The Junior Pipeline Is the Hidden Cost

A partner at the firm I've spent the most time embedded with said something to me eight months ago that I haven't been able to shake since. We were talking about how well the AI rollout had gone on the research and first-draft side of the business — genuinely well, by every metric the firm tracks — and she said, almost as an afterthought: "the work is better and faster now. I just don't know where my partners in 2031 are going to come from."

That sentence is the entire essay, and I want to spend the rest of it taking the sentence seriously instead of doing what most rollouts do, which is celebrate the productivity win and leave the staffing question for a future leadership team to discover the hard way.

The apprenticeship model, made explicit

Every services firm I've looked closely at runs, whether or not anyone calls it this, on an apprenticeship model. Junior people are given work that is, deliberately, slightly beneath their real cost to the firm — research memos, first-draft analysis, routine client communications — not because that work is valuable in itself, but because doing a lot of it, badly at first and then less badly, is how a junior person becomes a senior person. The mediocre first draft a second-year analyst produces isn't waste. It's tuition. The firm is paying for the analyst's judgment to develop, and the mechanism of payment is giving them real work, with real stakes, before they're actually ready for it.

AI is extremely good at exactly the category of work that apprenticeship model runs on. First-draft research, routine analysis, the unglamorous grunt work that used to be a second-year analyst's entire job — a model now does a meaningfully better version of it, faster, at a fraction of the cost. This is not a hypothetical tension. On the team I've watched most closely, the volume of first-draft work routed to junior analysts fell by more than half in eighteen months, not because anyone decided to change the training model, but because it was simply more efficient, task by task, to route that work to the tool instead.

"Nobody decided to stop training juniors the old way. It just stopped being the efficient choice, one task at a time, until the training had quietly disappeared."

What actually happens to the tuition

Here is the part that took the firm the longest to notice, because it doesn't show up in any quarter's numbers. The tasks that got automated away weren't randomly distributed across a junior analyst's job. They were, overwhelmingly, the repetitions — the fifth market-sizing memo, not the first. And repetition, it turns out, is not incidental to how judgment develops. It's the mechanism. The first memo teaches almost nothing except the shape of the task. The fifteenth memo is where a junior analyst starts to develop the pattern-recognition that eventually becomes senior judgment — the instinct for which numbers look wrong before you've checked them, which client questions are really asking something else, which structural choices in an analysis will hold up under partner scrutiny and which won't.

Take away the repetitions and you take away the mechanism, even while leaving the junior analyst nominally "doing the same job." They still touch every task. They just touch each one once, reviewing and lightly editing a model's draft instead of producing fifteen of their own and building the callus that comes from having been wrong fourteen times before getting the fifteenth one right.

This is not a new anxiety dressed up in new clothes, either — every generation of services leadership has some version of the "are we training the next generation properly" worry, usually aimed at whatever the current efficiency fashion happens to be, offshoring, outsourcing, junior-headcount freezes during a downturn. What makes this version different is the speed and the invisibility. Offshoring a task is a visible line item someone has to approve. Routing a task to a model that already sits on every analyst's desktop requires no approval at all — it happens one task at a time, made by the analyst themselves in the ordinary course of getting the day's work done, and by the time the pattern is visible in aggregate, eighteen months of quietly missing repetition have already accumulated.

The five-year problem, made concrete

I asked the partner who said the original line to help me pull the numbers, because I wanted to know whether this was a real trend or an anecdote dressed up as one. It was real, and it was sharper than either of us expected. The firm tracks, informally, how long it takes an analyst to be trusted with unsupervised client-facing judgment calls — the moment a partner stops double-checking their reasoning by default. For the cohort that joined five years ago, before any of this tooling existed, that milestone landed, on average, around the three-and-a-half-year mark. For the cohort that joined eighteen months ago — the first cohort whose junior work was substantially AI-assisted from day one — early signs suggest the same milestone is tracking meaningfully later, even though this cohort's output quality in year one was, on paper, noticeably higher than their predecessors' had been at the same point.

