Your company's world model is built from its communication

Jack Dorsey and Roelof Botha published an essay in March called From Hierarchy to Intelligence. It’s one of the clearest pieces of writing we have read on what AI does to the shape of a company, and we think it’s right. We also think the idea is much bigger than the company it describes.

Their argument starts with a question that sounds like it belongs in a history book. Why do organizations have layers of management at all? Their answer: hierarchy is a two-thousand-year-old technology for moving information across distance when communication is slow and expensive. The Roman army discovered that one person can only direct a handful of others, so it stacked people and their leaders into a pyramid. The Prussians invented the modern staff officer to route information and pre-compute decisions. The American railroads borrowed the same structure and drew the first org chart. Every reorganization since has been a negotiation with one stubborn tradeoff. If you narrow each manager’s span of control and add layers, the additional layers slow information down.

The essay’s move is to notice that the constraint has lifted. Routing information is exactly the thing AI is now good at. So the question is no longer “how many layers do we need,” it’s “do humans need to provide the coordination at all.” Dorsey and Botha say no. They describe a company built around four things instead of a chain of managers: capabilities, the atomic building blocks of what the company does; a world model that holds what’s true right now; an intelligence layer that reads the model and composes those capabilities into action; and the interfaces that deliver it. Two of those look different in every company. Block’s capabilities are its financial primitives, the payments and lending and card issuance and the rest, and its interfaces are the surfaces that carry them, Square and Cash App and Afterpay. Every company has its own version of both, whatever it actually does and the surfaces it does it through. What stays the same from one company to the next is the other two, the model and the intelligence layer. The model holds the current state of the business. The intelligence layer acts on it. Together they do what the middle of the org chart used to do, without the relay.

The part everyone will skip past

Block’s world model actually has two halves, and the distinction matters. One is a company world model, how the company understands itself, and that’s the half that stands in for the management layers. The other is a customer world model, a per-customer and per-merchant picture built, in the essay’s words, from proprietary transaction data. The transactions feed that second half. As the essay puts it, money is the most honest signal in the world: people can posture, but when they spend, save, send, borrow, or repay, that’s the truth. For a payments company sitting on an economic graph, that’s exactly right, and it’s the source of the compounding advantage the essay describes. Richer transaction data makes the model sharper, which drives more activity, which produces richer data.

Here’s the thing worth sitting with. The transaction-as-honest-signal claim is true because Block is Block. The essay even supplies the general test: what does your company understand that is genuinely hard to understand, and is that understanding getting deeper every day? For Block the answer is the flow of money. For almost everyone else, it’s not.

So the interesting question, the one the essay opens but doesn’t have to answer, is this. If you’re not a payments company, what’s your honest signal? What’s the raw material your own world model gets built from?

For the rest of us, the signal is in communication

We have argued before that communication is where intent lives. It’s buried in messages, in the email that says the deal slipped, the Slack thread where the launch date quietly moved, the customer reply that mentions a competitor. A company’s real state, as opposed to the state recorded in its systems, lives in what its people say to each other and to the outside world.

That’s the honest signal for a company that doesn’t run on transactions. Not because communication is purer than money. People posture in email too. But because for most businesses, communication is where the operational truth actually accumulates. Your CRM has the contact’s new title. The thread has the reason they’re unhappy, the favor they asked for, the thing they’ll churn over if it’s not fixed by Friday. One of those is a system of record. The other is the world model, and it comes in two layers: operational knowledge that changes by the hour, the live state of who owes what and what just slipped, which lives in the messages; and canonical knowledge that barely changes and only a human changes, the pricing and positioning a company writes down once. Almost nobody has built the operational one yet, because until recently nothing could read the whole stream and keep up.

This is the same point Dorsey and Botha make about why managers existed in the first place. The manager was the world model. They held the context, fielded the questions, and relayed the state of things up and down. When that context lives in a machine-readable model instead, the relay isn’t needed. The difference is where each reads the signal from. Block’s is money. For most teams it’s the message stream.

Even Block proves it. A transaction graph, however rich, is a system of record: it shows what a customer did, not that Block’s own launch slipped or a partner is cooling. That operational knowledge lives in Block’s messages like anyone’s, and the canonical knowledge above it is still kept by hand. The company model is those two layers, and neither is in the ledger. So communication isn’t only the signal for the rest of us. It’s the company-model signal everywhere, and a transaction graph is a system of record on top.

Two halves of one loop

If you take the architecture seriously, it has two parts that only work together, and they map onto the two jobs we think AI does for a team.

