Finn Brooks: Hey, Clara — okay, rough week for my inbox, honestly, because I've been staring at this Forbes piece since Tuesday and I cannot get the number out of my head.
Clara Bennett: The HBS working paper thing? I saw it too — the Y Combinator dataset.
Finn Brooks: Yes — okay, so today we're getting into this: Rembrand Koning at the Harvard Business School AI Institute and Hyunjin Kim at INSEAD looked at over 2,900 Y Combinator startups, 2020 to 2024, and AI-native firms are running 25% smaller than comparable non-AI startups. Twenty-five. And the question I can't shake is — if that's the future, where does the next generation of people actually learn to work?
Clara Bennett: And the part that stops me is the valuation line. These leaner firms aren't failing — they're hitting comparable funding and valuations. So the market is actively rewarding the organizational structure that cuts the bottom rung.
Finn Brooks: Which sounds like good news for founders and — wait, why are we both making that face?
Clara Bennett: Because 15% fewer entry-level roles and 15% fewer management layers — simultaneously — means it's not just lean at the bottom. The whole training pipeline is compressed. That's the tension we're sitting with today.
Finn Brooks: Okay but — hang on, that compressed pipeline thing, that's where I keep getting stuck, because it's not theoretical anymore. Nikkei Asia is reporting graduates from top U.S. schools applying to 8,000 jobs. Eight thousand. That's not a job search, that's a statistical long shot.
Clara Bennett: And the World Economic Forum and Cognizant put a number on the structural side of that: entry-level jobs in the U.S. down 35% in the last 18 months. That's the part that separates this from a bad hiring cycle.
Finn Brooks: Wait — 35% in 18 months?
Clara Bennett: In 18 months. Now, the analogy that comes to mind — imagine a restaurant that used to hire ten junior chefs to do all the prep work. Cutting vegetables, portioning, mise en place. Now a machine handles all of that. They only need the head chef. The food still gets made. But nobody in that kitchen ever learned to become a head chef, because that's not a role that exists anymore.
Finn Brooks: The ladder isn't just harder to climb — it's not there.
Clara Bennett: That's the core of it. And what's actually new — I mean, the 'AI will disrupt jobs' argument is not new, but what IS new is that it's measurable and it's landing right now. Through May 2026, companies have announced close to 90,000 job cuts directly tied to AI. That's not a projection. That's announced cuts, already on record.
Finn Brooks: Okay and the Revelio Labs piece is what actually — no, wait, this is the part that broke my brain. Early-stage tech startups raising 16% more capital, 16% fewer workers, compared to peers from five years ago. The money went up and the headcount went down simultaneously.
Clara Bennett: The Ramp and Revelio Labs data, yeah. And that's the signal, not the hype. The hype is 'AI kills all jobs forever.' The signal is narrower and more specific — the organizational model that venture capital is currently rewarding is structurally incompatible with building an entry-level talent pipeline. Those are different claims.
Finn Brooks: So what the projections — up to 15% of U.S. jobs eliminated over the next five years — what those might be describing is less 'the apocalypse' and more... the thing that's already quietly happened at the bottom of the org chart, just scaling out?
Clara Bennett: That framing is close — but the counter-signal actually makes the picture messier, not cleaner. Because TechCrunch flagged this explicitly: the Ramp and Revelio Labs report has a second finding that most people citing it just... skip past. Companies spending heavily on AI tools are growing headcount faster, including in junior roles. Which sounds like 'see, jobs are fine,' and I want to push back on that read hard.
Finn Brooks: Wait — the same Ramp and Revelio report? The one showing 16% fewer workers?
Clara Bennett: Same report. And that's not a contradiction to wave away — it's actually the tell. Because those are two different kinds of companies. An enterprise adopting AI as a tool, layering it onto an existing org, might genuinely hire more. But a startup where AI is the product — structurally, from day one — that firm has no reason to build the bottom rung at all. The distinction the Ramp data doesn't make cleanly.
