One Company Hit $50M in Revenue With Six Employees. Your Org Chart Is a Legacy System.
The benchmark that defined software for twenty years was $200,000 to $300,000 in revenue per employee. Good companies beat it. Great companies doubled it. The number was stable enough that investors used it to sanity-check a plan without opening the model.
That benchmark is now describing a world that no longer exists.
AI-native companies are posting $3.48 million in revenue per employee — roughly six times the traditional SaaS average, and in the extreme cases far beyond it. Cursor reached $500 million in ARR with fewer than fifty people, more than $10 million per head. Midjourney runs roughly $200 million a year with about eleven employees. One company crossed $50 million in revenue with six full-time staff and has stated it intends to reach $100 million with fewer than ten.
These are not rounding errors at the tail of a distribution. They are the new ceiling, and the ceiling moved by an order of magnitude in about three years.
What actually changed
It is tempting to attribute this to AI writing the code, and that is part of it — roughly 86% of organizations now use AI coding agents for production work, not experiments. Grindr reports that AI writes about 70% of its code, with engineering output up 2.5x since mid-2025.
But coding throughput is the least interesting part of the story. Engineering was never the majority of headcount at a $50 million software company. Support, sales ops, implementation, QA, content, finance ops, and the coordination layer required to keep those functions aligned — that was the headcount. A traditional org chart is mostly an apparatus for moving information between people.
What collapsed is the coordination cost. When a six-person company can run support, onboarding, and billing operations through agents that a single person supervises, the functions that used to require forty people require four. The org chart was never the work. It was the overhead of distributing the work across humans, and a large fraction of that overhead has become optional.
Companies above $5 million in revenue now average 20 to 50 employees, and the top AI-native performers run teams 40%+ smaller than comparable traditional peers. The median moved, not just the record.
The line item nobody modeled
Here is the part that is genuinely new, and it is showing up in operating reviews across the industry this year: for the leanest companies, the token bill now rivals payroll.
That is a structurally different business. Payroll is fixed, predictable, and slow to change in either direction. Inference cost is variable, scales with usage, and can move 10x in a month if a customer changes behavior or someone ships a prompt that loops. You have swapped a stable cost base for a volatile one, and most finance functions at small companies have no instrumentation for it.
This connects directly to the pricing problem we wrote about in $285 Billion Vanished From SaaS in 48 Hours. If your costs are variable and your revenue is per-seat and fixed, you have built a business that loses money precisely when it succeeds. Lean headcount does not save you from that; it makes the exposure sharper, because there is less fixed margin absorbing the variance.
Where lean actually breaks
We staff founding teams for a living, so we have watched the failure mode up close, and it is consistent enough to describe.
Small teams do not break on throughput. They break on surface area.
A six-person company can ship like a sixty-person company. It cannot be in as many places as a sixty-person company. Enterprise procurement, a SOC 2 audit, a security questionnaire from a bank, a regulated-industry deployment, a partner integration with a quarterly business review attached — each of those consumes a human who must hold context over months. Agents do not yet absorb that work, because the other side of the table is an institution that requires a named, accountable person.
The second failure mode is key-person concentration. At six people, every person is load-bearing. One departure in a critical seat is an existential event, not a hiring problem. Traditional orgs carried redundancy inefficiently but on purpose.
The third is the judgment bottleneck. Agents multiply execution, not discernment. A small team's output is gated by how fast its people can review, decide, and take responsibility. We have watched teams add agent capacity and get slower, because the humans became a review queue. If one person is approving everything, you have built a system whose throughput is one person.
How we staff a company now
Our actual working practice in the studio, as of this quarter:
- Start at three to five, not one. Solo is romantic and fragile. The minimum viable team is the smallest group that survives someone getting sick.
- Hire for judgment, pay for it, and give it leverage. The highest-leverage hire in 2026 is a senior operator who can supervise a fleet of agents across a function. That person replaces a department and costs less than four of the people they replace.
- Staff the surface area, not the throughput. Add humans where an institution needs a counterparty — enterprise sales, compliance, partnerships. Do not add humans to go faster; add agents.
- Instrument tokens as COGS from week one. Per-customer, per-feature. You cannot price what you cannot see, and this is the line item that will surprise you.
- Keep a written list of what the agents handle well, and revise it monthly. The teams that sustain their gains all do this. The boundary moves constantly, and an unexamined division of labor silently becomes wrong.
The honest version
Revenue per employee is a seductive metric because it makes a company look efficient without saying anything about whether the company is good. A business can post $10 million per head and be one competitor away from irrelevance.
What the lean-team era actually changed is not that small companies are better. It is that headcount is no longer evidence of anything. For twenty years, team size was a usable proxy for scale, seriousness, and capability — the same way seats were a usable proxy for software value, right up until they were not.
Both proxies broke for the same reason, in the same cycle. Measure the work.