The first job is maintaining operational knowledge, the model of the business. We call it the brain: the canonical knowledge a wiki would hold, plus the operational layer that today lives nowhere, scattered across inboxes and threads and the heads of whoever was on the call. The brain is meant to be the company world model for a team that runs on messages rather than payments.

The second job is automating the work, the intelligence layer. We call these workflows: the thing that, grounded in the brain, moves a piece of work forward without a person relaying it. A message comes in, the workflow reads it against what the brain knows, and the next step happens. Draft the reply, route the task, update the record, escalate the one that needs a human. That’s the coordination the middle of the org chart used to do by hand.

Neither half is much use alone. An intelligence layer with no model is a clever agent improvising without context, which is most of what “AI in the inbox” means today. A model with nothing acting on it is a wiki nobody reads. The loop is the point: the model grounds the action, and the action keeps the model current.

Hierarchy doesn’t disappear, it moves

It would be too neat to say this kills the org chart, and it doesn’t. Hierarchy survives in at least two places.

The first is the work you stop doing in-house. It still gets done somewhere, and that somewhere usually has an org chart of its own. A team leaning on an outside firm is delegating to a hierarchy, not abolishing one. The pyramid moves off your chart and onto someone else’s.

The second is happening inside the intelligence layer itself. The way AI does a large piece of work today is already hierarchical: an orchestrator model takes a goal, splits it into parts, hands each to a sub-agent, and stitches the results back together. A lead agent directing a swarm of workers is a span of control with layers, the same shape Dorsey and Botha trace back to the Roman army, rebuilt out of models instead of people.

And those are just the structural reasons. The org chart did other jobs the essay sets aside. Someone still has to run performance reviews, train the juniors, and be the one name accountable when something breaks. A company can keep a thin layer of managers for those long after AI has taken over the routing.

So “from hierarchy to intelligence” isn’t the end of hierarchy. It’s hierarchy moving off the human org chart and into the machine. And why that matters goes back to where the essay started. The pyramid got tall because a person can only directly coordinate a handful of others. An orchestrator doesn’t hit that wall in the same place. It can fan out to far more sub-agents at once, and a layer is nearly free to add or collapse. The structure survives. The human limit that made it tall and slow is what gives way, and what’s left is a flatter, faster, cheaper hierarchy, most of it no longer made of people.

Where we actually are

We want to be precise about maturity here, because the architecture is easy to describe and hard to finish, and Dorsey and Botha are describing a destination too.

The brain exists today and holds operational knowledge. Workflows write to it as they run, so the model gets richer as work happens. The part we haven’t finished is the most important part of the whole idea: the brain updating itself automatically from the full message stream, the way Block’s model updates itself from every transaction. That continuous, communication-fed update is where we’re headed, not where we are. It’s coming, and we’re building toward it deliberately rather than claiming it early. The honest signal only becomes a compounding advantage when reading it is automatic, and we’re not there yet.

So we’re not announcing that we have built the company world model for everyone who isn’t Block. We’re saying the essay describes the right shape, that the shape generalizes past payments, and that for the rest of the economy the raw material is communication. That’s the bet this+that’s built around: the brain as the model, workflows as the intelligence layer, the message stream as the source that feeds both.

Why this is the same essay we keep writing

We have made versions of this argument before. One framed agentic AI as the third wave of productivity software, with communication as the layer everything reorganizes around. Another argued that AI is moving into the work surface itself rather than staying in a chat box. This is the organizational version of the same idea. When the model holds the state of the business and the intelligence layer acts on it, the work surface and the org chart are describing the same shift from two directions.

Dorsey and Botha wrote it about a company with a proprietary economic graph, and the specifics of how Block organizes around transactions are theirs, not ours to extend. What we take from it is the general claim underneath: the company of the next decade is organized around a model of itself plus something that reads the model and acts. For Block the model is built from money. For most teams it will be built from what they say to each other, which is also the thing AI can finally read at the speed it arrives.

Key takeaways

• Hierarchy exists mostly to route information; Dorsey and Botha argue AI can now do that, so companies can organize around a machine-readable model of the business plus an intelligence layer that acts on it.
• Block builds its model from transactions because money is its honest signal. That’s true because Block is a payments company.
• For teams that don’t run on transactions, the honest signal is in communication, where operational truth actually accumulates.
• We map this to two jobs: the brain (the model) and workflows (the intelligence layer), which only work as a loop.
• The brain holds operational knowledge today; the piece still coming is the brain updating itself automatically from the full message stream. We’re building toward it, not claiming it yet.