Building an AI Company

Tom Blomfield, YC General Partner, on how AI changes the shape of a company — not just its speed. Watch the talk on YouTube.

The Old Shape vs. the New Shape

The old company was a Roman legion. Information moved up and down the hierarchy, and human beings were the conduit. AI breaks the assumption that this is the natural shape of a company.

Copilots are the small idea. Making engineers 20% more productive is just putting a bigger engine on the old machine. The bigger idea is to reimagine what a company is and how it acts.

The Company Brain

The company's real operating system is its domain knowledge. It lives in people's heads, Slack messages, emails, Notion, office hours, support tickets, product telemetry, and code changes. Make that knowledge legible, and the company can start becoming intelligent.

AI is not something you bolt onto the side of a company. The AI-native company is a set of recursive, self-improving loops:

1. Sense the world
2. Make a decision
3. Use tools
4. Pass quality gates
5. Learn from the result
6. Loop again

Self-Improving Loops

The goal is a company that improves while you sleep. If the system can see where it failed, ask why, update the tool, change the skill file, open the pull request, review it, merge it, and deploy it — the company is no longer waiting for a manager to notice.

The holy-shit moment is not an agent answering a question. It is a monitoring agent watching every failed query and making the next version of the system better. That is not AI making one person 20% more effective; that is AI learning how to improve the company.

Product can become a self-optimizing loop. An agent finds the highest-friction part of the sales funnel, researches best practices, launches an A/B test, runs it for a week, picks the winner, deploys it, and does it again.

Customer support can become a product loop. Suggestions come in; agents triage them like a CPO and CTO, discard what does not fit the roadmap, and ship what does. Overnight, without waiting for a meeting.

Burn Tokens, Not Headcount

The next constraint is not how many people you can hire; it is how much intelligence you can apply to repeated work. Token usage is dumb and gameable, but it still points at who is actually experimenting.

Middle management was a coordination layer. If AI can summarize, route, monitor, escalate, and improve workflows, the coordination problem changes. Everyone becomes an IC: a builder, an operator, a directly responsible individual.

Committees are the wrong primitive. For anything important, you need a named human — not a group, not a committee, not a vague owner. AI can coordinate more work, but accountability still needs a person.

Recording and Knowledge

Record everything. If it was recorded, it happened to the AI; if it was not recorded, it did not happen to your intelligence. The company brain cannot learn from conversations that disappear into the air.

Raw recordings are not enough. You cannot shove 100,000 hours of meetings into a context window. You have to aggregate, synthesize, categorize, and leave breadcrumbs the intelligence can actually use.

YC took thousands of hours of recorded office hours, synthesized them into a new manual, and can now update it every month. Every new piece of advice is compared with the old manual and either incorporated or thrown away.

Preserve the data; throw away the software. Store the emails, DMs, skills, context, and know-how with care. Dashboards and internal tools are ephemeral: generate them, use them, discard them, and regenerate them when the models improve.

Where Humans Belong

Humans move to the edge of the company brain. They handle the places where intelligence touches reality: novel situations, ethical calls, high-stakes moments, co-founder breakups, emotional conversations, and sales rooms where a human still matters.