Picture a first-year analyst, three weeks into the job, when a partner asks a simple question about the market-sizing model in the appendix: "Why did you assume 4% penetration in year two?"
She didn't assume it. The agent did. She read the output, checked that the numbers summed, formatted the chart and moved on. The model was competent. The question was fair. And she has no answer. She isn't careless. Nobody ever made her build one from a blank spreadsheet at 1 a.m., get it wrong, and get torn apart for it.
Some version of this moment is going to become the defining talent problem of AI-first professional services. Most firms are treating it as a staffing question. It's actually a curriculum question.
The grunt work was never just grunt work
The consulting pyramid always had an open secret: junior analysts weren't only cheap labour. They were apprentices. The hundred-tab model, the 200 interview notes coded by hand, the third rewrite of slide 14 — these were how judgment got manufactured. You learned what a sensible assumption looked like by defending a stupid one. You learned how a client thinks by reading every transcript, not a summary of them.
Agents are now very good at exactly that layer of work, and the market has responded rationally. One firm cited by Middle Market Growth hired 15 people into its incoming class in 2021 and is planning for three or four in 2026. McKinsey, BCG and Bain have frozen starting salaries for the third year running. Even where hiring is growing — McKinsey says it will take on 12% more graduates in 2026 — the job those graduates walk into is not the one their partners did.
So we have a quiet structural break. The work that used to train seniors is being automated, but the need for seniors hasn't gone anywhere. If anything, it has grown.
Review is a senior skill being handed to juniors
Here's the part that should worry us most. When an agent drafts the deliverable, the human role shifts from producing to reviewing. That sounds like a promotion. It isn't.
Reviewing well is the hardest thing a consultant does. A partner can glance at a model and smell that the churn assumption is off because she has built forty of them and been burned by six. Review is compressed experience — pattern-matching against a library of mistakes you personally made.
We are now asking 23-year-olds to do the partner's job, evaluating output they couldn't have produced, without ever giving them the library. That isn't leverage. It's a credential with nothing underneath it.
The data hints at what this feels like from the inside. ActivTrak's 2026 State of the Workplace study, covering 163,000 employees across 1,111 companies, found burnout risk falling while disengagement risk rose 21% in a single year, to nearly one in four employees. Their reading: people whose capacity isn't being used. I'd add a sharper version. People who are busy but not growing stop caring. A job made entirely of approving things you don't understand is a slow way to check out.
Designing friction back in
The answer is not to ban agents from junior work. That's nostalgia, and clients won't pay for it. The answer is to treat apprenticeship as something that now has to be designed deliberately, because it no longer happens by accident.
Build before you delegate. Every new joiner builds one core artefact per engagement by hand — the model, the synthesis, the storyline — before handing that class of work to an agent. It's slower. That's the point.
Make the agent's reasoning the training material. Instead of only checking the final output, have juniors critique the intermediate steps: which assumption would you have changed, and why? Disagreeing with a machine sharpens judgment faster than agreeing with one.
Grade the defence, not the deliverable. In an agent-heavy team, polish is free. What isn't free is the ability to stand in front of a client and explain why year-two penetration is 4% and what happens if it's 2%. Assess that, explicitly and often.
Budget for it. If a firm captures, say, 30% efficiency from agents and reinvests none of it in development, it is quietly liquidating its future partnership. Apprenticeship is now a line item, not a by-product.
The bill arrives around 2032
None of this shows up in this year's margins, which is what makes it dangerous. A firm that hires four instead of fifteen and lets agents do the formative work will look brilliantly efficient for five or six years. Then it will go looking for engagement managers who can sense when a model is lying, and discover it never grew any.
The pyramid isn't collapsing. It's being hollowed out from the bottom rung up, and every rung we remove is one someone was supposed to climb. The question for anyone leading teams right now isn't how much junior work an agent can absorb. It's what, exactly, your juniors will have done by the time you need them to be senior.
