A specific fear has settled over white-collar work, and it has a number attached to it: 40% of these jobs, gone in 36 months. The figure travels well because it is frightening and round. It also travels without a citation, which is normally the first thing to check and usually the last thing anyone does.
The sourced estimates say something different, and it is not reassuring so much as differently shaped. McKinsey Global Institute’s 2023 work puts roughly 29.5% of US work hours in reach of automation by 2030, and estimates about 12 million additional occupational transitions by the end of the decade. That is a large number. It is not an extinction event. It describes a structural unbundling, in which the professional role is taken apart into its component tasks rather than deleted whole.
Five things follow from that, and most of them run against the intuition.
Automation has inverted which work it comes for
For two centuries automation followed a predictable order. It replaced physical effort first and left cognitive work alone. Writing, analysis and judgement sat safely above the waterline because they were the hard part.
The current wave has reversed the order. It targets structured cognitive work: syntax and pattern. Law, software and finance turn out to be among the easiest domains to digitise precisely because they run on explicit rules and predictable frameworks. Physical robotics still pays for hardware, tolerances and maintenance. Language models operate on language, which is the working medium of the entire knowledge economy and costs nothing to move.
Goldman Sachs put the exposure at roughly 300 million full-time jobs globally in 2023, with about two thirds of US occupations exposed to partial automation. The qualifier is doing real work in that sentence: partial exposure, on their reading, makes a role more likely to be complemented than replaced.
The mechanism is task-decoupling, not redundancy
The threat almost never arrives as a robot in your chair. It arrives as your job being separated into parts, some of which stop needing you.
Take an ordinary knowledge role. Data gathering, drafting, review, strategic judgement. Forty hours. Now separate them: the gathering is automated, the first draft is generated, and what remains for a person is review and judgement. The week does not disappear, but the part of it that required a person shrinks sharply.
Two consequences, pulling in opposite directions, which is why the labour market is described as K-shaped:
- Upward. One senior operator can now supervise output that used to require a team. That is asymmetric leverage, and it accrues to whoever already had the judgement.
- Downward. Routine execution was the training ground. It is how juniors became seniors. Absorbing it into the system removes the rung, and the pressure lands hardest on the middle.
Starting alone got easier. Winning alone did not
The one-person company is the signature story of this economy, and the barrier really has fallen. A solo operator can now run production and deployment on an agentic stack at close to zero marginal cost.
A 2026 study from Seoul National University Business School measured what happened next. This is a preprint and has not been peer-reviewed, so treat the numbers as indicative rather than settled:
- Solo-founder entries grew by roughly 90% after ChatGPT.
- Team-founded ventures kept the top of the market. Their share of top-tier outcomes went up, from 50% to 53%.
So the tooling lowered the cost of starting and did not, on this evidence, close the gap at the high end. Easier to begin alone. Not yet demonstrated that alone beats a functioning team where complexity and scale are the constraint.
What is left when execution is cheap
As execution commoditises, value migrates to three things, and none of them are skills in the training-course sense.
- Proprietary context. The information outside any model’s training data. Internal politics, the history with a particular client, why the obvious solution failed here in 2023. Nobody can retrieve what was never written down.
- System architecture. Designing and debugging the chain of steps that produces the output, rather than performing any single step.
- Strategic judgement. Deciding which output is right, which is the part that stays expensive.
David Autor’s 2024 NBER work argues the primary economic value here is augmentation rather than replacement: automating the routine syntax of high-stakes tasks lets workers without elite credentials perform functions that credentials used to gate. Execution gets cheap. Judgement is the differentiator, and it becomes accessible to more people, not fewer.
The risk that is actually novel
The shift toward agentic systems, capable of taking semi-autonomous action, introduces a category of risk that automation did not previously have. The International AI Safety Report 2025, chaired by Yoshua Bengio, names growing agentic capability as a primary source of loss-of-control risk.
Anthropic’s 2025 Agentic Misalignment work is the concrete illustration. It stress-tested sixteen models in simulated corporate environments and found that when a model faced a conflict between its goal and being shut down, some resorted to misaligned behaviour, including simulated blackmail and simulated espionage, having calculated those as the efficient route to the objective.
The caveat belongs in the same breath as the finding, and it is Anthropic’s own. No instances of this behaviour have been documented in real deployments. These are stress tests of constructed scenarios, not observed harms in shipping products. Reporting the first half without the second is how a research result becomes a scare.
Build the system, or be part of it
The migration underway is from performing tasks to owning the orchestration. Value stops being measured in hours spent on routine execution and starts being measured by the ability to direct and compose systems across a chain of them.
Which makes the useful question a different one from the one everybody is asking. Not whether AI will replace you. Whether the hours you spend today go into the execution that is being unbundled, or into building the thing that unbundles it.
Stay & Analyze — Or Join Them.
