Blomfield and AI: Revolutionizing Business Organization

A Striking Analogy with Roman Legions

Blomfield, drawing inspiration from recent tweets by Jack Dorsey, uses a powerful analogy to describe modern companies. He compares them to Roman legions, which were designed to project the power of Rome to the farthest reaches of the empire through nested hierarchies. In this model, individuals had fixed spans of control, transmitting orders and relaying information. Blomfield emphasizes that most companies today are still organized this way, with humans serving as conduits for information flowing from top to bottom. This is not merely a productivity issue, but a structural assumption deeply embedded in our way of thinking about organizations.

According to Blomfield, AI does not just improve this structure; it breaks the underlying assumption. The old mental model involved adding AI copilots to your existing workflows, making engineers 20% more productive and delivering more code. Blomfield critiques this approach, likening it to adding a more powerful engine to a horse-drawn cart. He proposes a new mental model: redefining what a company is.

The Five Layers of the Self-Improving AI Loop

Blomfield suggests viewing the company as a set of recursive and self-improving loops, composed of five layers. If each step operates with minimal human intervention, the system improves while you sleep.

• Layer 1 — Sensor Layer: This is where real-world data enters, such as customer emails, support tickets, code changes, subscription cancellations, and product telemetry. This sensory layer captures signals from the outside world.

• Layer 2 — Decision Layer (Policy Layer): This is where the rules reside. The system can act autonomously, log certain actions, or request human intervention. It is the policy and decision layer, with safeguards for what the AI can and cannot do on its own.

• Layer 3 — Tools Layer: This is where the code executes. Deterministic APIs allow querying a database, checking a calendar, sending an email, or updating a record. Blomfield describes this as the layer of 'skills and code,' the actual execution mechanisms.

• Layer 4 — Quality Gate: Before an action is taken, it passes through a quality gate. This includes evaluation checks, security filters, or human review for high-risk actions. It is the checkpoint between "the AI has decided to do something" and "the thing actually happens."

• Layer 5 — Learning Mechanism: This is what makes the whole system self-improving. The system interacts with the real world, identifies where it has failed, and loops back. Failures become training signals, and the loop tightens over time.

The key idea is that if you can execute each step of this loop without human intervention, or with minimal oversight, your system improves automatically.

The Pivotal Moment at Y Combinator

Blomfield shares a significant experience at Y Combinator (YC), where they built a simple internal agent. This tool allowed partners to ask questions like "when did I last meet with this founder?" and received answers more quickly. While this was useful, it was not revolutionary, resembling a smarter search bar for their internal database.

Then, they added a monitoring agent on top. This agent tracked every request made by each YC employee, noting when they succeeded or failed. When they failed, it asked why, what could have made it work, and whether different deterministic tools or a new database view were needed.

Overnight, the agent wrote the fix, opened a pull request in YC's code, had it reviewed by another agent, and merged and deployed it. By the time a human arrived the next morning to make the same request, it worked. No human was involved.

"For me, it was like the 'Holy Shit' moment. It's not just AI making you 20 or 30% more valuable. It's AI going through this loop to discover how to self-improve."

This is the shift. Not a productivity gain, but a system that detects its own failures and corrects them.

Practical Application in Businesses

Blomfield did not stop at the internal example from YC. He described how this same loop structure applies to every business function.

• The Self-Optimizing Product Loop: Imagine an agent that scans your product analytics to determine which part of your sales funnel has the most friction. It searches for best practices, sets up an A/B test, runs it for a week, chooses the best-performing version, and deploys it. Then it starts again and again. A self-optimizing product loop, operating continuously, without a product manager needing to be involved in every cycle.

• The Automated Customer Support Loop: Customer suggestions come in constantly. An AI "CPO/CTO" makes decisions: this suggestion doesn't fit the roadmap, discard it. This one fits — write the code, deploy it, send it to the customer. No human involved in the decision, building, or delivery. The loop closes overnight.

These are not speculations. Blomfield described loops that exist or are currently being built in YC companies.

Recommendations for Integrating AI

Blomfield proposes four actions to prepare companies for this transformation:

• Make Your Company Readable for AI: If an interaction is not recorded, it does not exist for your AI system. This means every email must be in a searchable database. Every Slack message. Every recorded meeting. Every decision logged. YC took 2,000 hours of recorded office hours from the last three months and used AI to regenerate their entire user manual — a living document that updates with each new piece of advice given. This is what "readable for AI" means in practice.

• Burn Tokens, Not Headcount: The constraint is changing. You will hit limits on token usage before you hit limits on hiring. The directional measure of a high-performing team member right now: how much are they actually using AI? Who is maximizing tokens?

• Eliminate Middle Management: Middle management exists to solve a coordination problem. AI solves this problem better and faster. What remains: individual contributors (ICs) who build and operate, and a directly responsible individual (DRI) for each significant outcome. Blomfield: "I don't think you need middle management for this coordination problem. I think AI should handle that."

• Treat Software as Ephemeral, Context as Valuable: Models improve every few months. The dashboard you built six months ago? Regenerate it. The internal tool? Rewrite it with the latest model and better instructions. What you should never throw away: business context. Domain knowledge. The reasoning behind decisions. Blomfield: "The valuable part is the understanding in people's heads of how the function works. The software that overcomes it is ephemeral."

The Persistent Role of Humans

Blomfield is clear: humans are not disappearing. They are moving to the periphery — where your intelligence system interacts with reality. This includes new situations that models have not yet encountered, high-stakes and emotionally charged moments, such as a founder considering parting ways with their co-founder, or a tough negotiation with a client. Ethical decisions require genuine moral reasoning, and sales conversations remain essential. Humans will focus on tasks where AI cannot yet compete, ensuring a complementarity between AI and human intelligence.