The Manager's New Brief

13.1

Ch. 13 — Reskilling, New Roles, and the Changing Shape of Teams

Part IV — Management in the AI era

💡 Core idea

The manager’s job is to grow capability at least as fast as the work changes — including a path for juniors, not only higher output from the already skilled.

Adoption of a tool is not skill. A team can be busy with assistants and still be less able to understand, check, and maintain the work. If that happens, you have not upskilled anyone. You have borrowed speed and taken on cognitive debt (Storey, as discussed by Willison; see Chapter 12): work the team can no longer explain.

Expertise is not optional

A paradox sits at the center of the author’s research digest: tools make generalist production cheaper, and they reward expertise. Sean Goedecke, writing about how an expert actually uses a model — and using Terence Tao as an illustration — argues that you cannot prompt like an expert without the underlying domain mastery. The best “prompt” is still knowing the work.

Cong Wang treats taste as a scarce advantage once production is cheap: judgment compressed by time. That is the same scarce resource Chapter 6 called judgment, now seen as a career and team-design issue.

Goedecke also argues that getting good takes practice: when to add context, when to abandon a session, when to verify. These are long-term skills, not a tip you paste into a handbook once.

An Anthropic analysis of about 400,000 Claude Code sessions (October 2025–April 2026) found “persistent returns to expertise”: in a typical session, people made about 70% of the planning decisions and the model about 80% of the execution decisions. Sessions rated novice reached verified success 15% of the time, versus 28–33% for intermediate and expert sessions. Most of the gain was from novice to intermediate; the further gap to expert was modest. Occupation mattered less than task-specific expertise. The measures are transcript-based, not real-world outcome data, so do not turn them into a culture speech. They do support the claim this chapter is written for: expertise still pays.

What to grow, what will thin out

Skills that tend to appreciate: judgment, taste, domain depth, collaboration, problem framing, and the ability to verify someone else’s — or a model’s — work.

Skills that can commoditize: first-draft production, routine transformation of information, some forms of coordination.

Emerging work is better described as functions than as faddish titles: workflow owner, evaluator or editor of model output, conductor of agents (Fowler), data and quality steward — and, as Chapter 12 notes, Levie’s “agent deployer and manager” and Mensch’s managers who give agents specifications. Rewrite roles in writing: what you stop doing, what you start owning, how contribution will be evaluated. If you still reward volume alone, you will get volume (Chapter 15).

Molly Graham, updating “Give Away Your Legos” for the AI era, says the original advice was about change, not only hypergrowth: “The problem is that holding on doesn’t stop the change.” You can “believe in the opportunity and still hate the change,” and, as she quotes Chip Conley, “sometimes you need to throw the funeral” for a job or craft that ended. “Giving work to AI doesn’t work that way.” You still have to explain, review, and own the result: “You still own the cognitive load for the agents and workflows that you parent.” Do not outsource “judgment. Taste. Accountability. Relationships and trust.” Chapter 16 returns to the change conversation.

The apprenticeship problem

Marc Brooker asks, directly, what happens to juniors when senior work changes. If routine junior tasks disappear into tools, people still have to learn the craft somehow. Hiring fewer juniors without a new learning path is a choice, not a law of nature.

Gorintin and Le Bras make the same design point from the other side: collapsing technical execution is not a reason to stop hiring beginners. “Les juniors d’aujourd’hui sont les gardiens du discernement de demain. Mais seulement si nous leur donnons les conditions pour le développer.” See Chapter 8.

A SIEPR / Stanford brief (Mahoney, McEntarfer, and Wahal, 2026) concludes that AI’s impact on aggregate employment is likely small right now, while a tough market for recent graduates may be partly due to AI: unemployment among new graduates reached 5.6% in early 2026, up 1.6 percentage points from three years earlier. The authors cite Brynjolfsson, Chandar, and Chen on early-career declines in some AI-exposed occupations, and they also note serious confounders (interest-rate hikes, pandemic over-hiring). With added controls, those authors found the entry-level decline was not notable until 2024. Treat this as an open research question, not a settled fact — and do not use the uncertainty to wave away a missing junior path on your own team.

Seniors become teachers and reviewers. Juniors need a defined slice of real work, plus evaluation, exception handling, and someone accountable for their learning. If that sentence has no names on your team, the path does not exist.

Learn in the flow of the work

A 90-day upskilling plan should be short, practiced on live work, and reviewed. Pair skeptics with enthusiasts. Run drills: can people catch a bad output, ask a better question, or refuse a tool when the task is wrong for it?

An Anthropic randomized trial (Shen and Tamkin, 2026) with 52 mostly junior software engineers learning an unfamiliar Python library found that the AI-assisted group scored 50% on a follow-up quiz, versus 67% for those who coded by hand — 17 percentage points lower, or nearly two letter grades (Cohen’s d = 0.738, p = 0.01). The time saving was not statistically significant. A qualitative look at how people used the assistant found that those who asked for explanations or conceptual help scored better than those who delegated production; the authors do not claim that pattern is causal. The sample is small and the quiz was immediate, so it does not prove long-term skill loss. The safer practice is still: use tools to explain and to check, not only to finish faster.

Tobi Lütke, in “Learning on the Shop floor,” states the risk directly: “The risk is not that AI does the work. The risk is that AI does the work and we never learn from it.” If every interaction happens in a private window, “the only person who learns anything is the person at the keyboard.” His response at Shopify is to put the company agent in public channels so others can watch and reuse the work. You do not need that exact setup. You do need a way for the team to see, and learn from, how the work is done.

Mike Fisher adds the knowledge-transfer half. Quoting Michael Polanyi, “we can know more than we can tell.” “Documentation matters, but it only gets you so far.” Knowledge “survives when people have the chance to use it, teach it, and decide what is worth carrying forward.” Schedule a deliberate rebuild while the person who understands the work is still nearby; not everything should be preserved. AI makes this more urgent: “Code can now enter a system that nobody on the team ever had to understand.”

Andrej Karpathy makes a related cut: “The core idea is that this lets you skip writing but it doesn’t let you skip reading and thinking.” The surprising result, he writes, is that this works — if you still do the reading.

PwC US announced, in November 2026, a Critical Thinking and Judgment certificate program for more than 85,000 partners and employees, as AI takes on more of the foundational tasks through which people used to build judgment. That is a firm’s investment in practicing judgment, not evidence that judgment can be certified at scale. The usable signal is the bet: once routine production moves to tools, the scarce skill is evaluating information, challenging assumptions, and standing behind a call.

A field complement, not a contradiction, is Brynjolfsson, Li, and Raymond (2025): among 5,172 customer-support agents, a conversational assistant raised issues resolved per hour by 15% on average, and by 30% for less experienced and lower-skilled agents. Agents with two months of tenure and AI access performed as well as untreated agents with more than six months of tenure. The most skilled agents saw little speed gain and small declines in quality. Tools can flatten early performance and still starve the expertise the next model — and the next junior — will need.

Tools

🤖 AI-era tip: Measure learning, not only adoption

Treat literacy as a team sport. Prefer using the tool to explain and to check, not only to complete the task — that is the safer reading of the Anthropic learning trial, especially the qualitative patterns around explanation. Adoption metrics (“everyone has a license”) are not skill. Getting good takes practice. Do not shame people who are slower to trust a tool; they are often the ones who will find the failure mode.

Takeaways

Evidence