Teamday
← Back

Tom Blomfield: How to Build an AI-Native Company

YC Root Access
agentsfuture-of-workenterprisestartupautomationbusiness

Why AI-Native Companies Replace Human Coordinators With AI Loops

Tom Blomfield, a Y Combinator General Partner, gave this talk at Startup School Paris. His opening caveat matters: "no one knows how to do this. So this is largely theoretical", drawn from hundreds of YC companies. His claim is that most firms bolt AI onto a structure designed 2,000 years ago, and that the real change is removing humans as the information relay.

The Roman org chart: Blomfield starts with the legion's tent-group-and-century hierarchy. "A man was the bridge in this entire process", sending orders down and reports up. Modern org charts, he says, still run on the same pattern, and AI-as-chatbot leaves it intact.

Humans as the gating mechanism: Copilots and agents that stop and wait for input just make each person 20% faster. "When you were sleeping, that system didn't work." A human who must approve every step is a bottleneck, whatever the tooling.

The AI loop: He describes a company as loops: real-world inputs (telemetry, support tickets, billing), a policy layer (permissions and logging), tools (internal APIs, MCP), quality gates, and a learning step. His view on gates: "There shouldn't be humans except in the most critical cases". A second, adversarial LLM can do the checking, such as reviewing code or asking whether a bank bot is giving financial advice.

The data agent that fixes itself: YC built an English-to-SQL agent over data on 7,000 companies and 20,000 founders. A second agent then reviews the day's queries overnight, finds failures such as permission gaps or broken indexes, and sends pull requests. Next day the same query works. Blomfield calls it the moment his thinking changed.

Office hours become a living manual: YC now has 3,000-4,000 hours of recorded office hours. Transcribing them lets AI rewrite the 500-page, 15-year-old internal manual, then power an advice agent that draws on all 16 partners rather than one person's memory.

The company brain: Agents with a virtual machine, web search, Slack history and file storage become "an AI employee". Agents then share what they learn, so intelligence lives in the system rather than the hierarchy. Humans stay at the edge for sales calls, trust, ethics and the room's atmosphere.

5 Instructions Blomfield Gives Founders Building AI-Native

  • Reduce humans, use tokens - Founders reach Demo Day and Series A with revenue that used to need far larger teams.
  • Kill middle management - Everyone is an individual contributor who brings working prototypes, not decks.
  • One directly responsible person - A single owner per outcome, not a committee.
  • Make everything legible to AI - Record meetings, ban Slack DMs, and make every action leave a written artifact. If it isn't recorded, it didn't happen for the AI.
  • Loop on real-world data - Record sales and investor calls, learn what fails, and even simulate a specific investor's questions.

What This Means for AI-Powered Organizations

The useful idea is not headcount reduction but the removal of humans as message relays: write policy, give agents tools, gate with a second model, and feed outcomes back in. It is a vision, not a proven practice, and Blomfield says so himself. For teams deploying agents today, the near-term step is making company data legible and letting a loop, not a manager, close the feedback cycle.