YC Root Access

How to Build a Self-Improving Company with AI

In this recent batch talk, YC General Partner Tom Blomfield breaks down how to build a self-improving company using AI. He'll cover how to create a series recursive, self-improving AI loops and explain why the founders who get this right will run companies that can improve while they sleep. 00:00 — Companies Are Roman Legions 00:54 — Copilots Are the Wrong Mental Model 01:55 — Extract the Domain Knowledge 02:24 — The Recursive Self-Improving Loop 04:12 — The Holy Shit Moment at YC 05:50 — Self-Optimizing Product and Support Loops 06:29 — Burn Tokens, Not Headcount 07:23 — Middle Management Is Over 08:05 — Make Everything Legible to AI 09:40 — Regenerating the YC User Manual 11:19 — Software Is Ephemeral, Context Is Valuable 12:18 — Where Humans Still Matter

TBTom Blomfieldhost

May 19, 2026·13m·Watch on YouTube ↗

📖 CHAPTERS

1. 1
   0:00 – 0:54

   Roman legions as the default company operating system

   Tom frames most modern org charts as Roman-legion-style hierarchies built to move information up and orders down through humans. He argues that this structure is an inherited assumption, not a law of nature.

2. 2
   0:54 – 1:55

   Why “AI copilots” are the wrong mental model

   He rejects the idea that AI’s main value is making individuals 20–30% more productive in existing workflows. Instead, he argues AI enables a redefinition of what a company is and how it operates.

3. 3
   1:55 – 2:24

   Extracting domain knowledge into shareable, machine-legible context

   The core asset becomes the company’s know-how: the messy, distributed knowledge across people, Slack, email, and docs. If that knowledge is made legible to AI, it can power new operating models beyond hierarchy.

4. 4
   2:24 – 4:12

   The recursive self-improving loop: a new blueprint for companies

   Tom introduces the idea of the company as a set of recursive loops that sense the world, decide actions, execute via tools, validate, and learn. When the loop runs with minimal human intervention, improvement continues even while you sleep.

5. 5
   4:12 – 5:50

   YC’s “holy shit” moment: monitoring agents that fix the system overnight

   He contrasts a helpful internal agent with the breakthrough: an agent that monitors usage, detects failures, and proposes system improvements automatically. The system can write code, open merge requests, get reviewed, and ship fixes so the next day’s queries succeed.

6. 6
   5:50 – 6:29

   Self-optimizing loops for product growth and customer support

   Tom generalizes the approach: build loops that continuously improve funnels, features, and service outcomes. Agents can triage inputs, run experiments, ship changes, and repeat—while humans define constraints and supervise high-impact decisions.

7. 7
   6:29 – 7:23

   “Burn tokens, not headcount”: new constraints on scaling

   He argues the limiting factor will shift from hiring to compute usage. Early signals show startups reaching dramatically higher revenue per employee, implying experimentation and token allocation become core management concerns.

8. 8
   7:23 – 8:05

   Middle management declines; IC ownership and DRIs become central

   Coordination work traditionally done by middle management can be offloaded to AI systems. Tom argues orgs will bias toward individual contributors with clear responsibility rather than committees.

9. 9
   8:05 – 9:40

   Make everything legible to AI: record, store, and summarize the organization

   To power a “company brain,” everything important must be captured as data. Recording conversations, emails, Slack, and meetings matters because unrecorded work is invisible to the system; then summarization/diarization makes it usable at scale.

10. 10
    9:40 – 11:19

    Regenerating the YC user manual as a living, self-updating artifact

    Using thousands of hours of recorded office hours, YC can regenerate a dramatically improved user manual quickly and update it continuously. The manual becomes a living knowledge base that can also serve as context for agents.

11. 11
    11:19 – 12:18

    Software is ephemeral; context and data are the durable asset

    Tom argues internal software (dashboards, workflows) should be treated as disposable because models can regenerate it quickly. The enduring value is the stored data and the organization’s encoded know-how, not the specific UI built this month.

12. 12
    12:18

    Where humans still matter: the edge of the company brain

    Humans shift to interfacing the AI “company brain” with the real world, especially in novel, ethical, and emotionally complex situations. He expects humans to remain crucial in high-stakes decisions and relationship-driven work like sales.

13. 13

    Closing challenge: would you build your company in this shape today?

    He ends with a provocation: founders still early enough can design for self-improving loops from day one rather than retrofitting later. The implication is that org design is now a competitive choice, not an inevitability.