AI-first company

People and skills · Run

Keep the entry-level tasks that newcomers learn on

Before you automate routine work, check whether it is how juniors learn the job. If it is, keep some of it for them, or replace the learning another way.

By
Last reviewed
Corrections
Report an error
When Do later. When AI is an established part of the work.
Effort Medium (weeks, a few people)
Evidence Emerging Contested

Two case studies and one measured study of clinicians show skill wearing down when a tool does the work. One randomised trial with school students tested a remedy: a tutor limited to hints. Whether AI is cutting junior hiring is disputed. Other evidence shows AI helping newcomers get better faster.

Fits
Industries All industries
In the guide Roles and leadership

What it is

Before you automate a task, ask one question: who learns from doing it today, and how will they learn once it is gone? Routine tasks are often the first to be handed to AI. They are also where new staff learn the trade.

What the evidence shows

Two case studies show how it happens, and one study has measured it.

  • A two-year study of robotic surgery found that the new technology sharply reduced what trainees did in the operating room. The approved way of learning stopped working. Only a minority of trainees became competent, by practising outside the rules.
  • A case study of an accounting firm found that years of relying on automated software wore down staff skills. Nobody noticed, because the software kept working. The staff were still answerable for the results.
  • A study at four endoscopy centres in Poland measured it on routine audit records. Three months after AI was brought in to help spot growths in the bowel, experienced doctors working without the AI found fewer of them: the detection rate fell from 28.4% to 22.4%.

One remedy has been tested. In a randomised trial with nearly 1,000 secondary-school students in Turkey, students who practised with unrestricted GPT-4 scored 17% lower on the exam without AI than students who never had it. A tutor limited to hints avoided that loss. The trial was with students, not employees.

A study of career moves in Brazil and the UK adds the wider picture. Young graduates often start in the jobs AI is most likely to replace and move up from there. If those jobs go, the path up goes too.

Whether AI is already cutting junior hiring is disputed. The evidence points both ways, and all of it is from the United States.

  • Signs that it is. In payroll records, employment of 22–25-year-olds in the occupations most exposed to AI was about 19% below the path of less-exposed peers, in the authors’ revision of 12 August 2026. The gap came from fewer hires. Firms that advertised jobs for putting generative AI into their work had about 9% fewer junior employees than other firms six quarters later, in the May 2026 version of the study. Startups most exposed to AI, with about 11 staff on average, cut employment by about 8%, mostly juniors. In payroll data from Gusto, which sells payroll software to small firms, headcount for 22–28-year-olds fell within highly exposed roles.
  • Signs that it is not, or not yet. Unemployment among US recent graduates did not rise in summer 2026, compared with earlier summers. The Yale Budget Lab’s monthly tracker finds no clear sign of disruption across the economy.

The studies measure different things: employment in exposed jobs, unemployment, firms chosen by different markers. A national average can stay flat while one group of young people stops being hired.

The other side. In the study of 5,172 support agents, the AI assistant helped newcomers improve faster and appeared to help them learn. So AI does not always harm learning. The effect depends on whether the newcomer still does the work, with help, or only watches it being done.

How to do it

  1. For each task you plan to automate, ask who learns from it today.
  2. Sort the skills involved into two groups: skills you can let go, and skills people still need in order to check the AI and answer for the result.
  3. For the second group, keep a share of the work with people. Rotate it so everyone stays in practice.
  4. Have juniors do the task with AI help and explain the result, not only approve it. For people still learning, set the tool to give hints and explanations, not finished answers.
  5. Every few months, have staff do the task without the tool. Treat it like a fire drill. In the endoscopy study the drop appeared within three months. See check what staff can still do without the tool.
  6. Give seniors time to coach. Automation tends to remove the moments when coaching used to happen.

What to measure

  • Whether staff can still do the task unaided, on a sample of cases.
  • How long a newcomer takes to work independently.
  • How often people catch errors in AI output.