Finn Brooks: The wrong take that keeps circulating — and I've seen this, like, three times this week — is 'the Ramp data proves AI creates jobs, so the Harvard finding is overblown.' And I actually... I mean, I wanted to believe it? But it doesn't hold up because you're comparing, what, a toolkit adopter to a company that literally is AI.
Clara Bennett: Exactly — and then Fortune and PwC pile on with 'roles have morphed, not disappeared.' Which sounds reasonable. But morphed into what, specifically?
Finn Brooks: No, I don't buy that.
Clara Bennett: The PwC framing is 'AI-augmented work' — which, in practice, probably means you need two or three years of domain expertise before you can even operate at that level. A recent grad doesn't have that. So morphed roles might be real and also completely inaccessible as an entry point. Those aren't mutually exclusive.
Finn Brooks: And CNBC's been reporting exactly that — this ladder-rung problem where junior roles weren't just jobs, they were the training mechanism. Remove them and the pipeline just... stops regenerating. Anyway, the part that makes all of this worse — we haven't even touched who's supposed to fund rebuilding that pipeline, and the VC incentive structure is a whole other problem.
Clara Bennett: Right — and in the meantime, a grad applying to 8,000 jobs doesn't experience which dataset is technically correct. The outcome is the same either way.
Finn Brooks: And that VC incentive thing — valuation-per-employee is the metric investors love, which means every junior hire is literally a drag on the number that determines your next round.
Clara Bennett: That's the mechanism. And it's not subtle — if your valuation story depends on how much revenue or equity value each employee represents, then an entry-level hire who needs eighteen months to become productive is, in that math, a liability. Not a bad person. A liability.
Finn Brooks: It's not even malicious. It's just — the incentive is right there.
Clara Bennett: Now here's what I want people to actually sit with — the senior engineers those same AI-native startups are hiring right now? Rembrand Koning and Hyunjin Kim's data shows these firms recruit experienced talent. But those experienced engineers got their experience in entry-level roles. Roles that, increasingly, don't get posted.
Finn Brooks: Wait — so it's a delayed collapse. They're spending down savings they don't know they have.
Clara Bennett: Three to five years out, maybe. And CNBC has named this — the ladder-rung problem specifically — where if AI-native firms skip junior hiring entirely and big tech becomes the only institution that still runs a training ground, the whole pipeline concentrates dangerously. One bad hiring freeze at a few large companies and the mid-level talent pool just... thins.
Finn Brooks: Okay that actually — I mean, picture a Carnegie Mellon CS grad right now, May 2024, genuinely strong resume, would've had callbacks in 2019. She's not getting rejected because she's underqualified. She's getting auto-rejected because the role doesn't exist at the company she applied to. And that company will want a senior version of her in four years with no idea where that person is supposed to have come from.
Clara Bennett: That's the watch item. Not whether AI eliminates jobs in the abstract — it's whether the organizations built on this lean model realize, in about three years, that they've been drawing down a talent pipeline they didn't fund. And at that point the correction is slow. You can't rebuild an entry-level training cohort overnight.
Finn Brooks: And nobody's actually answered that. Like — who funds the rebuild? Not Y Combinator. Not a founder optimizing valuation-per-employee. Maybe big tech, but Rembrand Koning and Hyunjin Kim's data is about startups, not Google. So if AI-native is where the gravity is going... I genuinely don't know who's left holding that responsibility.
Clara Bennett: I don't either. And I think that's — mm — that's actually the honest place to land. The Harvard Business School AI Institute study tells you what the organizational model looks like. It doesn't tell you whether anyone in that model has thought past the next round of funding to ask where the senior engineers of 2030 are supposed to come from.
Finn Brooks: That's the question I'm walking away with. Not whether the 25% headcount gap is real — it is — but whether the people building on top of it have run the math far enough out.
Clara Bennett: Yeah. Thanks for sitting in this one with me — it didn't get tidier.
