AI Productivity Has an Apprenticeship Problem

The Missing Half of the Productivity Story

Imagine a professional-services team that introduces an AI assistant and immediately starts delivering cleaner work. Drafts arrive sooner, basic mistakes become less common, and managers can assign more projects without expanding the team. Judged by this quarter's output, the investment looks successful. Now ask a different question: which employees are becoming capable of handling the difficult cases without someone more experienced standing behind them? The first set of results cannot answer the second question.

That distinction deserves a place in the current debate about AI at work. October research from Google and Anthropic approaches it from different directions: one studies how professional judgment develops, while the other examines changes inside an unusually AI-intensive employer. Together with earlier experiments, they raise a practical management problem. An organization can improve what it produces while remaining uncertain about the people it will need to supervise production later. The response should be to make learning visible alongside delivery, rather than assuming one automatically follows the other.

A Better Draft Does Not Settle the Question

In an October 7 research report, Google described a three-month randomized trial involving 133 patent lawyers at eleven firms that did business with Google. Access to an AI drafting tool improved the quality of assisted drafting. Researchers also tested participants on a separate patent-editing exercise after 90 days, with AI unavailable. That unassisted assessment produced a different pattern: experienced lawyers benefited, while junior lawyers showed no clear average improvement. Junior scores spread toward both stronger and weaker results.

This is evidence from a particular professional setting, tool and time period, not a verdict that AI makes junior employees less capable. The important feature of the experiment is its separation of delivered work from independent judgment. Better assisted output did not provide a reliable substitute for measuring what people had learned. Google's account of the study is the primary source for these findings.

For a manager, that creates a useful distinction between approving a deliverable and approving a promotion. A report can be excellent because the worker made excellent decisions, because the tool supplied them, or because a reviewer repaired the important parts. Those possibilities may look identical in a folder of finished work. They imply very different training needs.

How the Tool Is Used Changes the Learning Opportunity

An earlier study by Judy Hanwen Shen and Alex Tamkin examined developers learning an unfamiliar asynchronous Python library. In randomized experiments, AI assistance reduced performance on subsequent assessments of understanding, reading code and debugging, without a statistically significant average improvement in completion time in the main study. The researchers also identified different interaction patterns: some participants remained cognitively engaged by seeking explanations or asking conceptual questions, while others delegated more of the task. Those patterns were associated with different learning outcomes.

The experiment is narrow. A short library-learning exercise cannot establish what happens over an engineering career, and its observed interaction categories are not themselves randomized proof that a particular prompting technique causes better learning. Still, the study provides a reason to look beyond whether an employee has an AI license. Two people can use the same assistant while doing very different amounts of thinking.

Consider an illustrative debugging task. One worker asks for a corrected file and checks whether it runs. Another first predicts the cause, tests that prediction, and then asks the assistant to challenge it. Both may reach the same functioning program. Only the second workflow necessarily leaves a record of the worker's own diagnosis. That record gives a mentor something concrete to examine.

There Is Also Evidence for Learning With AI

The evidence does not support a blanket choice between AI and learning. The published version of Generative AI at Work, by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, studied a staggered rollout among 5,172 customer-support agents. It reported a 15% average increase in issues resolved per hour, with larger benefits for less-experienced and lower-skilled workers. It also found evidence consistent with worker learning. This was a different workflow and research design from either the patent-lawyer or programming experiments; the results should not be pooled into a universal score for AI.

A useful interpretation is that exposure to effective solutions can be educational when the work repeatedly connects recommendations with consequences. But that is a hypothesis to investigate in a particular deployment, not permission to assume that every helpful suggestion teaches its recipient. A team should be able to describe the mechanism: what does the employee notice, decide, test or remember that they would otherwise miss?

This also changes the argument for retaining junior roles. Keeping people occupied with repetitive work is not a training strategy by itself. If an assistant can provide clearer examples and faster feedback, withholding it could waste a valuable learning opportunity. The design problem is to preserve the decisions that build competence while removing effort that contributes little to it.

The Employer Already Living With the Trade-Off

Anthropic's October 16 report on its own workforce provides a timely organizational example. It says junior technical employees fell from 21% to 14% of technical headcount since early 2025, while their share of technical hires fell from 27% to 9%. Technical hiring still grew in absolute terms. The report also describes greater emphasis on judgment and autonomy in hiring, and employees spending more effort coordinating and reviewing agent work.

These are company-specific observations from an AI developer experiencing rapid growth, not a representative forecast for other employers. Anthropic explicitly notes that adoption was not random in its analysis of productivity, so differences around adoption cannot automatically be attributed to AI alone. Its internal experience matters as a case study, with the company's commercial position kept in view.

The strategic question for other firms is less about copying its staffing ratios than explaining their own succession plan. If a business chooses to hire fewer beginners while asking experienced staff to supervise more automated work, where will its next generation of experienced supervisors come from? Recruiting them from competitors might solve one employer's problem for a while. It cannot be the only development plan for every employer at once.