Where it fails

  • Keeping work only for training costs money. Keep the tasks that build skills you still need. Let the rest go.
  • Obsolete skills. The accounting study itself notes that losing skills nobody needs is a normal cost of new technology.
  • AI may take the skilled part, not the routine part. Anthropic, which sells AI models, reports from its own customers’ usage that AI tends to take the more skilled tasks in some jobs, such as travel agents and technical writers. Ask which part of the job the tool takes, not only which tasks look routine.
  • Very small teams may have no juniors at the moment. The question returns with the next hire.

Outside the firm

The entry-level work that goes is often a freelancer’s. On a large freelance platform, demand for writing and translation fell 20–50% against trend after ChatGPT, and demand for novices fell even in skills that AI helps with. A separate study of another global freelance platform found posts for writing and coding jobs fell 21% within eight months of ChatGPT. Among translators with under two years’ experience, the share who see a viable career fell from 69% in 2024 to 41% in 2025, according to the European Language Industry Survey as cited in a study of translators in Belgium and Switzerland.

By company size

  • Micro and small: the risk is one person who alone knows how a task is done without the tool. Make sure a second person can do it. If you sell what AI makes (code, copy, design), expect to hire fewer juniors. Decide where the next person you hire will learn.
  • Enterprise: build the “who learns from this?” question into the approval step for any automation project.

Sources

  1. Beane (2018). Shadow Learning: Building Robotic Surgical Skill When Approved Means Fail. Administrative Science Quarterly, 64(1), 87–123. What this study did and found
  2. Rinta-Kahila, Penttinen, Salovaara, Soliman and Ruissalo (2023). The Vicious Circles of Skill Erosion: A Case Study of Cognitive Automation. Journal of the Association for Information Systems, 24(5), 1378–1412. What this study did and found
  3. Budzyń, Romańczyk, Kitala and colleagues (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology, 10(10), 896–903. What this study did and found
  4. Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. What this study did and found
  5. Cazzaniga, Pizzinelli, Rockall and Tavares (2025). Exposure to Artificial Intelligence and Occupational Mobility: A Cross-Country Analysis. Economía, 24(1), 314–339. What this study did and found
  6. Brynjolfsson, Chandar and Chen (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab working paper (revised 12 August 2026). What this study did and found
  7. Hosseini and Lichtinger (2025). Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data. SSRN working paper 5425555 (version of 6 May 2026). What this study did and found
  8. Gupta, Qian, Simintzi and Sun (2026). Generative AI and Entrepreneurship. NBER conference paper, version of 14 April 2026. What this study did and found
  9. Teutloff, Einsiedler, Kässi, Braesemann, Mishkin and del Rio-Chanona (2025). Winners and losers of generative AI: Early Evidence of Shifts in Freelancer Demand. Journal of Economic Behavior & Organization, 235, 106845. What this study did and found
  10. Demirci, Hannane and Zhu (2025). Who Is AI Replacing? The Impact of Generative AI on Online Freelancing Platforms. Management Science, 71(10), 8097–8108. What this study did and found
  11. Brynjolfsson, Li and Raymond (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. Limits or contradicts this practice. What this study did and found
  12. Chamberlain, Gusto (February 2026). AI's impact on small businesses: payroll analysis. Gusto sells payroll software to small firms. gusto.com
  13. Fairlie and Wu (2026). The Early Impacts of AI on Employment among Recent College Graduates. IZA Discussion Paper 18945. iza.org
  14. The Budget Lab at Yale (updated September 2026). Tracking the Impact of AI on the Labor Market. budgetlab.yale.edu
  15. Anthropic (January 2026). Anthropic Economic Index report: Economic primitives. Anthropic sells AI models; the figures come from its own customers' usage. anthropic.com
  16. Girletti and Lefer (2025). Lost in compensation: pricing methods, rates, and income satisfaction among freelance translators in Belgium and Switzerland. Perspectives. Cites the European Language Industry Survey 2025. research.dial.uclouvain.be