Jack Dorsey: Saying goodbye to traditional corporate hierarchies and moving toward intelligent agent architecture with the help of AI

Written by: Jack Dorsey and Roelof Botha, Block

Edited by: Yangz, Techub News

At Sequoia, we’ve found that speed is the best predictor of startup success. Most companies treat AI as a productivity tool. But few people focus on AI’s potential to change how we collaborate. What Block is demonstrating is a fundamental rework of organizational design—ultimately using AI to increase speed, turning it into a compounding competitive advantage.

Two thousand years ago, when the first organizational charts appeared in enterprises, the Roman army solved a problem that still troubles every large organization to this day: how to coordinate thousands of people spread across vast regions when communication is constrained.

Their answer was a nested hierarchical structure that maintains a consistent span of control at each level. The smallest unit was called a “contubernium,” consisting of eight soldiers who shared a tent, equipment, and a mule, led by a decanus. Ten contubernia formed a “century,” totaling eighty people, commanded by a centurion. Six centuries made up a cohort. Ten cohorts made up a legion, about 5,000 people. At each level, there was a designated commander with clear authority—responsible for aggregating information from subordinates and transmitting decisions back down. This structure (8 → 80 → 480 → 5,000) is, in essence, an information routing protocol built on one simple human constraint: the number of people a leader can manage effectively is roughly between three and eight. The Romans learned this through hundreds of years of war. Even today, the hierarchical chain of command in the U.S. Army follows a similar pattern. We now call it “span of control,” and it remains a core constraint for every large organization on Earth.

The next major transformation came from Prussia. In 1806, Napoleon’s army destroyed the Prussian army at the Battle of Jena. After that, a group of reformers led by Scharnhorst and Gneisenau rebuilt the army around an unsettling reality: you can’t rely on the genius of individuals at the top—you need a system. They created the general staff, a specialized class of trained officers whose job wasn’t to fight, but to plan operations, process information, and coordinate different units. Scharnhorst wanted these staff officers to “support incapable generals, providing the talents that leaders and commanders might lack.” Before the term “middle management” emerged, this was its prototype. The role of these professionals was to route information through complex organizations, pre-calculate decisions, and keep everyone aligned. The military also formally distinguished between “line” authority and “staff” authority. Line units advanced the core mission; staff units provided specialized support. Even today, every company still uses these terms.

This military hierarchy structure entered the commercial world through American railroads in the 1840s and 1850s. The U.S. Army borrowed engineers trained at West Point and assigned them to private railroad companies. These officers brought military organizational thinking. The line-and-staff hierarchy, divisional structure, and bureaucratic reporting and control systems—all of these were developed in the army, and only later were adopted by railroad companies. In the mid-1850s, Daniel McCallum of the New York and Erie Railroad Company created the world’s first organizational chart to manage a system stretching more than 500 miles and employing thousands of workers. The informal management methods that had worked on smaller railroads began to fail. Train collisions became frequent, causing casualties. McCallum’s organizational chart formalized the hierarchical logic the Romans had used: hierarchical authority, clear reporting relationships, and structured information flow. It became the blueprint for modern companies.

Later, Frederick Taylor (1856–1915), known as the “father of scientific management,” optimized how work was handled within that hierarchy. Taylor broke jobs into specialized tasks, assigned them to trained experts, and managed through measurement rather than intuition. This gave rise to a functional, pyramid-shaped organization—an optimized structure for the efficiency of information transmission systems—initially pioneered by the military and then commercialized by railroad companies.

During World War II, the functional hierarchy structure was tested for the first time under real pressure. The Manhattan Project required physicists, chemists, engineers, metallurgists, and military personnel to cross disciplinary boundaries and work toward a single goal under extreme secrecy and time constraints. Oppenheimer organized the Los Alamos Laboratory into several functional divisions, but he insisted on open collaboration across divisions, resisting the division patterns the military typically used. In 1944, when the nuclear bomb problem became the key bottleneck, he reorganized the laboratory around it and created cross-functional teams—unprecedented in mainstream U.S. corporate circles at the time. This approach worked, but it was an exception dominated by an extraordinary leader during wartime. The question facing the business world after the war was: could this kind of cross-functional collaboration become the norm?

As companies grew and globalization accelerated after World War II, the scalability limits of functional design became increasingly obvious. In 1959, Gilbert Clee and Alfred di Scipio of McKinsey published “Creating World-Class Companies” in the Harvard Business Review, providing a theoretical framework for matrix organizational structures that combine functional expertise with divisional structure. Under the leadership of Marvin Bower, McKinsey helped companies like Shell and General Electric implement these principles, striking a balance between central standards and local flexibility. This became the early form of “specialized” or “modern” companies that would drive global economic development after the war.

