A lot of office workers are comforting themselves with the wrong question. They keep asking whether AI can do their whole job. That is not the threshold their employer will use. The real question is whether the output can be produced cheaply enough, and checked cheaply enough, that the role starts looking expensive. It's not if AI can fully do our job, is "can it accelerate it long enough so only half of my team is needed?". Because the answer to that, sadly, is yes.
That matters because a large amount of office work already arrives in reviewable form. A market note, a draft, A documentation pass, A research summary. A slide deck. A routine code fix with clear acceptance criteria. The manual work behind those outputs may still be real, but the finished product is often legible enough that a more senior person can inspect it, correct the obvious failures, and still spend less than the old fully loaded labor cost.
That is the mechanism people do not want to stare at. AI does not have to replace trust, judgment, or context all at once. It only has to make enough of the first pass machine-producible that one reviewer can supervise what used to require several people producing from scratch. In practice that means fewer analysts, fewer coordinators, fewer junior writers, fewer junior coders doing clean-up work, and more pressure on the remaining people to validate machine output instead of generating every line themselves.
You can already see the pattern in ordinary workflow. A manager used to need an analyst to gather source material, draft the internal memo, and shape the first recommendation. Now the analyst may still exist, but can probably server multiple managers at once. Or a manager needs less analysts. The same thing happens in code review. A human still matters, sometimes a lot, but the human gets pulled upward into validation, edge cases, and responsibility while the cheap first pass gets generated elsewhere.
That is why office work is more exposed than people want to admit. It is because information work used to be expensive. Organizations had to pay humans for the first pass because there was no other way to get one. Once the first pass gets cheap, the value of the role gets judged more harshly. The role is no longer priced around producing organized language. It gets priced around ownership, verification, and consequences.
The contrast with art and other hard-to-specify work should stay narrow. Good visual work is still harder to describe precisely and harder to verify cheaply than text, spreadsheet logic, or routine code changes. That does not make creative work immune. It just means the compression logic is strongest where success is easy to describe and failure is cheap to inspect.
The work that survives better sits closer to reality. It owns systems, signs off on outcomes, absorbs consequences, and handles messy context that does not fit neatly inside a review queue. It is harder to compress the job of the person who has to validate a physical system, manage a live client conflict, own an outage response, or make a decision when the inputs are incomplete and the cost of error is real. If anything, it makes those remaining human bottlenecks more visible. For engineers for example, it reduces the cost of writing code (even designs) a lot, while increasing the value of people that can adapt and drive something to completion. Being an end-to-end person, someone who can get an idea or a feature broken down in a lot of components and then being able to prioritize and execute on them is what software engineering is now about. It's unlikely that you will code much anymore, specially the better the AI gets at it. But you will need to understand patterns, designs, tooling... and put all of that together.
Yes, AI on it's own won't build anything. But a person with AI will build what a team of 10 used to do. So, to those 9, you can definitely say that AI is replacing them.