. Whether the number of people in a role shrinks, holds or grows also depends on demand, on cost, and on which parts of the service still need a person. A capability score should never be converted directly into a staffing target. This is not a reassurance. Some work will be automated outright, and an outcome that is stable on average can still be a loss for the people in a particular role. It is an argument about what a workforce plan has to look at before it writes a number.

What exposure measures

Exposure, in this kind of research, is a statement about tasks. It says that a technology can carry out a specified activity, under stated conditions, well enough to matter. The OpenAI report builds its exposure measure from earlier research that judged, task by task, which work activities large language models could affect, and it adds a second measure of where people are actually using ChatGPT for work.² Both are useful. Neither says what an employer will do.

Between "can" and "will" sit several ordinary business facts. The system has to be reliable enough on this firm's cases, not on a benchmark. It has to be integrated with the systems the work already runs through, and its errors have to be caught by someone. It has to be cheaper, all in, than the way the work is done now. The OpenAI article lists these frictions itself: employers still have to redesign workflows, integrate AI with existing systems, manage errors and regulatory risk, and decide whether automation is actually cheaper than people.¹ An exposure score is a measure of possibility. It is the start of an analysis, and a good one, but it is not the analysis.

The arithmetic of hours

The clearest way to see why is to do the arithmetic that sits under any staffing number. The hours a team needs are, roughly, the volume of service it delivers multiplied by the hours each unit of service takes, plus the work that does not scale with volume. Exposure acts on only one of those terms.

Take an invented team that handles 1,000 cases a month at 2 hours a case, plus 400 hours a month of coordination, training and supervision. That is 2,400 hours. Now suppose AI halves the hours per case, to 1. If volume stays at 1,000, the case work falls to 1,000 hours and the total to 1,400, a drop of about 42 percent. If cheaper, faster service draws more customers and volume rises by half, the total is 1,900. If volume doubles, the total returns to 2,400 hours, exactly where it began, though the work inside it has changed. Figure 2 lays the three cases side by side.

None of these is a forecast. They are conditional statements: if hours per case halve and volume does this, required hours do that. Their value is in showing which number the outcome depends on. The halving was the same in every column. What differed was demand, and the exposure score says nothing about demand.

What the arithmetic leaves out

The OpenAI framework's second and third questions are, in effect, the two terms the simple sum hides. The third is demand. The report calls the link between productivity and employment ambiguous: when a service becomes cheaper, people often buy more of it, and employment can rise. Its examples of jobs that could grow are services many people find expensive or hard to reach today, such as tutoring, counseling and personal financial advice.¹ The economist David Autor made the general case a decade ago: automation substitutes for labor, but it also complements it and raises output in ways that can increase the demand for labor, and commentators tend to overstate the first effect and overlook the second.³

The second question is about the person. The report sorts what it calls human necessity into three kinds: regulatory and accountability (a licensed person must decide or answer for the result), relational (the service depends on trust, care or teaching) and physical (someone has to be there).² Its example is the court reporter, an occupation near the top of some exposure rankings, where in most jurisdictions a person is still legally required to certify the transcript.²

But the report is careful not to treat a necessary person as a safe job. A lawyer may remain necessary for accountability and representation, it notes, yet if each lawyer can handle much more work and lower prices do not raise demand much, employment can still fall.² That is the arithmetic of Figure 2 again, in the fixed-demand column. Necessity keeps a person in the job; it does not set how many people.

And the fixed row matters more than its size suggests. Coordination, handoffs, review of the system's output, the exception nobody planned for: these are often the tasks that grow when the routine ones shrink, because someone has to check the work and own the result. Which of these binds in a given business is an empirical question about that business. A model of the occupation cannot answer it.

Read the framework as a map

The report is unusually frank about its own limits, and they belong in any plan that borrows from it. Two of its measures are estimates produced by a model. The authors asked GPT-5.4 to classify each occupation's need for a person, and to estimate how much demand for its output would rise after a 10 percent fall in price over two to three years.² They call the demand figure the least directly observed part of the framework: a structured approximation, they write, not a substitute for high-quality causal evidence. They add that the framework simplifies how tasks are organized inside jobs and may overstate substitution, and that it does not fully capture effects on wages, spillovers between occupations or the wider economy.²

The report also tests its groups against recent data, with an instructive result. From the first quarter of 2024 to the first quarter of 2026, unemployment among people in the occupations it expects to change least rose by 0.6 percentage points, more than the 0.2 points for the group it rates at higher automation risk.² The authors do not read this as proof that AI has had no effect, and neither should anyone else. Many forces move unemployment, and some effects show up first in hiring, entry-level openings, wages or the mix of work rather than in layoffs.¹ But it is a useful caution in both directions: an exposure ranking is not evidence that jobs have already been lost, and a quiet unemployment rate is not evidence that a role is unchanged.

A worksheet, not a percentage

What does this mean for an organization deciding how many people it needs? It means the plan should be a scenario sheet rather than a single reduction percentage. The sheet needs a row for demand, with more than one assumption about how customers will respond to faster or cheaper service. It needs the task mix, including the work that does not scale with volume and the new work of supervising the system. It needs a measure of service quality, because hours saved by lowering quality are not savings. It needs the costs of the transition: integration, error handling, retraining. And it needs a time horizon, because the column that is right for next quarter may be wrong in three years.

It also needs a line for the people. Total employment can hold steady while particular workers lose: the routine work that used to be a newcomer's way in, the hours of part-time staff, the wages of a role whose skilled core has been automated. These changes often arrive before any change in headcount. The March 10 dispatch on entry-level work as training is one example; here the point is simpler. A plan that reports only the total has not looked at who bears the transition.

Objections, considered

Some work will simply be automated. Yes. The framework itself places a group of occupations at higher automation risk, where the person is not central and demand is unlikely to absorb the gain.¹ Nothing here guarantees that a job survives. The point is the reverse: because outcomes differ, a plan has to find out which case it is in rather than assume.

Demand does not always respond. Markets saturate, prices may not fall, and a firm may keep the productivity gain as margin rather than pass it to customers as lower prices or more service. Then the fixed-demand column is the right one. That is a reason to model demand explicitly, not to assume it will rescue employment.

New jobs may not go to the people who lost the old ones. Work created by cheaper service may need different skills, or sit in a different place, from the work that was cut. History does not promise that a new technology will repeat the pattern of older ones. This is why the distribution line on the worksheet matters as much as the total.

This framework comes from a company that sells AI. It does, and that deserves scrutiny. A vendor has reasons to stress that its product reorganizes and expands work rather than replacing it. The distinctions, though, are older than the vendor: they come from the labor economics of tasks, complements and demand. A reader can use the distinctions without accepting the vendor's percentages, which the vendor itself declines to call forecasts.

Automating part of the work does not specify the size of the workforce.

Before the next headcount plan

• Which tasks change? List them by name, with the conditions under which the system does them well enough, and the ones that remain.
• What happens to demand? Write at least three assumptions about how customers respond to faster or cheaper service, and carry each to a number of hours.
• What grows when the routine shrinks? Count the review, coordination and exception work the new process adds.
• Who bears the transition? Name the roles, entry routes and schedules that change, even if the total does not.
• When will you look again? Set a date to compare the scenario you chose with the volume and hours you actually see.

Back to the envelope on the desk. The marked strips are real information, and a planner who ignored them would be making a mistake. But the payroll line is written on the envelope, not on the strips. It waits on the service: how much of it customers will want, what it will cost to deliver, and who has to be there when it goes wrong. Rewrite the job after that, and the number on the envelope will mean something.