A Job Is a Bundle of Tasks

When ATMs spread through American banks, they looked like the end of the bank teller. A machine could hand out cash and take deposits, which sounded like most of the job.

The opposite happened, for a while. The number of tellers needed to run an average urban branch fell from 20 to 13 between 1988 and 2004. But ATMs made each branch cheaper to run, so banks opened more of them to compete for customers: urban branches grew 43 percent in the same period. Fewer tellers per branch, many more branches. Teller jobs held up.

Then mobile banking arrived, and this time the outcome flipped. You could deposit a check from your kitchen, so there was less reason to visit a branch at all. The Bureau of Labor Statistics counted about 497,000 tellers in 2016 and about 341,000 in 2023, and it projects a further 13 percent decline by 2035.

Same occupation, two automation waves, opposite results. The technology didn't decide either outcome. Demand did.

That's the idea this piece is about, and it's the one most AI-and-jobs arguments skip. AI takes over tasks, and a job is a bundle of tasks. What happens to the job depends on three things: which tasks remain, how much more people buy when the work gets cheaper, and how the leftover tasks get recombined into jobs. The share of tasks a machine "could do" tells you almost nothing about any of them.

What a job actually is

An occupation is a label for a bundle of tasks that one person gets paid to do together. A paralegal reads records, drafts documents, looks up law, talks to clients, files with courts and keeps a calendar. Those tasks sit in one job because they share context, a client and a supervisor, and doing them together is cheaper than splitting them across six people.

Technology doesn't arrive at jobs. It arrives at individual tasks. Labor economists have worked this way since 2003, when David Autor, Frank Levy and Richard Murnane showed that computers replaced people in tasks that follow explicit rules and helped people in problem-solving and complex communication, and that this happened even inside jobs that kept their names. A job title can stay put while what the job consists of changes underneath it.

So when you read that a model can do 40 percent of the work in some job, the useful response is to ask which 40 percent, and what happens to the other 60.

One paralegal, ten tasks

Take a paralegal at a mid-size firm and break the job into ten tasks. The task names follow the official US occupational database, O*NET; the time shares and the judgments about AI are invented for illustration. A real analysis would time the work.

Scroll sideways to see the whole figure

Sort the tasks three ways.

AI produces, a person checks. Summarizing records, first-pass document review, drafting routine documents: 36 percent of the time. A current model can produce a usable first pass of each.

AI assists. Tailored drafting, legal research, trial preparation, court filings and deadlines, case files and calendars: 44 percent. AI speeds them up without replacing the person's part.

Stays with the person. Client and witness contact, judgment calls with the attorney: 20 percent.

Now assume the AI-produced tasks lose 90 percent of their human time (what's left is checking), the assisted ones lose 40 percent, and the person-only ones lose nothing. The same output now takes 50 hours of human work instead of 100.

That halving is the number an exposure study would point to. Look at what else happened.

The job changed shape. The person-only work went from 20 percent of the job to 40 percent. The job tilted toward clients, judgment and the attorney. It also gained a task that didn't exist before: checking the machine's output, about 7 percent of the new job.

The tasks that vanished were the beginner's tasks. Summaries, first-pass review and routine drafting are where a new paralegal learns what a case file looks like and what an attorney cares about. Hold on to that; it comes back below.

There's still no headcount. If the firm handles the same number of matters, it needs half the paralegal hours. If clients, facing lower prices, bring twice as many matters, it needs the same hours as before. Nothing about the model tells you which. That's a question about demand.

The rule that decides headcount

Here's the mechanism behind the tellers, written in plain words.

When a task gets cheaper to do, the price of the output can fall. When the price falls, people buy more. Whether employment goes up or down depends on a race between two numbers: how many fewer hours each unit of work needs, and how many more units people buy.

Economists call the second number the price elasticity of demand. If a 1 percent price cut makes people buy more than 1 percent more, demand is elastic, and productivity can raise employment. If they buy less than 1 percent more, it's inelastic, and the same productivity gain costs jobs.

