Vijayanta Gupta

Field note

What I Learned Building an AI-Native Company With a Small Team

.5 min read

We set out to build SoftwareValue.ai with a small team and a working assumption that AI agents would behave like cheap engineering capacity. The assumption was half right. It was right enough to ship the first product in months, and wrong in the place that mattered most, which turned out to be the more useful lesson.

What we tested was simple. Could a handful of people, with agents doing the bulk of the production work, attempt a build that would once have required a team several times the size. The economics said yes before we started. Inference cost for a model of equivalent performance has fallen by roughly an order of magnitude a year since 2021, from sixty dollars per million tokens to a few cents over three years (Andreessen Horowitz, November 2024). When the marginal cost of generating a draft, a function, or a test approaches zero, the constraint that used to govern a small team, how much production it could afford, effectively disappears. The market has already shown what that unlocks. The leading AI-native companies run at three to four and a half million dollars of revenue per employee, above Microsoft and Meta and near Nvidia (CB Insights, July 2025). A small team is no longer a ceiling on ambition.

What we expected was that this capacity would compound cleanly. More generation, more throughput, more shipped. For the structured parts of the work it did. Agents are strong where the task is well-defined and the output is checkable: scaffolding, refactors, test coverage, first drafts of almost anything. There it has never been faster to build high-quality software (IEEE Spectrum, December 2025). The capacity multiplier is not a slogan. It is the part of the story that works.

What surprised us sat on the other side of that line. The more the team produced, the more of its time moved into deciding whether the production was any good. Generation got cheap. Judgment did not. And the gap between the two is wider than it feels in the moment, which is the part worth dwelling on, because it is where small teams quietly lose the time they think they are saving. In a controlled trial of experienced developers on their own codebases, AI tools made them nineteen percent slower, even though the same developers predicted they would be twenty-four percent faster and believed afterward they had been twenty percent faster (METR, July 2025). The perception gap is the trap. The work feels accelerated while the clock says otherwise, because the cost has moved from typing to reviewing, and reviewing is the part nobody instruments.

For us the same pattern showed up away from code. Copy, positioning, the editorial line on what the product even claimed: agents produced plausible versions of all of it instantly, and plausible was exactly the problem. The output cleared the bar for fluent and missed the bar for right. Deciding which draft was correct, which claim we could stand behind, which version sounded like us rather than like everyone, took the judgment it always took. AI removed the cost of producing the option. It did not remove the cost of choosing among options, and choosing is most of the job.

What changed as a result was how we staffed and sequenced the work. We stopped treating the small team as a production team with agents bolted on, and started treating it as a judgment team that happened to have near-infinite production. That inverts the usual hiring instinct. The scarce resource is not the person who can make the thing. Agents make the thing. The scarce resource is the person with the taste to know when the thing is wrong, and the domain depth to know why. That is consistent with where the broader market is moving: the work that survives automation is the higher-order, design-oriented judgment, while the routine production it used to sit on top of is what compresses (IEEE Spectrum, December 2025).

The operating implication is one line. In an AI-native build, the binding constraint moves from production to judgment. Capacity stops being the bottleneck almost immediately, and review becomes the thing that gates throughput. A small team that does not plan for this will feel fast and ship slowly, generating more than it can properly evaluate and mistaking volume for progress. The discipline is to invest in judgment ahead of production: clear standards for what good looks like, a named owner for the editorial and quality line, and review treated as real work with real time, not as a tax on the real work.

This is an early pattern from one build, not a law. The boundary between what agents can decide and what still needs a person will keep moving, and some of what gets reviewed by hand today will be safely delegated tomorrow. The test I am running now is in my own work, the writing and the advisory: how much of the judgment layer itself can be encoded into the standards and evaluations the agents are given, so that taste scales a little further than the single person currently holding it. The honest answer so far is some, not all. The capacity was the easy part. The judgment is the work.

Sources

Back to writing