At a Fortune 500 software company, a generative AI system began sitting beside customer-support agents while they worked. It followed each conversation, drew on patterns from earlier exchanges, and proposed a response. The agent could ignore the suggestion, revise it, or send something close to what appeared on the screen.
Researchers followed 5,172 agents as the system was introduced. Productivity rose by about 15 percent, with much larger gains among less experienced and lower-performing workers. New agents with two months on the job began performing about as well as unassisted agents who had been there for more than six. During software outages, agents who had used the tool retained some of their improvement, suggesting that the system had taught them something rather than merely carrying them through each exchange.1
This is close to the best case for AI at work. The system spread the practices of strong employees to people who had not yet accumulated them. New workers became useful sooner. Some of what looked like experience may have been knowledge waiting for a better delivery system.
A six-month ramp-up became two months. The obvious response is to celebrate the missing four. We should celebrate the missing four!
Experience is an expensive and poorly organized teacher. A novice encounters problems in whatever order the day supplies them, receives coaching when a manager has time and patience, and slowly reconstructs lessons the company already possesses somewhere, diffuse as it may be. AI can offer the right example when the problem appears and spare a new employee months of rediscovering what the organization already knows. There is no reason to preserve the delay merely because older workers endured it.
The harder question is what, exactly, has disappeared here. What shape is outlined in the absence?
The study measured that job as the company had defined it: speed, resolution, customer response, and some persistence after an outage. It could not tell us everything the workers learned during the four months that vanished. Perhaps the old curve was mostly a waste. Another possibility remains: the agents crossed the company’s performance threshold before crossing an experience generation threshold the company did not measure.
Junior work has often produced two things at once. One is the immediate product. A new analyst assembles background material. A young lawyer prepares routine language. A programmer fixes an ordinary bug. A manager tracks decisions and discovers that “circling back” is less a commitment than an eternal recursion, depending on the employee.
The other product, the one we are focused on here, appears later. The analyst starts recognizing a number copied from the wrong year. The lawyer notices that the standard clause behaves differently in this agreement. The programmer sees that the ordinary bug belongs to a larger design problem. The researcher becomes suspicious of data that arrived cleaner than the institution producing it has ever been or even aimed for.
The organization pays for the first product. It receives the second gradually, as a fringe benefit, and rarely figures out how to put it on the invoice. Production supplies cases, senior people correct them, and repetition becomes adept pattern recognition. Training hides inside useful work and finances part of itself.
Generative AI can separate these two products.
When a system prepares the background, drafts the language, repairs the simple code, and cleans the data, the organization has less reason to hire beginners to do those things. It may still need people who can handle unusual cases, revise procedures, train others, and accept responsibility when the standard answer fails. But that need arrives later and isn’t tied to dollars cleanly. The immediate savings and the future need appear in different types of budgets.
This is different from ordinary deskilling, which we have discussed earlier in this series. A worker who delegates something she once knew how to do can lose an existing ability. Here the person may never receive enough experience to develop it. The missing competence belongs to someone who was never hired, or who entered after the work that once developed it had been removed.
An institution can live on stored experience for a long time. Senior people learned through the previous arrangement and continue reviewing difficult work. Managers still know whom to call when the dashboard begins insisting on something everyone in the room knows is wrong. But the gap rears its head when those people leave.
Calling the old arrangement “apprenticeship” can make it sound more deliberate than it was. Many organizations assigned low-risk work to the cheapest available person, placed that person near someone more experienced, and allowed development to occur where it could. Professions have also shown great creativity in describing their least appealing tasks as formative. A senior employee remembers suffering through several thousand hours of routine work and concludes that the suffering supplied the lesson. He may be right. He may also be mistaking the long circuitous walk through the hallway for the destination because he spent a long time in it.
Some scut work teaches only that suffering becomes tradition once the person who suffered gains control of the assignments.
The useful distinction is between repetition and practice. Repetition reproduces an action. Practice changes what a person can perceive or decide. The novice forms an expectation, acts on it, sees the result, and adjusts. Feedback arrives while the decision is still recoverable. Responsibility grows with performance.
Citation formatting is repetitive. Reading the cases behind the citations may be practice. Moving numbers between systems is repetitive. Reconciling two sources that should agree and do not can be practice. Producing the hundredth routine draft may mostly teach endurance and several methods of appearing mentally present on the conference call when the ask comes down.
Fifty straightforward reviews look like fifty completed reviews. They may also have taught the reviewer what straightforward looks like. That second result becomes valuable when the fifty-first case contains an exception. AI gives us a chance to improve this process. It also makes it easier to eliminate the volume before understanding what the volume was actually doing beyond traditional metrics.
A study involving nearly one thousand high-school students shows the distinction unusually cleanly. Students used one of two GPT-4 systems while solving mathematics problems. A general assistant made answers easy to obtain; a guarded tutor was instructed to guide students rather than supply them. Both improved performance during practice. Once the systems were removed, students who had used the general assistant scored 17 percent below the control group, while students who used the guarded tutor did not show the same penalty.2
The experiment establishes a narrow point with broad relevance: better performance while assistance is available does not prove that the person acquired the ability the performance appears to demonstrate. The design mattered because one system displaced more of the student’s role in reaching the answer.