In the paralegal example, hours per matter halved. For employment to hold, the firm has to handle twice as many matters. If clients only buy somewhat more legal work when it's cheaper, the firm ends up with about 70 percent of its paralegal hours. If they buy a lot more, it could end up with more paralegal hours than it started with. Same technology, same tasks, opposite outcomes.

Headcount rises or falls according to how many more units people buy, set against how many fewer hours each unit needs.

This is the Jevons paradox applied to work: the economist William Stanley Jevons noticed that more efficient engines made Britain burn more coal, not less. It also explains the long history. James Bessen studied two centuries of textiles, steel and cars, and found the same arc each time: automation first created jobs, because demand for the cheaper product was enormous, and later destroyed them, once people had as much cloth or steel as they wanted.

Scroll sideways to see the whole figure

Real cases are messier, and almost always in the same direction. Paralegal time is a small part of what a client pays a law firm, so halving it moves the client's price much less than half. A firm that keeps the saving as margin generates no extra demand. And nobody buys unlimited legal work. Each pushes the outcome toward fewer jobs than the simple rule suggests. What survives is the shape: you can't predict employment from the share of tasks automated without knowing what happens to demand.

For the jobs people ask about most, nobody knows yet. Machine translation plausibly raises the volume of text translated while cutting human hours per word; government projections have translators growing slightly, with people reviewing machine drafts. For legal documents, the question is whether a contract that costs a tenth as much to draft leads people to buy ten times as many. The same projections have paralegal employment flat to 2035, held back by technology including AI. Both are forecasts. Elasticity has to be measured market by market, and for AI-assisted legal work nobody has measured it.

The prediction that skipped the bundle

The most famous wrong prediction in this field is a clean example of skipping the bundle.

In 2016 the AI pioneer Geoffrey Hinton said people "should stop training radiologists now." The argument rested on one task, reading images, where AI was getting very good. But a study of radiologists at three hospitals found only about a third of their time went to reading images. The rest was overseeing exams, talking to clinicians and patients, teaching, reviewing orders.

Radiology didn't collapse. In 2025 US residencies offered a record number of radiology training positions.

There are two explanations, and the honest one starts with the task itself: the reading hasn't actually been handed over. Most approved models detect one finding on one type of image, they can do much worse on another hospital's data, and regulators and insurers are wary of diagnoses made by software alone. But even a model that read every image perfectly would leave most of a radiologist's day untouched, and cheaper reading could raise demand for imaging. Either way, the prediction jumped from "AI can do this task" to "the job will go" without asking whether the task would really be delegated, what else the job held, or what would happen to demand.

How to read an exposure number

The headline figures in this debate come from "exposure" studies. The most cited, published in 2023 by researchers from OpenAI and the University of Pennsylvania, rated about 19,000 task descriptions by whether a language model could cut the time to do each one by half at the same quality. Its result: about 80 percent of US workers could have at least a tenth of their tasks affected, and about 19 percent could have at least half.

Those numbers measure one thing, technical feasibility on paper, and the authors say plainly that they make no predictions about adoption or timing. They don't measure adoption, quality, where the saved time goes, or demand. A Danish study that linked chatbot use to official employment records ruled out, two years after ChatGPT's launch, any average effect on earnings or hours larger than 2 percent, even though most employers in exposed jobs had adopted chatbots. What it did find was reorganization: employers moved tasks around, including new ones like overseeing and integrating AI.

So when you see an exposure number, ask what it is measuring: feasibility, adoption, usage or employment. Headlines swap them constantly.

Where it's already hurting: the bottom rung

Here is the part that should worry anyone who runs a team.

The best labor-market evidence so far comes from a Stanford study using payroll data for millions of American workers, revised in August 2026. Overall, it finds no wave of displacement: employment in the most AI-exposed occupations kept growing, a little slower than the rest. But workers aged 22 to 25 in those occupations are about 19 percent below where they would be had they kept pace with peers in less exposed jobs, and the gap has been widening. Experienced workers show nothing comparable. The shortfall comes mostly from fewer hires rather than more firings, and it concentrates where people use AI to do the task for them, not to help them do it.

