Tasks, not jobs
The most useful mental model is that jobs are bundles of tasks, and AI is unbundling them rather than removing them wholesale.
A marketing role might involve research, drafting, analysis, stakeholder management and judgement about brand. Current tools handle parts of the drafting and research well, some of the analysis competently, and the judgement and stakeholder work barely at all.
The outcome is a role whose composition changes: less time producing first drafts, more time directing, editing and deciding. That is a real change in what the job is, and it changes what employers pay for — but it is not the same as the job disappearing.
Where pay has actually risen
The clearest wage effect so far has been upward, in two groups.
The first is people who build these systems — machine learning engineers, data engineers, infrastructure specialists and researchers. Demand has substantially outpaced supply and compensation has risen accordingly.
The second, and larger, group is people who apply the tools credibly inside an existing domain. A lawyer who uses them well for review and drafting, an analyst who automates a reporting pipeline, a designer who compresses iteration cycles. That combination — real domain expertise plus genuine fluency — is currently scarce and valuable, and it is available to people already in a field rather than requiring a career change.
Where the pressure is
The roles under most pressure are those whose output is primarily routine text, data manipulation or standardised administration.
Basic content production, first-line customer support, routine translation, data entry, simple document review and entry-level administrative coordination have all seen either reduced hiring or expanded expectations for the same headcount.
Entry-level work is disproportionately affected, which creates a genuine structural problem: many professions trained juniors by having them do exactly the tasks now automated. How fields rebuild that on-ramp is an unresolved question, and it is worth being aware of if you are early in a career.
What is most insulated
Several categories remain difficult to automate for reasons that are structural rather than temporary.
Physical work in unpredictable environments — trades, maintenance, construction, most healthcare delivery — requires dexterity and situational judgement that remain hard problems. Licensed roles carrying legal accountability need a person answerable for the decision. Work built on trust and relationships is resistant because the relationship is the product.
Also resistant: anything requiring genuine responsibility under uncertainty, where the cost of a confident wrong answer is high. That is precisely the current weakness of these systems.
The honest uncertainty
Anyone claiming precision about the next decade is overstating what is knowable, and it is worth saying so plainly.
Previous technology waves eliminated occupations and created others, usually with a painful transition and a net increase in employment. Whether this follows the same pattern depends on how fast capability improves and how quickly institutions adapt — neither of which is predictable.
What can be said with reasonable confidence is narrower: routine cognitive work is under sustained pressure, complementary skills are rising in value, and the transition is happening faster than previous ones. Planning around those three is reasonable; planning around a specific timeline is not.
What to actually do
The practical response is less dramatic than the discourse suggests.
Learn the tools properly within your own field. Not superficially — enough to know where they are reliable, where they fail, and how to verify output. That capability is currently rewarded and costs weeks, not years.
Deepen the parts of your work that are hardest to automate: judgement, accountability, relationships, physical skill, domain knowledge that is not written down anywhere. Keep checking your role against the market annually, because in a fast-moving period internal pay lags external reality more than usual.
And be sceptical of both panic and complacency. The people who do worst are usually those who assumed nothing would change.