November 20, 2026 · AI writer · Oak City Intelligence

How Many Workers Does Your AI Add Up To?

A flat vector illustration in Oak City Intelligence's house style for "How Many Workers Does Your AI Add Up To?."

Two Ways to Count

The numbers that frighten people count exposure, meaning what a model could in principle touch. In January 2024 the IMF said that almost 40% of global employment is exposed to AI, about 60% in advanced economies. Goldman Sachs estimated in 2023 that roughly two thirds of US occupations are exposed to some degree, and that the equivalent of 300 million full-time jobs worldwide could be exposed to automation. Researchers at OpenAI and Penn found that about 80% of US workers could have at least a tenth of their tasks affected by language models.

The same sources carry their own small print: the IMF says roughly half the exposed jobs may benefit. Goldman says most exposed jobs are only partially exposed and "more likely to be complemented rather than substituted by AI." The ILO, which puts one worker in four in an occupation with some exposure, says "transformation of jobs is the most likely impact."

The other way to count is realised use, meaning what AI is actually doing. The St. Louis Fed's survey found that in August 2025, 37.4% of employed Americans used generative AI for their job, that 5.7% of all work hours involved it, and the time it saved came to 1.6% of all work hours.

Exposure and use are different kinds of number, and most of the fear comes from reading one as the other. An exposed job is a job a tool could touch, and a used tool is a job a tool has partly done.

Run the Census

Turn the use figures into people: the US had about 163 million people employed in July 2025. 1.6% of their hours is the working time of roughly 2.6 million full-time people, and 5.7% is roughly 9 million. Those are our multiplications of the cited figures. They assume hours spread evenly, and rest on what workers reported saving.

Counted by work actually done, then, AI in 2025 was a workforce of a couple of million people. Counted by what it could touch, it was hundreds of millions, and the gap between those two numbers is the real story.

The small number is growing: by August 2026 the same survey put the time saved at 2.1% of work hours, and work use at 40.5% of employed people.

Counting in people makes the number earn itself in hours rather than headlines. A small, growing number is reassuring now, and is the trend line to watch.

Who the Small Number Lands On

The economy-wide measurements show little so far. Yale's Budget Lab found no discernible disruption to the broad US labor market in the 33 months after ChatGPT, and its tracker, updated in October 2026, still finds none. In Denmark, Anders Humlum and Emilie Vestergaard followed about 25,000 workers in 11 exposed occupations and could detect no effect on earnings or hours, ruling out effects larger than 2%.

The small number has an address. Stanford economists tracking payroll records found that by June 2026 employment of workers aged 22 to 25 in the most AI-exposed occupations stood 19% below where it would have been had it kept pace with their less-exposed peers, while experienced workers in the same occupations showed no comparable gap. That shortfall works through reduced hiring and is concentrated where AI substitutes for human tasks rather than complementing them, and the authors call these early, descriptive indicators, not causal estimates.

The shape of the new work explains why. On Anthropic's business API about three quarters of interactions are automation, where a person sets the task, then reviews and approves. And the tasks the model takes tend to be the more educated parts of a job, so the net first-order effect, in Anthropic's words, is "to deskill jobs."

A job made of directing and checking suits people who already know the work. It removes the rung on which people used to learn it by doing it.

What Each Worker Brings With It

A new hire brings habits. So does a model, in the form of the training data, objectives and defaults of whoever built it. Every task handed over carries them.

The habits show up as sameness. In a study of 293 writers, access to AI story ideas made stories rate as more creative, most of all for the less creative writers. But the AI-assisted stories were more similar to one another than the stories written unaided. Doshi and Hauser call it a social dilemma, individually better off and collectively narrower (Doshi and Hauser, 2024). In a 36-person study, people brainstorming with ChatGPT produced ideas less distinct from one another's than people using a different tool, and felt less responsible for the ideas they came up with (Anderson, Shah and Kreminski, 2024).

A business that routes its drafting through one vendor's model starts writing in that vendor's habits. And since its competitors route theirs through the same two or three models, the habits converge. A census asks how many, and the better question is whose defaults they run on.

What I'd Actually Do

Run your own census for one week, listing every task you or your staff hand to an AI tool. Mark three things against each task: whether anyone checks it, whether you could still do it yourself, and whether you no longer could.

For each automated task, write one line on whose defaults it is running: what the tool was trained and rewarded for, and what it would take to switch vendors.

Keep at least one way for a junior person to learn the work by doing it, because the rung you remove is the one your next senior person was standing on.

Count both columns, staff and staff-equivalents in software. Read the second one in hours rather than headlines. The first says how many people you have; the second says whose habits you have hired without an interview.


Sources

Written by Oak City Intelligence's AI writer, from an outline and a facts sheet by John. Reviewed before publication.