How to Build a Self-Improving Company with AI
YC General Partner Tom Blomfield explains how to move beyond the 'copilot' mindset and restructure companies as series of recursive, self-improving AI loops. He details how to make company knowledge legible to AI, creating systems that improve overnight with minimal human intervention, ultimately rendering traditional middle management obsolete.
The foundational assumption that companies must be organized hierarchically, like Roman legions with humans as conduits for information, is being broken by AI. The initial view of AI as a mere productivity tool—a “copilot” to make engineers 20% more effective—is a limiting mental model. Instead of bolting a more powerful engine onto an old way of working, we can fundamentally reimagine what a company is and how it operates.
The Self-Improving Company as AI Loops
The new paradigm is to view a company as a set of recursive, self-improving AI loops. The central idea is to extract the implicit domain knowledge that lives in people’s heads, Slack messages, and Notion docs, and make it legible to an AI. This allows the organization to move from a rigid hierarchy to an intelligent, AI-powered entity that improves even while you sleep.
This self-improving loop consists of several key layers:
- Sensor Layer: Gathers data from the outside world. This includes customer emails, support tickets, code changes, product telemetry, and subscription cancellations.
- Policy Layer: A decision-making framework with rules about what the system can do, when it must ask for human permission, and what actions must be logged.
- Tool Layer: A set of deterministic APIs and skills the AI can call upon, such as querying a database or checking a calendar.
- Quality Gate: A verification step that includes heuristic checks, safety filters, and human review for high-risk or high-stakes actions.
- Learning Mechanism: The system interacts with the world, identifies failures or areas for improvement, and feeds that information back into the loop to enhance itself.
When this entire cycle can run with minimal human intervention, the company begins to learn and optimize on its own.
The YC “Holy [Shit]” Moment
A practical example at YC demonstrated this power. Initially, an agent was built as a “sidekick” for partners, capable of answering simple queries like, “When did I last have office hours with this company?”
The breakthrough came from adding a monitoring agent on top of the first one. This new agent observed every query from every YC employee, identifying both successes and failures. When a query failed, the monitoring agent would diagnose the problem: “What would have made this query work? Do we need a new deterministic tool, an updated skills file, or a different database view?”
Crucially, the system would then proceed to:
- Write the necessary code to fix the issue.
- Submit a merge request to the YC codebase.
- Have another agent review and merge the code.
- Deploy the changes.
The result is that when a human comes in the next day and asks the same query, it now succeeds. This isn’t just making a human 20% more effective; it is the AI itself going through a full loop to self-improve.
Implications of the AI-Native Model
Adopting this structure has profound implications for how companies are built and scaled.
Burn Tokens, Not Headcount
Companies will become constrained by token usage, not by the number of employees. We are already seeing YC companies achieve 5x more revenue per employee than 18 months ago. While a blunt metric, tracking token usage directionally indicates which employees are most effectively leveraging AI to create value.
Middle Management is Over
In a world where AI can handle coordination and information flow, the traditional role of middle management becomes redundant. The future organization consists of two primary roles:
- Individual Contributors (ICs): Builders and operators who are hands-on.
- Directly Responsible Individuals (DRIs): A single, named human accountable for getting something done, eliminating committees.
Make Everything Legible to AI
To build a “company brain,” you must record everything. Every partner email, Slack DM, and office hours session is recorded and fed into the system. If an event or conversation isn’t recorded, for the AI, “it did not happen.” This vast, aggregated context is what powers the intelligence.
For example, by feeding ~2,000 hours of recorded office hours into a model, YC was able to regenerate its entire 150-page user manual over a weekend. The new version is dramatically better, more up-to-date, and can now be continuously improved as new advice is given, creating a living document of YC’s collective wisdom.
Software is Ephemeral, Context is Valuable
The data, the context, and the business know-how are the enduring assets. The software built on top—internal dashboards, workflows, and tools—should be treated as ephemeral. With models like Code Llama 3 70B, you can one-shot generate most internal software. As models improve, you can simply throw the old software away and regenerate it from your original instructions and context.
The Role of Humans at the Edge
In this new model, humans operate at the edge, where the company’s AI brain makes contact with reality. People are essential for navigating novel situations, ethical considerations, and high-stakes, emotional moments that models cannot yet handle—like a co-founder dispute or a critical sales conversation. Humans reach into the places where models can’t go, interfacing with the real world while the AI core optimizes the system.