Rethinking Companies: From Roman Legions to Self-Improving AI Systems

Most contemporary companies are structured like Roman legions, using nested hierarchies for power projection. In these organizations, human beings serve as conduits for information flowing up and down the chain. Jack Dorsey has questioned the foundational assumption that hierarchy is the only effective method for organizing value, suggesting a need for alternative models in the age of AI.

Initial discussions around AI adoption often centered on achieving a 20% productivity boost for engineers. However, simply adding a powerful engine to existing processes is seen as a flawed approach to AI integration. The new objective is to fundamentally reimagine the company as an autonomous entity, rather than just improving old ways of working.

A company's domain knowledge, business knowledge, and know-how often reside within the heads of its employees, or are scattered across platforms like Slack, emails, and Notion. The core challenge for AI transformation is translating this internal information into a legible format that AI systems can understand and process. This move is crucial for going beyond merely bolting AI tools onto existing structures and truly transforming how a company operates.

The vision for a modern company is to function as a collection of recursive, self-improving AI loops. This concept is deemed highly important because it enables the company to continuously refine its operations and improve itself even when human oversight is not active. Such systems can continue to evolve and enhance their processes autonomously.

An effective AI loop consists of five essential layers. First, a sensor layer is responsible for collecting data from the real world, such as support tickets or telemetry. Second, a policy layer defines the rules, permissions, and logging mechanisms. Third, a tool layer provides deterministic APIs and database access for AI actions. Fourth, a quality gate incorporates evaluations, safety filters, and human review to ensure quality. Finally, a learning mechanism closes the loop, enabling the system to iterate and improve based on failures and successes.

A pivotal realization occurred with the implementation of monitoring agents that track every query performed by employees, identifying successes and failures. This continuous monitoring is crucial for the system's ability to self-correct. When failures occur, the system triggers automated updates to relevant tools, databases, or indexes, ensuring immediate rectification and improvement. This mechanism allows the AI system to autonomously submit merge requests to the codebase, effectively closing the improvement loop without human intervention, marking a significant step towards fully self-improving organizations.