- YouTube

### TL;DR AI enables companies to evolve from rigid hierarchies into self-improving recursive loops, where token usage replaces headcount and humans handle only high-risk, high-touch interactions at the edge. ## The Roman Legion Model Is Dead Traditional companies are built like Roman legions: nested hierarchies with uniform spans of control, power projected from a center to the periphery through command-and-control chains. People become pipelines for moving information up and down. As Jack Dorsey pointed out, the fundamental assumption was that hierarchy is the natural organizational form for economic value. AI shatters that. Most people still think of AI as a productivity tool—a copilot that makes engineers 20% more efficient. That’s like putting a more powerful engine into an old chassis. The real opportunity is to reimagine what a company *is*, not just turbocharge existing workflows. ## The Self-Improving AI Loop Instead of layering AI onto a hierarchy, think of a company as a set of recursive, self-improving AI loops. Here’s the cycle: 1. **Perception** – Sensors pull data from the outside world: customer emails, support tickets, code changes, cancellations, product telemetry. 2. **Strategy / Decision** – Rules about what can be done, when to ask a human, what must be logged. 3. **Tools** – Deterministic APIs that the AI can call: query a database, check a calendar, execute code. 4. **Quality Gates** – Evaluation checks, safety filters, human review for high-risk actions. 5. **Learning** – The system interacts with the real world, spots gaps, and feeds back to the top of the loop. If every step runs with minimal human intervention, the system gets better while you sleep. ## Real-World Examples **Example 1: The Monitoring Agent** Start with an agent that answers queries using deterministic tools (e.g., “When did I last have office hours with this company?”). Next, it learns to make introductions by combining database queries with RAG. That’s last year’s version—a 20-30% efficiency boost. The real leap came when we added a *monitoring* agent that watches every query: success or failure. If failure, it analyzes why—maybe a new tool is needed, a skill file must be updated, or a new database view is required. Now it writes code overnight, opens a pull request, has another agent review, merge, and deploy. By morning, the same human query succeeds. That’s AI finding its own improvement path, not just boosting productivity. **Example 2: Product Analysis** An agent analyzes the sales funnel for the biggest drop-off, researches best practices, runs an A/B test, picks the winner, and deploys it. It loops continuously—self-optimizing product improvement. **Example 3: Customer Service** Customer suggestions arrive. Agents role-playing as Chief Product Officer and CTO evaluate them: some are discarded, others align with the roadmap. Those get coded, deployed, and delivered to customers overnight—no humans in the loop. ## What This Means ### Burn Tokens, Not Heads In the last 18 months, YC companies at Demo Day generate about 5x more revenue per employee. That trend will continue through Series A and B. Soon your constraint is token usage, not headcount. Rough proxy: measure tokens per person. Obviously it can be gamed, but directionally it tells you where to invest your time. ### Middle Management Is Over Coordination problems are solved by AI, not by layers of managers. Only two roles remain: the **IC** (independent contributor, builder, operator). Every piece of work needs a **DRI** (Directly Responsible Individual)—no committees, no groups. You can build a company on ICs alone. ## How to Build a Self-Improving Company ### 1. Make Everything AI-Readable Record everything. All partner emails go into the YC database. Every Slack message, every office hour recording from the last 3-4 months. If it isn’t recorded, AI can’t see it. Use devices (phones, glasses, room mics) to capture audio, then apply speaker diarization and summarization to create “breadcrumbs” for the AI. **Case in point:** The YC user manual was 5-10 years old and stale. Haj took 2000 hours of recorded office hours, fed them with instructions to summarize by category (fundraising, hiring, co-founder disputes), and regenerated the manual. In a weekend he produced 150 pages of far higher quality. Now it can be updated every month—every new piece of advice is compared, integrated, or discarded. That manual becomes a living brain, and can be pumped as context into an AI agent that embodies the collective wisdom of 16 partners. ### 2. Generate Self-Improving Artifacts If an artifact can be improved by the AI loop, keep it. If not, discard it. Everything that isn’t generating a loop is noise. ### 3. Treat Dashboards as Disposable Internal operational software is ephemeral. Codex 55 is already good enough to generate most simple internal dashboards in one shot. All internal ops teams should sit on top of the AI understanding layer and create on-demand dashboards. Store data forever (like Gary storing every email as Markdown), but treat software as throwaway. Business context and skills are valuable; the software is a temporary instantiation. Rebuild it when models get smarter. ## The Human Role: At the Edge The company’s “brain” consists of all data, emails, DMs, skills, know-how. Humans sit around that brain, interacting with the real world at points AI can’t yet reach. That includes meetings (though AI handles many calls), novel situations, ethical considerations, high-stakes moments (e.g., breaking up with a cofounder, sensitive sales conversations). For at least the next 20 years, sales will likely need humans. Humans are the touchpoint between the AI brain and messy reality. ## A Closing Question If you were to re-create your company today, would you build it this way from scratch? Most of you are small enough that you have no excuse. Some teams I know are already tearing down and rebuilding. --- **Source:** https://www.youtube.com/watch?v=t-G67yKAHBQ