Tutor CoPilot points in the more encouraging direction. In a preregistered randomized experiment, AI supplied real-time suggestions to human tutors, such as asking a guiding question or checking understanding. Students working with assisted tutors became more likely to master the topic, with the largest gains among tutors who had previously received lower quality ratings. The tutors also became more likely to use stronger instructional strategies and less likely to give the answer away.3
AI can complete the novice’s work and leave the person to inspect it. It can also help the novice perform a better version of the work. The second use preserves the encounter: the person still has to ask, notice, decide, and recover from being wrong. That may be a better apprenticeship than the one it replaces. The concern begins when organizations assume the replacement happens automatically.
Most workplace systems are purchased to increase output. Success appears as shorter turnaround, lower staffing costs, and fewer errors. Those are sensible measures. An employer does not buy software mainly to create the employee it may need in eight years. Learning may become visible only when a worker faces an unfamiliar problem or inherits responsibility for the method itself.
Technology can also rearrange who gets close enough to consequential work to learn from it. Matthew Beane and Callen Anthony found forms of “inverted apprenticeship” in urological surgery and investment banking: senior professionals learned from technically fluent juniors while preserving their senior position, and several pathways weakened learning among the most junior workers.4
The pattern is easy to imagine with generative AI. A junior employee can generate the draft without joining the client conversation that determines what it must accomplish, or prepare the analysis without attending the meeting where its assumptions are challenged. Technical fluency moves downward while authority stays where it was.
She may then be told that AI has allowed her to begin at a higher level. Sometimes it has. The phrase becomes suspicious when the higher-level assignment is reviewing output that she has never learned to produce. Review looks more sophisticated than drafting and is also cheaper: the system creates the document, the junior employee checks it, and a senior person approves it. Everyone has moved upward on the organization chart without the inconvenience of moving through the work.
A reviewer can catch an incorrect name, unsupported claim, or obvious contradiction. She may struggle with the omission that becomes visible only after trying to construct the answer from the underlying material. If production no longer supplies that experience, the organization needs another route to develop it.
Evidence that junior opportunities are already shrinking is early and conflicted. A Stanford working paper found a 16 percent relative employment decline among workers aged twenty-two to twenty-five in the most AI-exposed occupations after firm-level changes were considered, while more experienced workers in the same occupations remained stable or grew.5 The authors later emphasized that the timing and causes were less clean than the headline suggests and that AI could not be established as the sole explanation.6
A separate working paper using resume data covering nearly 62 million American workers found slower junior hiring after firms began actively implementing generative AI, while senior employment changed little.7 Danish administrative data offer a counterweight: researchers found no detectable average effect larger than 2 percent on earnings or recorded hours during the first two years, including among early-career workers, even as tasks changed.8
Current evidence does not establish an apprenticeship collapse. It gives us an early employment signal, a plausible mechanism, and no long-term measure of what happens to the supply of experts. A decline in junior hiring would not settle the argument anyway. An occupation could lose routine work and improve its training. The cliff appears only when paid entry routes shrink without a replacement for what they taught.
There is reason to expect that replacement to be underprovided because training has always had an incentive problem. The employer pays the novice, accepts slower work, assigns someone experienced to review it, and absorbs mistakes. Much of the benefit arrives later, perhaps after the employee has moved elsewhere. Economists have long studied why firms fund transferable training anyway; imperfect labor markets and firm-specific arrangements can let employers capture enough of the return to make it worthwhile.9
Junior production has supplied a more practical subsidy. The new lawyer’s imperfect draft still advances the matter. The research assistant’s data cleaning requires supervision and still moves the project forward. The analyst’s clumsy background review may save a senior employee several hours. The organization receives a modest product while the worker develops.
AI can make that bundle come apart. Once the system produces more of the immediate value, hiring a novice begins to look like paying explicitly for someone else’s future expertise. A firm can still protect time for independent attempts, pay senior people to teach, and accept lower short-term output while someone learns. The expense simply becomes more visible at the same moment the immediate business case becomes weaker.
Each firm may prefer to automate junior work and hire experienced people later. That decision can be rational for every organization and impossible for the occupation.
These contradiction would take time to show up. Senior workers would remain, and firms might pay a premium for people trained before automation or inside the shrinking set of institutions that maintained serious apprenticeships. Eventually, employers would complain that candidates no longer possess the judgment their entry-level systems had stopped producing.
That is my forecast, not a measured outcome. AI may instead make training faster, cheaper, and less dependent on sitting near the right senior person. The old system distributed opportunity badly and supplied cases almost at random. A trainee might spend a year seeing the same easy problem; another might enter during a crisis and receive five years of education in six months, along with several new opinions about sleep.
