Generative AI has not yet produced a demonstrated broad white-collar wage collapse, despite the confidence of its loudest evangelists and its loudest critics. The strongest early causal evidence says so: a study linking survey data from 25,000 Danish workers in 11 AI-exposed occupations to administrative records found no effect on earnings or hours through December 2024, either for workers using chatbots or for firms adopting them.[1] That result should be the starting point of the debate, not its inconvenient footnote.
But “no wage collapse yet” does not mean that nothing has changed. This essay’s claim is narrower: before AI removes a job title, it is likely to cheapen the paid first draft—the routine memo, market scan, slide deck, product description, or low-stakes coding task that once justified a junior employee’s hours or a freelancer’s fee. Jobs are bundles of tasks, responsibility, relationships, and institutional liability. AI does not need to erase the bundle to weaken the price of the most visible, standardised part of it.
The labour-market evidence is still stubbornly quiet
The International Labour Organization’s 2025 global index puts one in four workers in occupations with some exposure to generative AI, but only 3.3% of global employment in its highest exposure category. Its central conclusion is not that most jobs vanish; it is that job transformation is the more likely result because occupations contain tasks that still require human input.[2]
That distinction matters. A transformed job can still become a worse bargain. An analyst who once spent two days assembling a research brief may now be expected to deliver it in one day. A junior designer may still have a job, but the market may be less willing to pay for iterations that can be generated in minutes. A developer may retain responsibility for code, while the amount of code that can be produced per person rises sharply.
There is no global wage series proving that AI has already caused this outcome across white-collar work. The Danish result is a serious counterweight, not a footnote. It shows that task-level speed does not automatically become lower pay, fewer hours, or smaller payrolls. The commercial pressure described here is therefore a forward-looking mechanism, not a completed statistical fact.
The corporate payoff is still harder to find than the corporate rollout
Stanford’s 2025 AI Index reported that 78% of surveyed organisations said they used AI in 2024, up from 55% a year earlier. Seventy-one percent reported using generative AI in at least one business function. Yet, among organisations reporting financial effects, the most common cost-saving result was below 10%, and the most common revenue increase was below 5%.[3]
Adoption is not the same thing as a durable return. A company can buy licences, run pilots, and proudly announce an AI programme without changing its cost base very much. The proof of the pudding is in the P&L, not in the number of employee accounts created.
The OECD reaches a similar practical conclusion from its enterprise research: managers often struggle to identify where AI solves a real workplace problem, and they underestimate the organisational and cultural changes implementation requires.[4]
So the near-term story is not “AI works everywhere” or “AI is a bubble.” It is more uncomfortable. The tools can be very good at selected tasks while the organisation is still bad at turning that speed into a business result.
The first pressure falls where the work looks interchangeable
Anthropic’s January 2026 Economic Index is not a survey of the whole labour market; it is an analysis of Claude usage, and it must be treated as vendor-specific evidence. Still, its pattern is informative. In its November 2025 sample, computer and mathematical tasks made up roughly one third of Claude.ai conversations and nearly half of API traffic. The ten most common work tasks accounted for 24% of Claude.ai use. Augmentation was slightly more common than automation on Claude.ai, 52% to 45%.[5]
The point is not that one vendor’s users represent every worker. They do not. The point is that AI use is already concentrated in tasks that are easy to describe, digitised, and repeated. Those are precisely the conditions under which clients can compare outputs and ask why a task still costs what it cost before.
This is how work gets repriced: not through a boardroom declaration that “humans are obsolete,” but through small changes in the quote, the turnaround time, the staffing plan, and the definition of entry-level competence.
The part of the job that does not fit inside a prompt
Speed alone does not settle value. The worker who merely produces a generic first draft is exposed. The worker who knows which facts can be trusted, which exception changes the answer, who bears responsibility when the answer is wrong, and who can obtain cooperation from real people holds something more difficult to automate.
That is not a sentimental defence of “human creativity.” It is a commercial distinction. Context, judgment, accountability, access, and trust are costly to build because they are not contained in a prompt.
This is also why the debate over whether AI is “good for workers” is too broad to be useful. It can improve the value of a professional who owns a client relationship and has authority to make a decision. It can lower the price of a junior worker whose contribution is mainly a standardised first pass. Both effects can occur in the same office.
The statistic that will arrive late
Public debate tends to react to visible layoffs. By then, the quieter adjustment has already happened: fewer junior openings, flatter fee schedules, more unpaid tests, higher output expectations, and a larger share of the final reward captured by the owner of the client, the workflow, and the distribution channel.
The old saying is that a rising tide lifts all boats. Labour markets are less generous: before the tide reaches the worker, someone decides who owns the dock.
The policy mistake would be to wait for a dramatic employment collapse before measuring what matters. Governments, firms, and labour researchers should track entry-level hiring, task-level productivity, fees, wage progression, and who captures the savings. Counting job titles alone will miss the first round of damage.
The individual mistake is simpler: treating AI as either a magic assistant or an alien enemy. It is becoming a pricing tool. If a client can obtain your visible output from ten people using the same model, your value must increasingly come from what remains hard to standardise—and from owning a relationship that the model cannot call by itself.
Sources
- Anders Humlum and Emilie Vestergaard, “Early Labor Market Transformation under Generative AI,” NBER Working Paper 33777 (2026), NBER paper; see also the authors’ public summary, Brookings, 6 October 2026 (accessed 8 October 2026). ↩
- International Labour Organization, Generative AI and Jobs: A Refined Global Index of Occupational Exposure (2025), ILO publication (accessed 8 October 2026). ↩
- Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report 2025: Economy, Stanford HAI (accessed 8 October 2026). The underlying figures are survey-based and should not be read as causal estimates of AI’s effect on wages or output. ↩
- OECD/BCG/INSEAD, The Adoption of Artificial Intelligence in Firms: New Evidence for Policymaking (2025), OECD publication (accessed 8 October 2026). ↩
- Anthropic, “Economic Index: New building blocks for AI use” (January 2026), Anthropic Economic Index (accessed 8 October 2026). Anthropic’s data reflect its own products and are used here as evidence about observed tool usage, not as a representative labour-market sample. ↩