How to Build a Self-Improving Company with AI

The speaker opens by challenging the foundational assumption that companies should be organized like Roman legions — hierarchical structures where humans serve as conduits for information flowing up and down. Drawing on ideas from Jack Dorsey, the speaker argues that AI fundamentally breaks this model and enables a reimagining of what a company is.

The core concept introduced is the 'self-improving AI loop,' composed of five layers: a sensor layer (customer emails, support tickets, product telemetry), a policy/decision layer (rules governing AI behavior and human escalation), a tool layer (deterministic APIs and callable functions), a quality gate (safety filters and human review for high-risk decisions), and a learning mechanism that feeds real-world outcomes back into the top of the loop. The speaker argues that if every step of this loop runs with minimal human intervention, the system improves continuously — even while employees sleep.

As a concrete example, the speaker describes YC's internal AI evolution: starting with a simple query agent, then adding retrieval-augmented generation for richer answers, and finally adding a monitoring agent that observes every employee query, identifies failures, determines fixes, writes code, submits pull requests, and deploys changes overnight — so the same query succeeds the next morning without human involvement. This, the speaker says, was the 'holy shit moment' that distinguished true self-improvement from mere productivity gains.

The speaker extends this model to other business functions: product analytics agents that identify funnel friction, run A/B tests, and deploy the winning version autonomously; and customer service agents that triage suggestions, write code, and ship features without human involvement.

On organizational structure, the speaker argues that middle management is obsolete in this model, replaced by AI coordination. The remaining human roles are individual contributors (ICs) with clear, singular ownership — directly responsible individuals rather than committees. Humans are positioned at the edges of the 'company brain,' handling novel situations, ethical decisions, high-stakes emotional moments, and real-world interactions that AI cannot yet fully reach.

The speaker outlines practical steps for building this kind of company: recording everything (emails, Slack messages, office hours, conversations) to make the organization legible to AI; distilling or 'diorizing' that data into synthesized context rather than raw dumps; and treating software as ephemeral while treating business context and domain knowledge as the durable, valuable asset. As an illustration, the speaker describes how 2,000 hours of YC office hour recordings were synthesized into a 150-page, continuously updating user manual — a living knowledge base that improves as new advice is given.

The talk concludes with a challenge to founders: if they are small enough to build their company correctly from scratch using this AI-native model, they have no excuse not to.