AI Won't Replace Your Juniors First. It Will Expose the Managers Who Press Play

The standard story goes like this: AI eats the entry-level work first, the career ladder loses its bottom rungs, and a generation of juniors never gets to learn the job. It is tidy, it is scary, and it fits the headlines about a soft graduate job market.

Karim Lakhani, the Harvard Business School professor who has run randomized AI experiments at BCG, Procter & Gamble and H&M, thinks the story points at the wrong people. Sitting across from Marina Mogilko on Silicon Valley Girl, he said the quiet part out loud:

"If anything, people that may be at risk are more in the middle management layers than in the junior layers from my perspective."

He was careful to say the question is not settled, and that the early-career market is soft for several reasons, AI being only one of them. But his observation from inside real companies is specific, and it deserves more than a shrug. You can read the full exchange in our timestamped breakdown of the interview.

The Contrarian Insight: Risk Follows Behavior, Not Seniority

Lakhani's argument has two halves. First, juniors are "fearless" with these tools. He describes undergrads who "whip up all these things so quickly." They have no legacy workflow to defend, so they just build.

Second, the middle layer is where he sees the most resistance. In his words, the 35-to-55-year-olds are the ones pushing back against AI usage and not changing, while the more enlightened senior people and the junior people "could run circles around most of the organization."

Read that again. The vulnerable group is not defined by pay grade or age. It is defined by one behavior: refusing to rebuild how you work. The title on your badge is a lagging indicator.

That framing connects to the sharpest idea in the whole conversation, which is a taxonomy of how people use AI.

The Press-Play Problem

In the BCG work, Lakhani and his colleagues saw three working styles. Cyborgs work jointly with the model, trading moves in a tight loop. Centaurs divide the labor: "I'm good at X, the AI's good at Y," then supervise what comes back. And then there is the third group:

"In terms of learning, the self-automatons do the worst because they have stopped thinking, they've stopped engaging in the activities"

In the interview, the host quoted a split of 60% cyborgs, 14% centaurs and 27% self-automators from the consultant study. Whatever the exact proportions in your own company, the shape of the failure is what matters. The self-automator "presses play" and forwards the result.

Now map that onto a middle manager's job. A large share of it is routing: summarizing what the team did for the layer above, translating strategy into tasks for the layer below, approving documents that someone else drafted. Every one of those activities is also something a model can do tolerably well on the first try. The manager who pipes model output upward is not a manager with AI leverage. In Lakhani's phrase, they are a meat interface to the silicon, and a meat interface is the cheapest component to swap.

Author's Comment: I find the junior-versus-manager framing less useful than the behavioral one. A 24-year-old who pastes model output into a deck without reading it is a self-automator. A 50-year-old who interrogates every claim is a cyborg. Seniority is a distraction; engagement is the variable.

First Principles: Why the Middle Is Structurally Exposed

Three mechanisms stack on top of each other.

1. The middle's product is information handling, which is exactly what LLMs commoditize. Individual contributors produce artifacts. Executives make a small number of high-stakes calls. The middle converts, compresses and relays. When the cost of compression drops toward zero, the relay role has to justify itself with judgment, not throughput.

2. Fluent output punishes passive review. Lakhani's research with Dell'Acqua and colleagues on the jagged technological frontier found that consultants with GPT-4 completed 12.2% more tasks and finished 25.1% faster on tasks inside the model's competence. On a task deliberately placed outside it, AI users were 19% less likely to reach the correct answer than people working without AI. The model sounded equally sure in both cases. In the interview he puts it plainly: the intelligence "is not perfect, but it talks really well," and it "can be very persuasive." A reviewer who skims a confident wrong answer is worse off than one who never had the tool.

3. Incumbency taxes adaptation. Juniors pay no switching cost. A manager with fifteen years of tuned judgment has to admit that part of that tuning no longer applies. Lakhani's observation is that the people with the most to lose are the least willing to experiment, which is the textbook shape of disruption.

There is a fourth effect that raises the stakes: the same research program found that a single professional with AI, in a P&G field experiment, performed on par with a two-person team without it. If one engaged person plus a model can match a pair, the coordination layer built to manage pairs and teams has to be rethought, not defended.

The learning side is just as ugly. Daron Acemoglu makes a related argument in our brief of his interview, where he explains why the pocket-calculator analogy breaks down for AI: a calculator never did anyone's thinking for them. Self-automation does. Lakhani's "stopped thinking" and Acemoglu's atrophy are the same failure seen from two disciplines.

Second-Order Effects: What Everyone Is Getting Wrong

If this is right, a lot of current corporate behavior is miscalibrated.

• Companies protect juniors and neglect managers. Training budgets go to "AI onboarding" for new hires, while managers get a memo. The people who most need hands-on exposure get the least.
• Leaders delegate the learning. Lakhani's email story (card 8 of our interview brief) is the best line in the interview: bosses' bosses once had assistants read their email aloud and dictated replies, and some "never did email or did the email 10 years later." Same pattern today, he says: "Somebody else's job, not my job."
• Founders mistake a thin layer for a moat. His blunt question to law firms and manufacturers is "why should you exist?" if all you are is a thin wrapper around a frontier lab's model. Greg Brockman makes the complementary argument in his interview brief: build products whose value grows as the models get smarter, rather than patching today's limitations.

The pattern is fractal. A person who only forwards model output, a team that only relays it, and a company that only resells it are all the same shape at different scales.

The Builder's Playbook

Here is what to change on Monday morning, whether you manage a team, lead a product, or write the code.

1. Run a delegation audit. Take your last ten AI-assisted tasks and label each one: cyborg, centaur or self-automator. The test is simple. If the model had quietly been wrong, would you have caught it? If not, you automated blind. Move those tasks to a tighter loop or take them back.

2. Make the model argue with you. Lakhani's own recipe is refreshingly mechanical: "don't believe what I'm saying either," counter-argue every proposal with three arguments against it, and always follow up with additional research that hunts for red flags and "black swans." Bake it into a reusable system prompt:

You are my adversarial reviewer. For every proposal I give you:
1. Give three independent arguments for why it is wrong or irrelevant.
2. Run a second pass looking only for red flags and low-probability failures.
3. Do not agree with me. Rank your objections by how expensive they would be if true.
4. List what evidence would change your mind.

If you want structured versions of this loop, our Recursive Reflection framework and the prompts in Effective Prompts That Make AI Admit It Doesn't Know are good starting points.

3. Never take the first answer. Lakhani repeats it twice in the interview: "I never take the first answer." He also advises spending tokens freely while the tooling matures: let a research agent go much deeper than its thrifty default.

4. Own the sign-off. Adopt a one-line rule: if you cannot defend a number or a claim with the chat window closed, you do not ship it. His line on termination is blunt: he has fired people who answered "the AI gave me that."

5. Redefine the job, not the tasks. The productive question is not "which of my tasks can AI do?" It is "what would this role look like if I did ten times more?" Managers who answer that, and then retrain their teams around it, are the ones Lakhani describes as accelerating.

6. Test your own wrapper-ness. Ask whether the next frontier model makes your role, product or company more valuable, or absorbs it. If you cannot answer, that is your first research project.

Practical Pitfall Avoidance Guide: Do not respond to this by making AI mandatory with a usage dashboard. Counting prompts rewards exactly the self-automator behavior you are trying to prevent. Measure whether people can explain, defend and improve what the model produced.

Takeaway

The question to ask about your career or your team is not "are we senior enough to be safe?" It is "are we thinking, or just forwarding?" Open your last AI-assisted deliverable and look for the place where you pushed back on the model. If you can't find one, you have already pressed play.