AI can provide cases on demand, alter one fact at a time, explain the source of an error, simulate rare events, and revisit a weakness until the learner improves. It can give detailed feedback to someone at a small firm, public university, rural hospital, or overextended government office whose senior staff have been meaning to establish a mentorship program since 2018. A novice can ask the elementary question she would conceal from a supervisor and rehearse before another person bears the cost of her first attempt.
The old system is not entitled to survival merely because learning sometimes happened inside it. The real choice is whether the new system is designed to produce people as well as answers.
That requires identifying the developmental function of work before removing it. Which tasks expose novices to useful variation? Which force them to make a judgment rather than approve one? Where does feedback come from? Some tasks will fail the test and can disappear without ceremony.
Where the function matters, it has to be rebuilt deliberately. A learner may form an initial view before seeing the system’s answer. Assistance can recede as performance improves. Practice cases can include plausible AI mistakes and rare exceptions. Real responsibility still matters because simulation cannot fully reproduce a decision that counts.
Senior instruction also has to count as an output. An experienced employee who spends an hour examining how a beginner reached a conclusion has produced something even when no client document appears. An organization that records only throughput will treat the hour as a loss, remove it, and later circulate a survey asking why mentorship has declined.
None of this requires preserving a fixed number of junior jobs. The route can become shorter and less exclusive. The customer-support study suggests that some compression is real: new employees became better faster, and some improvement survived without the tool. The system may have replaced several months of weak, accidental instruction with something more useful.
The same result contains the temptation. Once a worker performs like someone with six months of experience, the company has little immediate reason to investigate whether she acquired everything else those six months once supplied. The measured problem has been solved.
The unmeasured problem belongs to the future.
* * *
The Second Order is a series within The Slow Panic. View the section alone at https://slowpanic.substack.com/s/the-second-order, or subscribe to the full publication.
Notes
1. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work,” The Quarterly Journal of Economics 140, no. 2 (2025): 889–942, https://doi.org/10.1093/qje/qjae044. The study examined 5,172 customer-support agents at one Fortune 500 software company. It found a roughly 15 percent average increase in issues resolved per hour, with larger gains among less-experienced and lower-skilled agents, and evidence that some gains persisted during AI outages.
2. Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman, “Generative AI without Guardrails Can Harm Learning: Evidence from High School Mathematics,” Proceedings of the National Academy of Sciences 122, no. 26 (2025): e2422633122, https://doi.org/10.1073/pnas.2422633122. A correction was published in Proceedings of the National Academy of Sciences 122, no. 34 (2025): e2518204122, https://doi.org/10.1073/pnas.2518204122.
3. Rose E. Wang, Ana T. Ribeiro, Carly D. Robinson, Susanna Loeb, and Dora Demszky, “Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise,” arXiv preprint arXiv:2410.03017, initially posted October 3, 2024, https://doi.org/10.48550/arXiv.2410.03017. The preregistered randomized experiment involved 900 tutors and 1,800 students and remained a preprint at the time of this essay.
4. Matthew Beane and Callen Anthony, “Inverted Apprenticeship: How Senior Occupational Members Develop Practical Expertise and Preserve Their Position When New Technologies Arrive,” Organization Science 35, no. 2 (2024): 405–431, https://doi.org/10.1287/orsc.2023.1688. The comparative ethnographic research examined urological surgery and investment banking.
5. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab working paper, November 13, 2025, https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/. The authors reported a 16 percent relative employment decline among workers aged twenty-two to twenty-five in the most AI-exposed occupations after controlling for firm-level shocks.
6. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, “Canaries, Interest Rates, and Timing: More on the Recent Drivers of Employment Changes for Young Workers,” Stanford Digital Economy Lab, February 9, 2026, https://digitaleconomy.stanford.edu/news/canaries-interest-rates-and-timinga-more-on-recent-drivers-of-employment-changes-for-young-workers/. The authors emphasized that their analysis could not establish AI as the sole cause and discussed the later timing of the employment divergence under their most demanding controls.
7. Seyed Mahdi Hosseini Maasoum and Guy Lichtinger, “Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data,” working paper, revised November 11, 2025, https://doi.org/10.2139/ssrn.5425555. The paper used résumé data covering nearly 62 million workers at about 285,000 firms and identified adoption through job postings seeking workers to implement generative AI.
8. Anders Humlum and Emilie Vestergaard, “Large Language Models, Small Labor Market Effects,” NBER Working Paper no. 33777, revised October 2025, https://doi.org/10.3386/w33777. The study linked representative adoption surveys to Danish administrative employment records and reported precise null effects on earnings and recorded hours, ruling out average effects greater than 2 percent during the first two years.
9. Daron Acemoglu and Jörn-Steffen Pischke, “Beyond Becker: Training in Imperfect Labour Markets,” The Economic Journal 109, no. 453 (1999): F112–F142, https://doi.org/10.1111/1468-0297.00405. The article explains how labor-market imperfections can give firms incentives to finance general training despite the portability of the skills produced.