Over time, other frameworks emerged to deal with the complexities, rigidity, and bureaucratization of matrix structures. In the late 1970s, the McKinsey 7S framework proposed by Tom Peters and Robert Waterman distinguished between “hard S” (strategy, structure, systems) and “soft S” (shared values, skills, people, style). The core idea was: relying on structural elements alone isn’t enough. Organizational effectiveness requires alignment between cultural traits and the human factors that determine whether strategy can truly succeed.

In recent decades, technology companies have run bold experiments in organizational design. Spotify popularized the cross-functional squad model with short-cycle sprints. Zappos tried Holacracy, eliminating management titles entirely. Valve adopted a flat structure with no formal hierarchy. These experiments all revealed some limitations of traditional hierarchical structures, but none solved the fundamental problem. After Spotify scaled up, it returned to traditional management. Zappos experienced significant employee churn. Valve’s model also proved difficult to scale beyond a few hundred people. When organizations grow to thousands, they revert to hierarchical coordination, because no alternative information routing mechanism is strong enough yet to replace it.

This constraint is identical to what the Romans faced, and what the U.S. Marine Corps rediscovered during World War II: reducing the span of control means adding more layers of command—yet the more layers you have, the slower the information flow becomes. Two thousand years of organizational innovation have, at their core, been attempts to avoid this trade-off, but they’ve never managed to break it.

So what’s different now?

At Block, we’re questioning a basic assumption: that organizations must be built through hierarchy, with people as the coordination mechanism. Instead, we aim to replace the functions that hierarchy takes on. Today, most AI-using companies give everyone a “copilot,” which makes existing structures work slightly better without changing their essence. But what we’re pursuing is completely different: a company built as an “agent” (or mini AGI).

We’re not the first to try to go beyond traditional hierarchy. Haier’s RenDanYíShí model, platform-style organizations, and “data-driven” management are all real explorations of the same problem. What they lacked was a technology capable of truly executing the coordination function that hierarchy performs. AI is that technology. For the first time ever, a system will be able to maintain and continuously update the business model and use that model to coordinate work. Before this, humans had to transmit information through layers of management.

To do this, a company needs two things: first, some kind of “world model” about its own operations; second, sufficiently rich customer signals that make the model useful.

Block is remote-first. Everything we do produces artifacts. Decisions, discussions, code, designs, plans, problems, progress—all exist in the form of recorded actions. These are the raw materials for building the company’s world model. In traditional companies, managers’ job is to understand what their teams are doing and to pass that contextual information up and down through the hierarchy. In a remote-first company where the work is machine-readable, AI can continuously construct and maintain this picture: what’s being built, what’s stuck, how resources are allocated, what works, and what doesn’t. In the past, this information was carried by hierarchical structures; now it’s carried by the company’s world model.

However, a system’s capability depends on the quality of the customer signals you feed it. And money is the most honest signal in the world.

People may lie in surveys, and they may ignore ads. But when they spend, save, transfer, borrow, or repay—that’s fact. Every transaction is a fact about someone’s life. Block obtains data from the buyers and sellers of millions of transactions every day through Cash App and Square, plus data generated by operating merchants’ businesses. This gives the customer world model a rare capability: an understanding of the financial reality of each customer and each merchant built on continuously accumulating honest signals. The richer the signals, the better the model; the better the model, the more transactions; the more transactions, the richer the signals.

The company’s world model and the customer’s world model together form the foundation of a new type of company. In this kind of company, it’s no longer product teams building predetermined roadmaps, but building four things:

First: capabilities—atomic financial primitives. Fundamental financial functions such as payments, lending, issuing cards, banking, buy-now-pay-later, and payroll services are not “products”; they are hard-to-get and hard-to-maintain foundational building blocks (some have network effects and require regulatory approvals). They don’t have a user interface. They have goals for reliability, compliance, and performance.

Second: the world model. It has two parts. The company world model allows the company to understand itself—its operations, performance, and priorities—replacing information that previously flowed through management hierarchies. The customer world model is a representation of each customer, each merchant, and each market built from proprietary transaction data. It starts from raw transaction data and, over time, evolves gradually into a complete causal model and predictive model.