That gap between output quality and judgment development is the whole hidden cost. A junior analyst assisted heavily by AI produces better-looking work sooner, which reads on every dashboard as faster development. But the work being better sooner is exactly what's masking the fact that the underlying judgment — the thing that actually needs to develop before someone can be trusted as a senior reviewer, let alone a partner — isn't accumulating at the same rate the output quality suggests it is.

The comparison that made it undeniable

The number that finally moved leadership past "interesting observation" into "budgeted problem" wasn't the milestone-timing gap on its own — it was a comparison the partner ran almost by accident, against a smaller sister practice inside the same firm that had, for unrelated reasons, been slower to adopt the tooling. That practice's junior cohort from the same eighteen-month window had none of the AI-assisted drafting the main practice had rolled out. Their output was, predictably, slower and rougher in year one. But when the two cohorts were put in front of the same panel of senior reviewers for a blind judgment assessment — evaluating each analyst's reasoning on a novel case they hadn't seen before, with no AI assistance allowed in the room — the slower cohort scored measurably higher on independent judgment, at the same point in their tenure, than the AI-assisted cohort did.

That result travelled around the firm faster than any of the earlier, softer evidence had. It's one thing to say "the milestone seems to be arriving later." It's another to put two cohorts of analysts side by side, strip away the tooling for an afternoon, and watch the assisted cohort visibly struggle with reasoning the unassisted cohort handled comfortably. Nobody in the room could credibly argue the AI-assisted analysts were less talented — the hiring bar for both cohorts had been identical, and by every other measure the assisted cohort had been the stronger performers on paper. The only variable that had changed was how much unassisted repetition each cohort had actually logged.

"We put both cohorts in a room with no AI and a case they hadn't seen. The gap wasn't subtle. That's when the finance team stopped calling this a training-philosophy debate and started calling it a risk."

It's worth being honest about what this comparison doesn't prove. It's one cohort, one practice, one blind assessment — not a controlled study, and the partner running it would be the first to say so. But it was enough, inside one firm's own risk appetite, to justify the deliberate re-insertion of unassisted repetition that follows, and I think that's the right bar for a result like this to clear: not academic proof, but enough of a signal that a firm stops treating the risk as hypothetical.

What the firm is trying instead

To the partner's credit, once the numbers were in front of her, the response wasn't to roll back the tooling — that would have been its own kind of mistake, giving up a genuine productivity gain to solve a problem that has a cheaper fix. Instead, the firm started deliberately re-inserting repetition into junior development, on purpose, in places where it had quietly disappeared. Junior analysts now do a rotating set of tasks without AI assistance, specifically chosen because they're judgment-building rather than output-critical — internal memos, not client-facing work, where a rough draft costs nothing and the repetition is the entire point. It's a strange thing to have to schedule deliberately, work that used to just exist as the ordinary texture of a junior job, but that's precisely what the shift required: the tuition had to become an explicit line item once the market stopped generating it as a side effect of ordinary business.

"We had to put the training back in on purpose. It used to just happen, as a byproduct of the work being genuinely too hard to hand to a machine. Now it doesn't happen unless someone schedules it."

The question every rollout plan should carry

I don't think this is an argument against the tooling, and I want to be careful not to let it read as one, because the productivity gains on the output side are real and worth having. It's an argument for a specific line item that almost no rollout plan currently carries: an explicit accounting of which tasks, among the ones being automated, were quietly doing double duty as junior training, and a deliberate plan for how that training happens now that the task itself is gone. Skip that accounting and the bill doesn't disappear. It just moves five years down the road, arrives all at once, and lands on whoever is trying to staff a partner track with a cohort that got faster output and slower judgment, in exactly the proportion nobody was measuring at the time.

I've started asking every leadership team I talk to on this desk a version of the partner's original question, early, before they've had a chance to celebrate their productivity numbers uncritically: if this tooling does the repetitions your juniors used to learn from, who is doing the learning now, and have you actually checked, or are you assuming it's fine because the output looks fine? Most teams haven't checked. The ones that have, almost without exception, found some version of the same gap the partner found — quietly, a year or two into a rollout that everyone had already filed under "success." The productivity win is real. It just isn't the whole ledger, and the part that's missing is the part that doesn't send an invoice until the cohort it was underinvesting in is the cohort you're trying to promote into the partnership.