Reviewing Work Is a Skill That Needs Its Own Practice

Moving an employee from production to review sounds like a natural promotion. In practice, the two activities require different evidence of competence. A person who can make a plausible document more readable may still miss a wrong assumption embedded in its argument. Someone who can run a software test may not know whether the test covers the failure that matters. Assigning responsibility for verification does not create the ability to perform it.

This suggests a more precise definition of AI supervision. A capable reviewer should be able to identify the source of an important claim, describe what would make the recommendation fail, distinguish a harmless discrepancy from a consequential one, and know when to stop and seek help. None of that requires memorizing every detail that a computer can retrieve. It does require enough domain understanding to choose what deserves attention.

Training can therefore focus on deliberately selected decisions rather than preserving an entire manual workflow. For example, a team could take a completed, approved project and alter one assumption in a safe training copy. The learner's task would be to identify the change and explain its consequences before consulting an assistant. This is a proposed exercise, not a method validated by the studies above. Its value would have to be tested against actual learning and review performance.

The key is that an answer should come with a defensible reason. Simply asking employees to explain an AI-generated answer can produce a polished explanation of the same mistake. A stronger exercise requires them to check against independent evidence, show a conflicting possibility, or predict what a test will reveal before it runs.

Build a Learning Route Into the Workflow

A practical starting point is to separate production assistance from development exercises. In live work, teams should use the tools and review standards that produce the best safe result for the customer. They should not deliberately make client work slower or less reliable merely to create a lesson. Learning opportunities can be designed around sanitized examples, supervised drafts, simulations and retrospective reviews, with an explicit allocation of time.

For a new employee, one possible route begins with making a small number of decisions before seeing the assistant's recommendation. The person identifies missing information, writes a provisional conclusion and explains their confidence. They then compare the model's proposal with their own, resolve disagreements using evidence, and discuss a selected case with a mentor. As demonstrated competence improves, the employer can reduce the scaffolding and expand the employee's authority.

The same approach can work in reverse for experienced staff entering an unfamiliar area. Seniority in one specialty should not automatically count as expertise in every task an agent can attempt. A strong reviewer of marketing copy, for example, may be a beginner when asked to assess a statistical analysis. Permissions and training should follow the task and the evidence of competence, rather than a broad job title.

There are real costs here. Mentors need time, practice material needs maintenance, and useful feedback is harder to provide than a completion badge. A company should budget for these costs when estimating AI's return. If the business case works only by treating employee development as somebody else's responsibility, the apparent efficiency may depend on a capability it is consuming without replacing. That is a management risk worth identifying early, even when the current automation is genuinely valuable.

Measure the Capability You Expect to Rely On

A balanced evaluation would track at least three different outcomes: the quality and cost of work delivered with AI, the employee's performance on selected independent tasks, and the amount of expert intervention required. Combining them too early can hide the problem. A team might deliver more projects while requiring a growing share of one senior person's attention. Another might show little immediate speed improvement while developing people who can handle a wider range of work later.

Independent assessment should be relevant to the job. If staff will always have access to documentation, testing obscure facts from memory measures the wrong thing. A better exercise might allow reference materials while withholding generated conclusions, then ask the employee to diagnose a problem or challenge a recommendation. The purpose is to examine the judgment the organization plans to depend on, not to recreate an arbitrary pre-AI working environment.

Evaluators should also look at the distribution of results. An unchanged average could conceal some employees advancing quickly and others becoming more dependent. That calls for different coaching, not necessarily one company-wide restriction. Before using such results for consequential personnel decisions, the employer would need to establish that the assessment is fair, reliable and relevant, with employees understanding how it will be used.

Finally, compare progress over time rather than celebrating a single successful demonstration. Revisit similar reasoning in a different case, check whether the person recognizes a changed assumption, and record which interventions were necessary. These are proposed management practices, not guaranteed solutions. Their effectiveness should be judged with the same skepticism applied to a vendor's productivity claims.

A Sustainable Productivity Gain Leaves Expertise Behind

There is no reason to preserve every task simply because previous generations learned through it. Better tools should eliminate unnecessary effort. The harder question is which parts of that effort were doing a second job: giving people the experience to recognize an exception, challenge an answer or make a decision when the usual process stops working.

The research now emerging makes that second job harder to ignore. It does not establish a single future for work, or prove that employees must choose between speed and competence. It does make an output-only definition of success inadequate. An employer needs evidence that its people are developing the abilities its operating model still requires.

The strongest AI deployment plan should therefore explain two things together: how the organization will deliver better work now, and how it will develop the people trusted to direct that work later. When both improve, productivity becomes more than a temporary advantage borrowed from today's experts. It becomes a capability the business can keep building.

• Coding Agents Need a Workplace, Not Just a Model