Insight · May 19, 2026
Burn Tokens,
Not Headcount.
The Roman legion org chart is the default shape of every company. AI breaks it. Here is the shape I am building Bttr into instead.
The Premise
Most companies today are still organized like a Roman legion. Nested hierarchy. Consistent spans of control. Humans wired as the conduit for information flowing up and orders flowing down.
The shape is so default we forget it is a choice. It made sense when humans were the only intelligence in the building. AI breaks the assumption underneath it. The talk this is built on is Diana's, with a thread from Jack Dorsey shoved into the front of it. Both arrived at the same place from different angles. The hierarchical company is not a law of nature. It is a coping mechanism for low-bandwidth communication between humans.
A company is not an org chart. It is a set of recursive, self-improving loops.
01 · The Default Shape
The legions were designed to project power from Rome to Hadrian's Wall. Information went up the chain. Orders came back down. Humans were the wire. Open any org chart today and you are looking at the same diagram with worse uniforms.
The first wave of AI adoption did not change this shape. It bolted a faster engine onto it. Copilots inside engineering. Drafting assistants inside marketing. Twenty percent productivity on the existing process. That is the wrong frame.
If every step of the loop runs without you,
the company compounds while you sleep.
02 · The Loop That Replaces It
AI is not something you bolt on. You reimagine what a company is. The unit of execution is not a person reporting to a person. It is a loop that runs whether you are awake or not.
Sensor layer.
Customer emails, support tickets, code changes, product telemetry. The loop only knows what it can sense.
Policy layer.
Rules about what the system is allowed to do without asking. What requires a human in the loop. The guardrails.
Tool layer.
Deterministic APIs. Query the database. Read the calendar. Send the invoice. Boring infrastructure.
Quality gate.
Safety filters. Human review on high-risk moves. The line between autonomy and oversight is set here, not at the prompt.
Learning mechanism.
The part most companies skip. The system looks at where it failed, writes the fix, ships it. By morning it knows more than it did last night.
03 · A Loop That Rewrites Itself Overnight
Diana's working example at YC was the moment the idea stopped being a slide. They built an agent that could query their database for relevant founders. Useful, but a sidekick. The shift came when they put a monitoring agent on top of every query every YC employee ran. When a query failed, the monitor asked why. Wrong tool. Missing skills file. Stale index. Bad view. It diagnosed the cause, wrote the patch, opened a merge request, had a second agent review and merge it, and shipped it.
The next human to ask the same question gets a system that already learned. Not twenty percent more productive. The system itself is structurally better than it was last night. Throw tokens at it and it gets better while you sleep.
04 · The Implications
Burn tokens, not headcount.
The constraint moved. Companies are getting to Demo Day with about five times more revenue per employee than eighteen months ago. The blunt measure now is everyone's token usage. Gameable, but directionally right.
Middle management is done.
Coordination is a problem AI does for free. Two roles really matter. A builder who is also an operator. A directly responsible individual to ship the thing. No committee.
Make the company legible.
Record everything. Partner emails. Slack messages. DMs. Office hours. If it was not recorded, it did not happen to the intelligence. The unrecorded company is invisible to the loop.
Software is ephemeral.
Code is cheap enough to one-shot internal dashboards. Store the data preciously and never throw it away. Throw the software away when the models get smarter and regenerate it from the same instructions.
05 · Legibility Is the Unlock
The YC user manual was five to ten years out of date. The fix was not a writing sprint. They had two thousand hours of recorded office hours from the prior three months. They asked the AI to regenerate the manual. By Sunday they had a hundred and fifty page document that was dramatically better than the one written by humans. Now it regenerates every month.
That only works because the input existed. The lesson is to capture the substrate. Office hours, partner emails, Slack threads, decision memos, client calls. Anything that defines how the company actually works. If it lives only in a head, it is not in the loop.
The company is a brain in the middle. Humans sit at the edge.
06 · Where Humans Sit
The brain is your data, your skills, your know-how. All of it legible to the system. Humans sit around the edge of that brain, where intelligence makes contact with reality. The conference. The novel situation. The ethical call. The moment a founder shows up thinking about breaking up with their cofounder. High-stakes, emotional, ambiguous. That is the human zone.
Notice what is not in the diagram. No middle layer of people shuttling information between the edge and the brain. The loop does that. Removing that layer is not a layoff narrative. It is a shape change. The roles that remain are higher leverage by default.
If you were building your company today,
would you start it in this shape?
If you are small enough, you have no excuse.
Share this perspective
Built from a talk by Diana at Y Combinator and a thread by Jack Dorsey, with Pete's framing of AI as company architecture rather than productivity tooling. Stolen liberally and rebuilt around what I am shipping at Bttr.
More Insights
Perspectives on product design, technology, and the forces shaping how people interact with the things we build.