The authors are careful: these are early, descriptive patterns, and part of the gap may be explained by education rather than AI. But look at the paralegal table again and the mechanism is easy to see.

The tasks that went to the machine were checkable, codified and repetitive. That's exactly what made them good first assignments: cheap to supervise, cheap to get wrong, and the way a beginner learned the craft. The study's authors suggest something close: AI may be automating the checkable, process-intensive tasks that used to justify entry-level hiring, while making experienced people, whose tacit knowledge the machine complements, more productive.

The tasks that disappear first are the ones juniors used to learn on.

That creates a slow problem on top of a hiring problem. A firm can stop hiring juniors and see no loss at all, because seniors with AI cover the work. Five years later it discovers that nobody learned the job on the job (Hiring When Producing Is Cheap).

The leftovers need an owner

When part of a job goes, someone recombines what's left. Exposure studies can't see this step, because they rate tasks, not bundles. And it's where companies have the most choice.

The paralegal firm has at least three options. It can enlarge the role: the paralegal does the person-only work plus checking, covers more matters, and moves toward case management. It can split it: one person reviews AI output across many matters, another owns client contact. Or it can merge upward: a junior associate absorbs the checking and research, and the paralegal role shrinks to coordination. Same ten tasks, three different jobs, with different pay and different career paths.

A rebundled job can be worse than the old one even when it exists. If the easy, satisfying tasks go to the machine and the person is left with exceptions, complaints and verification, the work gets harder and may pay the same (Skills That Fade When the Machine Does Them).

Is this time different?

The serious objection: earlier machines automated narrow sets of tasks, and people moved to tasks machines couldn't do. If AI can do any cognitive task more cheaply than a person, there is no refuge, because the new tasks can be done by machines too.

My view is that the bundle logic holds as long as one condition holds: that AI's time is worth more somewhere else. Employment depends on comparative advantage, not on who is better at a task. A person keeps a task when the machine's capacity is better spent elsewhere, which stays true while computing power is scarce and weakens as it gets abundant. If the cost of AI per task keeps falling toward zero, the bundle offers no refuge, and the first sign will be falling pay for substitutable tasks, before any sudden wave of unemployment.

The two views predict different data, which makes this testable. If this time is different, declines should spread from young workers to experienced ones, and pay should fall where employment holds. As of mid-2026 the payroll data show neither: experienced workers show no comparable gap, and the adjustment is in hiring more than in pay. I'd watch for both.

What to do

Write down the real bundle for one role. Use time records, not job descriptions. Mark each task as AI produces, AI assists, or stays with a person. You'll learn more from that hour than from any exposure report.

Ask the demand question before the headcount question. If this work got half as expensive, would you sell more of it, use more of it, or just do the same amount with fewer hours? The answer decides whether AI grows the team or shrinks it.

Decide the rebundling on purpose. Who checks the machine's work? Who owns what's left? If nobody decides, the leftovers land on whoever is closest, usually the most senior person, who becomes the bottleneck (Where the Saved Time Goes).

Protect the bottom rung. If AI now does the tasks juniors learned on, design new ways for them to learn: checking AI output against a senior's judgment, owning small matters end to end, sitting in on the decisions. Cutting junior hiring because it's cheaper this year is borrowing against the next decade.

For your own work, look at the mix. Which of your tasks are checkable and codified, which carry accountability, trust or physical presence, and would people buy more of what you produce if it got cheaper? The tasks nobody is trying to automate are where your job is heading.

Check yourself

Why did teller jobs hold up after ATMs but fall after mobile banking?

ATMs made each branch cheaper to run, so banks opened many more branches and demand for tellers held. With mobile banking there was less reason to visit a branch at all. Demand decided both outcomes, not the technology.

Why can halving the hours per matter either shrink or grow paralegal employment?

It depends on how many more matters clients bring when prices fall. If they bring twice as many, the hours stay the same; if only somewhat more, the firm needs fewer hours.

Why do the first losses fall on entry-level workers?

The tasks machines took first are checkable and repetitive, which made them good first assignments for beginners. Firms can stop hiring juniors and see no loss until nobody has learned the job.