The entry-level job was always a bargain dressed as a chore: the newcomer did the routine work nobody else wanted, and in exchange absorbed the profession, one tedious rep at a time. Junior lawyers summarized documents, junior analysts built the first draft of the model, junior developers wrote the boilerplate. AI now does much of that faster and cheaper, which breaks both halves of the bargain at once. The chore is automated, and so is the curriculum hidden inside it.
The squeeze is no longer hypothetical. Using large-scale payroll data, Brynjolfsson, Chandar and Chen (2025) found that workers aged 22 to 25 in the occupations most exposed to AI experienced a relative employment decline of roughly 16 percent, while experienced workers in the same occupations, and young workers in less-exposed fields, held stable or grew. Two details sharpen the picture. The adjustment ran through employment, hiring and headcount, rather than through wages. And the losses concentrated where AI automates the work instead of augmenting it: the automation-versus-augmentation line, not exposure as such, predicted who got hurt. The canary, as the authors put it, is the person on the bottom rung.
The paradox: AI favors the novice it displaces
Here is what makes the moment strange rather than simply grim. Inside firms, the experimental evidence consistently shows AI helping junior people most. In a staggered rollout across 5,172 customer-support agents, access to a generative AI assistant raised productivity about 15 percent on average, but the gains concentrated among novices and lower-skilled workers, up to roughly 34 percent, with negligible effect on the most experienced; the system, trained on the patterns of top performers, effectively piped their tacit knowledge to beginners (Brynjolfsson, Li and Raymond, 2025). A field experiment with 758 BCG consultants found the same compression from the other end: with GPT-4, consultants completed 12.2 percent more tasks about 25 percent faster with roughly 40 percent higher quality on tasks within the model's competence, and the lower-performing half gained most (Dell'Acqua et al., 2025).
So the technology narrows the gap between novice and veteran even as the labor market cuts the number of novices it hires. Both findings are correct; they sit at different levels. The market prices the routine output a junior used to be hired for, and that price is collapsing. The experiments measure what a junior can produce with the tools, and that is rising. The gap between those two facts is where a career strategy has to live.
The new bottom rung is judgment
The BCG study contains the clue. Its authors describe a jagged frontier: on tasks beyond the model's competence, AI users did worse, because they trusted answers that were plausible and wrong. Someone has to know where the frontier runs, task by task, and catch the confident error. That skill, verification under uncertainty, is not a degraded version of the old grunt work. It is closer to what seniors always did, applied earlier: judging output rather than generating it. The entry-level role does not vanish; it migrates from producing the first draft to interrogating it, and the newcomers who treat AI output with editorial suspicion are practicing the profession's actual scarce skill from day one.
The employer surveys, whatever their limits as forecasts, point the same direction: the World Economic Forum's 2025 employer survey across 55 economies projects that nearly 40 percent of required job skills will change by 2030. A first-jobber should read that not as a threat but as an arbitrage: when the skill definition of a role is moving, tenure advantages shrink, and the person who learns the new definition fastest is not necessarily the person who mastered the old one.
Hiring without the proxy
There is an older precedent for what happens when credentials stop predicting. As Ricardo Semler has told it, he never looked at a diploma before hiring; Semco had factories run by engineers without degrees, and a maintenance technician led 200 people because everyone trusted him. When the standard proxy fails, organizations that still need to allocate responsibility fall back on demonstrated judgment and earned trust. AI is now doing to the routine-work résumé what Semco's practice did to the diploma: stripping a proxy of its information value. What remains legible is what always remained: whether this particular person catches what the machine misses, ships what they promised, and can be believed.
For the graduate staring at a thinner job market, the strategy follows from the evidence rather than from reassurance. Aim at augmented work, not automated work, since that boundary predicted who got hurt. Use the tools to close the competence gap the experiments say they close. And make judgment visible early: documented catches, verified output, small trusted calls. The bottom rung was never really the work. It was the chance to be observed becoming reliable, and that rung can be rebuilt anywhere someone is watching.
Sources
- Brynjolfsson, E., Chandar, B. & Chen, R. (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of AI. Stanford Digital Economy Lab. Workers 22-25 in the most AI-exposed occupations saw ~16% relative employment decline; adjustment via employment not wages; automation vs augmentation predicts impact. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
- Brynjolfsson, E., Li, D. & Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics 140. 5,172 support agents: ~15% average productivity gain, concentrated in novices (up to ~34%); AI diffuses top performers' practices to beginners. https://academic.oup.com/qje/article/140/2/889/7990658
- Dell'Acqua, F. et al. (2025). Navigating the Jagged Technological Frontier. Organization Science. Field experiment with 758 BCG consultants: within AI's frontier, +12.2% tasks, 25.1% faster, ~40% higher quality, with lower performers gaining most; beyond the frontier, more errors from plausible-but-wrong output. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- World Economic Forum (2025). Future of Jobs Report 2025. Employer survey, 1,000+ firms in 55 economies: ~40% of required skills expected to change by 2030 (survey-based projection). https://www.weforum.org/publications/the-future-of-jobs-report-2025/