AI Automation vs Augmentation: What the Job Data Shows

AI automation vs augmentation sounds like two competing theories about the same future. Real hiring data says both are already happening, at the same time, split by the kind of work a role actually does. A Harvard Business School working paper that tracked real US job postings before and after ChatGPT's public release found exactly that split, and it is sharper than either side of the debate usually admits.

What the study actually measured

The paper, "Displacement or Complementarity? The Labor Market Impact of Generative AI", published December 2024 by Chen, Srinivasan and Zakerinia, built two scores for every occupation in the O*NET task database: how automatable its tasks are, and how much those tasks could be augmented rather than replaced by generative AI. The authors then matched those scores against a near-universe dataset of actual US job postings and compared hiring before and after ChatGPT's release, using a synthetic difference-in-differences design against a matched control group. This is not a survey of employer expectations. It measures postings that already happened.

Postings split, they did not just shrink

Once occupations were grouped by task mix, the two groups moved in opposite directions. Job postings for occupations in the top quarter of automation potential fell 17% per quarter per firm after generative AI's introduction, relative to the control group. Postings for augmentation-prone occupations, the ones mixing automatable tasks with work that still needs a person, rose 22% per quarter per firm over the same period.

The same split shows up in skill requirements, not only headcount. In automation-exposed roles, the count of generative-AI-exposed skills a posting asked for fell 24% per firm per quarter. In augmentation-prone roles, that same measure rose 15%. Automation-prone hiring did not just shrink, it asked for less. Augmentation-prone hiring grew and asked for more.

What actually separates the two kinds of roles

The dividing line in the research is not the industry or the job title, it is the task mix inside the role. A role built from structured, repeatable tasks scores high on automation and low on augmentation. A role that pairs tasks generative AI can do with tasks that still require judgment, negotiation or reading a specific situation lands on the other side, and the paper treats that combination as the one most likely to see productivity gains rather than headcount cuts.

That lines up with how a task actually gets scored for automation on this site: task by task, not job title by job title. It also lines up with the six non-technical domains this site measures, since judgment and reading people are exactly the parts of a role a language model cannot absorb on its own.

What to do with this before it becomes obvious

None of this says a specific reader's job is automation-prone or augmentation-prone. That takes looking at the actual tasks inside a specific role, not the title on a business card. What the research does support is a practical move: build and show evidence of the judgment-heavy parts of a role now, rather than waiting to find out which side of the split it lands on. Proof beats a listed skill in a hiring process either way, and the six-domain check is one way to get a specific, checkable read on where a reader already stands.

The honest limit

This is one paper's read on US job postings through one dataset and one scoring method, not a universal law, and occupation-level scores still average over real variation between two people doing what is nominally the same job at two different companies. What the paper does support, directly, is that generative AI has not shrunk labor demand across the board. It has redirected that demand toward roles that pair machine speed with something a machine still cannot supply on its own.