The company I am building would have needed thirty people in 2019
By Meet Patel · 2026-09-06 · updated 2026-09-17
Summary
AI has sharply reduced the cost of building and operating a software company, but only for a specific class of work: implementation, first-draft production, research synthesis and internal operations. It has not reduced the cost of distribution, trust, or deciding what to build — those remain the binding constraints. As a result, headcount-based benchmarks such as revenue per employee lose meaning, and small teams gain leverage while losing slack, which raises rather than lowers the cost of a bad decision.
Key Metrics & Takeaways
- 43 users
- in the first ten weeks, zero paid acquisition
- 64 signups
- 12.05% of visitors
- 70+ interviews
- the part that did not get cheaper
Ten weeks in, no paid acquisition, no sales team. Sixty-four signups, which was 12.05% of the people who visited. Forty-three of them actually used the thing. A hundred and thirty-four investigations run.
Those are small numbers and I would rather publish them plainly than round them into a story. The interesting thing is not the numbers. It is what it took to produce them, and how different that would have been six years ago.
The same product in 2019 needs a data engineering function, a backend team, a frontend team, a designer, an analyst, someone running infrastructure, someone doing research, someone writing everything. Not a huge company. About thirty people, and eighteen months, and the capital that implies.
What actually got cheaper
It is worth being precise, because the general claim that AI makes everything cheaper is not true and believing it will cost you.
The work that collapsed has a specific shape: the specification is clear and verification is fast. Implementation against a known design. First drafts of anything. Research synthesis across many sources. Internal operations — the reporting, the reconciliation, the routine checking that used to consume a person.
Where both conditions hold, the reduction is not incremental. It is closer to an order of magnitude, and it is the reason the thirty-person estimate is not hyperbole.
Where either condition fails, almost nothing has changed. If the specification is unclear, AI produces a great deal of confident work in the wrong direction, faster than before, which is worse than slow. If verification is slow, you cannot tell good output from plausible output, and you end up doing the work twice.
The short list of things that did not get cheaper
Distribution. I published thirty-seven blog posts and they got forty-eight views between them. Thirty-two of them had zero. Producing writing is now nearly free; getting anyone to read it costs precisely what it always did.
Trust. Nobody connects their revenue systems to a product because it was well built. That is earned slowly, by a person, in conversations, and it does not compress.
Knowing what to build. Seventy-plus interviews. Every one of them a scheduled call with a human at a time that suited them, most of them telling me something I did not want to hear. There is no version of that which is automated, and the moment you try, you get a synthetic answer to a real question.
Judgment about what matters. Which brings the whole thing back around: the constraint on the company I am building is the same constraint the product addresses. Not a shortage of work produced. A shortage of attention available to decide what deserves it.
Why the benchmarks are about to break
Revenue per employee has been a reasonable comparative for a long time because headcount was a reasonable proxy for capacity.
It is not any more, and it degrades unevenly, which is the worst case. Two companies with the same headcount can now differ several times over in operating output depending on how much of their work is done by systems. So the ratio does not become wrong in a predictable direction you could adjust for — it becomes noisy.
It will survive far longer than it should, because it is one line of arithmetic and everyone already has the inputs.
The part people leave out
Small teams get leverage. They also lose slack, and slack was doing something.
In a thirty-person company, somebody notices. Not through process — just because thirty people looking at the same business will produce someone who says hang on, that assumption stopped being true. That redundancy of judgment is invisible until it is gone.
A small team has no version of it. Every wrong assumption survives exactly as long as the person holding it holds it, and there is no ambient friction to knock it loose. I spent six months not answering one question about a segment that had been unprofitable since roughly month two, and there was no one whose job it was to be irritated by that.
So the honest position is that a bad decision costs a small team more, not less. The leverage is real and it multiplies whatever direction you are already pointing.
Which is why the first thing I would build for a small company is not more capacity to produce. It is something responsible for noticing — because production is the part that got cheap, and noticing is the part that did not.
Frequently asked questions
What does AI actually make cheaper when building a company?
Implementation, first-draft production, research synthesis and internal operations — work where the specification is clear and verification is fast. It does not meaningfully reduce the cost of distribution, of earning trust, or of deciding what to build, because those depend on other people changing their behaviour rather than on work being produced.
Why are revenue-per-employee benchmarks becoming misleading?
Because headcount is no longer a reliable proxy for capacity. Two companies with identical headcount can now differ several times over in operating output depending on how much of their work is done by systems, so a ratio built on the denominator of people stops carrying comparative information.
Do small AI-native teams have an advantage?
They gain leverage and lose slack. A small team can produce far more, but it has no redundancy of judgment — there is no second person who happens to notice a wrong assumption. That makes a bad decision more expensive, not less, and it is the main reason small teams should invest early in noticing rather than in producing.
Written by Meet Patel — founder of Company 8, building Dan (usedan.com). Dubai, UAE.