Third: the intelligence layer. It’s responsible for combining foundational financial functions into solutions for specific customers at specific moments, and proactively delivering them. For example, a restaurant’s cash flow tightens when seasonal lows arrive, and the model has already seen this pattern. The intelligence layer stitches together a short-term loan from the lending capability, uses the payments capability to adjust the repayment plan, and presents the solution to them before the merchant has even thought about seeking financing. Or, if a Cash App user’s spending behavior changes, the model links it to “moving to a new city.” At that point, the intelligence layer assembles a new direct deposit setup, a Cash App card with cashback categories optimized for the new community, and a savings goal calibrated based on their updated income. No product manager decided to build these two solutions. The foundational financial capabilities already exist; the intelligence layer recognizes these moments and combines them.

Fourth: interfaces (hardware and software). Square, Cash App, Afterpay, TIDAL, bitkey, proto. These are delivery interfaces, through which the intelligence layer delivers the assembled solutions. They’re important, but the core of value creation isn’t here. The value lies in the model and the intelligence layer themselves.

When the intelligence layer tries to build a solution but fails because a capability is missing, that failure signal becomes the product roadmap of the future. Traditional roadmaps—product managers’ assumptions about what to build next—are the limiting factor for every company in the end. Under this model, the customer’s real needs directly determine the product backlog.

If the product the company builds is this, then the question becomes: what should employees do?

A new organizational structure emerges from this, and it’s the opposite of the traditional model. In a traditional company, people are the agents, and hierarchies route all kinds of information. In the new model, intelligence lives inside the system; people are at the edge—and the edge is where actual action happens.

The edge is where intelligence meets reality. People can go into domains the model can’t yet reach, sensing things the model can’t sense—such as intuition, opinions, cultural context, trust levels, the atmosphere in a room, and more. They can make decisions the model shouldn’t make on its own, especially ethical decisions, new situations, and high-stakes moments where the cost of being wrong could mean life or death. A world model that can’t touch the world is just a database. But the edge doesn’t need management hierarchy to coordinate. The world model provides each person on the edge with the context they need, so they can take action without waiting for information to travel up and down the chain of command.

In practice, this means we define roles as three types:

Individual Contributors (ICs): They build and operate various financial capabilities, the models, the intelligence layer, and the interfaces. They are deep experts at specific layers of the system. The world model provides the background context that used to be supplied by managers, so individual contributors can make decisions at their own level without waiting for instructions.

Directly Responsible Individuals (DRIs): They own specific cross-domain problems, opportunities, or customer outcomes. A DRI might, for example, be responsible for merchant churn in a particular sub-area during a 90-day project, with sufficient authority to pull resources as needed from the world-model team, the lending capabilities team, and the interface teams. DRIs can stay focused on certain problems for a long time, or pivot to solve new ones.

People who do both management and execution: They participate in building work directly and are also responsible for developing people. They replace the traditional “managers,” whose primary job used to be information routing. These people will still write code, build models, or design interfaces. At the same time, they invest effort in helping people around them grow. They don’t spend time in status-sync meetings, alignment meetings, or priority negotiations. The world model handles alignment, the DRI structure owns strategy and priorities, and they own technology and people management.

There’s no longer a need for a permanent middle-management layer. Everything else that the old hierarchy used to do is coordinated by the system. Everyone is empowered, and their role is closer to real work and customers.

Block is in the early stages of this transformation. It will be a difficult journey, and some parts will likely break before they truly work. We write it now because we believe every company will eventually have to face the same problem we faced: what is the thing your company understands that’s truly hard for others to understand? Is that understanding getting deeper every day?

If the answer is “nothing,” then AI is just a story about cost optimization. You can cut headcount and improve profit margins within a few quarters, but eventually you’ll be swallowed by a smarter competitor. If the answer is that it’s getting deeper, then AI won’t merely enhance your company—it will reveal what your company really is.

Block’s answer is an economic graph: millions of merchants and consumers, the buying-and-selling counterparties of every transaction, and the financial behaviors observed in real time. This understanding compounds every second the system runs. We believe the underlying pattern—building a company as an agent instead of a hierarchy—matters enough to reshape how companies operate over the coming years. What Block has built so far is enough to show that this idea isn’t just theoretical (of course, we also welcome discussion and feedback from all sides to test and improve our thinking).

How fast a company operates depends on how information flows, and hierarchy and middle management block that flow. For two thousand years, from the Romans’ “eight-person squads” to today’s global enterprises, we’ve never had a truly viable alternative. Eight soldiers sharing one tent need a decanus; eighty people need a centurion; five thousand people need a commander. The issue has never been whether you need hierarchy—it’s whether people are the only option for the functions hierarchy performs. Obviously, they’re not anymore. Block is building the next era of organization.