A new Harvard Business School/INSEAD study of over 2,900 Y Combinator startups reveals something most AI coverage misses: the companies being built around AI aren’t just smaller—they're architected around a fundamentally different principle. Instead of using AI to make existing workflows faster, they embed AI directly into their products, moving knowledge work that used to require internal teams into what customers interact with directly. The result: organizations that operate with 25% fewer people, flatter hierarchies, and comparable valuations to non-AI peers—because the productive work happens in the product, not in the organization.

Hyunjin Kim of INSEAD and Rembrand Koning of Harvard Business School published their findings early this month. The numbers are striking. AI-native startups—those built with AI capabilities embedded in their products—operate at 25% smaller headcount than non-AI peers in the same industry and cohort. Their hierarchies are half a level flatter. Their engineering density is 13 percentage points higher. Entry-level workers and management layers are each roughly 15% lower. And yet investors value them at comparable rates, translating to significantly higher valuation per employee.

What matters isn't that they're lean. It's why they're lean, and what that reveals about a new organizational form taking shape in front of us.

The Product Channel, Not Just the Process Channel

Most discussion of AI’s impact on firms focuses on the process channel—how workers use ChatGPT or Cursor or Claude to be more productive at their existing jobs. That's real. But the Harvard study identifies something fundamentally different: the product channel.

Two-thirds of AI startups in their data embed AI directly into what they sell. Forty-three percent build products that autonomously perform tasks people used to do. Another 24% build tools that make expert workers dramatically faster. When firms codify intelligence, they can locate it inside or outside their organization. The choice matters enormously.

When firms embed AI intelligence into a product, they move productive work from inside the organization to inside what is sold. A traditional presentation service company scales by hiring teams to scope requests, coordinate design work, and deliver decks. Gamma, an AI presentation startup, scales by letting customers generate complete slide decks through the product itself. The firm still needs engineers to build and improve the product. It doesn’t need a proportional team of designers and copy editors for each additional customer.

Here’s the architectural consequence: When you embed intelligence in the product, you move the locus of coordination from inside the firm to inside the interface. Customers interact directly with the intelligence. (See my previous Forbes article on why AI is the new UI.) Internal workflows manage exceptions and improvements. Radically fewer requests queuing. Vastly fewer hand-offs to oversee. No managers needed to route routine work through a knowledge hierarchy. That's why the hierarchy flattens. That's why coordination layers disappear. Much of the work that used to require supervision, approval, and human judgment now happens through the product itself.

This is the architectural shift. When knowledge work that historically required internal workflows gets baked into the product, firms can serve more customers with fewer people doing coordination work. The distribution of what the firm builds changes. The shape of the organization flattens because customer requests route through the product, not through a management chain.

What This Competitor Looks Like

Here’s what these AI-native organizations actually look like:

They’re technically dense.

Forty-five percent of their workforce is engineering or science roles, compared to 36% at non-AI startups. Not because they hire differently on a per-head basis, but because they need fewer people in sales, operations, finance, and administration. The work that would have happened in those functions is happening in the product instead.

AI-native startups employ 15 percentage points fewer entry-level workers—not because they don’t value junior talent, but because building systems around AI requires architects, not scaffolders. The people you hire need to make consequential decisions: how to embed intelligence into products, how to integrate foundation models, how to handle edge cases that automation can't solve. That's experienced-practitioner work. The firm shape that emerges reflects the seniority of the problems being solved, not a preference for seniority itself.

They're geographically concentrated.

These firms cluster in Silicon Valley at higher rates. That’s not incidental. It reflects the importance of being near specialized AI talent, infrastructure, and the research institutions driving capability advances. It's a signal about where the labor pool for this organizational form currently concentrates.

They're capital-efficient at scale.

When you control for cohort and industry, AI-native startups raise about 30% more funding per employee and achieve valuations roughly 30% higher per head. This isn't because they're better quality (they raise similar total funding to non-AI peers), but because investors expect them to grow more revenue on the same headcount base.

The Competitive Implications

Three things should matter to you:

First, new entrants into your market now have operational leverage you many not be able to match without radical redesigns. A competitor with thirty people built around an AI-embedded product can serve segments of your market that would require you to hire dozens. This isn’t theoretical. FazeShift, a ten-person startup, automated accounts receivable workflows that competitors had to staff with teams of analysts. Legion Health operates an AI psychiatry platform with twenty-eight people; a comparable non-AI mental healthcare network runs hundreds. The architectural advantage is real, and it compounds as the startup scales.

Second, the talent dynamics are shifting faster than most organizations realize. AI-native startups don’t need the same organizational depth. They need different talent—smaller teams of more senior, technical builders. But that means they're pulling from the same limited pool of specialized engineers and architects that established companies depend on. Your best people can see the equity upside and the technical leverage of building something that scales without proportional headcount growth. The question isn't whether you'll lose people to AI startups. It's whether you'll lose the right people, the ones who can architect these systems.

Third, and most important: most established organizations have architectural constraints that AI-native startups simply don’t face. Your product decisions were made years ago. Customer contracts lock in specific delivery models. Org incentives reward maintaining legacy workflows. Technical debt compounds. This isn't a design problem you can solve with internal dialogue. It's a systems problem.

Before a competitor forces the question, you need to audit whether your existing architecture could support AI-native density, or whether redesigning it would require breaking assumptions you’re not willing to break. Can you embed intelligence into your product the way these startups do? Can you reduce coordination layers without breaking customer commitments? Can you shift from hiring coordinators and junior analysts to hiring architects and senior builders?

For some businesses—particularly those selling knowledge services—the answer might be yes. For others, the answer is genuinely no. And if it's no, that's not a design challenge. That's a competitive disadvantage you need to understand clearly, because a startup built around a different architecture will eventually force you to reckon with it.

There’s one thing the data can't tell us yet: whether these ratios hold as AI-native startups mature and scale. Do they stay lean, or does the organizational gravity of coordination reassert itself? Early evidence from the fastest-growing AI companies suggests they can remain flatter longer than traditional software companies. But most of the firms in this study are three to four years old. We don't yet know what happens at fifty-person scale, or two-hundred-person scale.

That uncertainty is actually the point. This organizational form is still being invented. The winners will be companies that figure out how to stay architected around AI even as they grow. That's a design problem, not a headcount problem. And it's worth thinking about now, before your market defines the answer for you.