Key ideas
- A company's hierarchy is a technology for moving information, and it can be replaced like any other technology.
- When all the work leaves a written record a model can read, anyone can query the whole company, so fewer people are needed to relay information.
- Block reduced its structure to three roles: people who build, people who own an outcome, and coaches who help both get better.
- Building is now easy, so the hard part is judging what's worth shipping and keeping the work pointed at what the company is for.
- A fork of your open-source tool shows demand, and the business is in the parts people don't want to run themselves.
- To get someone started with a new tool, make the first step small enough that it costs them nothing.
Jack Dorsey argues that a company’s hierarchy exists to move information around, and that once all the work leaves a written record a model can read, most of that hierarchy isn’t needed. These are my notes on his interview at HBR’s Innovation Summit, with the names, dates and Block’s announcements checked. The parts I’d reuse are the three roles Block kept, his test for when an open-source tool is a business, and his one-minute rule for getting people to try AI.
The company as a technology
Dorsey’s starting point is that the way a company organises itself (departments, meetings, reporting lines) is a technology, even though nobody calls it one. Its job is to distribute information, tasks and accountability. He says Block borrowed its version from Google, which borrowed from older companies, in a line he traces back through the railroads to Rome. The essay he wrote with Roelof Botha fills in that history, from the Roman army’s fixed span of control to the first organisation chart at an American railroad in the 1850s.
In that design the company’s intelligence sits in people’s heads, and the hierarchy is the route information takes between them. Each hop is slow and loses something, because someone forgets a detail or decides not to raise it.
Block is remote-first, so nearly every interaction leaves something written behind. Dorsey’s claim is that a model can read all of it, which puts the whole company’s working record on one surface that anyone can query and act on. He describes the result as a company that’s almost perfectly legible to itself.
Two changes at Block sit behind this. In July 2024 it stopped organising by business (Square, Cash App and the others) and went back to organising by discipline, such as engineering and design. Dorsey says that gave the company one set of tools and standards, which made the AI tools easier to roll out. In February 2026 Block cut about 40% of its staff, taking it from over 10,000 people to just under 6,000, and tied the cut to these tools.
Three roles
With the record open to everyone, Dorsey says the roles came down to three.
| Role | What the person does |
|---|---|
| Individual contributor | Builds, makes or operates something |
| Directly responsible individual | Owns an outcome, sets the roadmap for it and takes accountability |
| Coach | Is assigned to an individual contributor or a directly responsible individual to help them master their craft |
The coach is what used to be a manager, and each discipline has them. The difference he stresses is that it’s an assignment, not a chain of command. The other two roles get the agency: they can see everything, build most of what they want, and decide close to the customer.
People at the edge
Asked how bottom-up work stays on strategy, Dorsey describes a central intelligence that represents everything the company is doing, with people around its edge. The people do the aligning. They hold the values, the purpose and the sense of what’s strategic, and they decide which customer requests fit.
His reasoning is that building has become easy, so the hard part moved. What’s scarce now is the judgment to say whether something should go out to every customer, and whether it’s worth the cost of building and maintaining it. He compares it to aligning an AI system: people keep pointing it at particular goals.
He makes the same point about tools. He calls harnesses such as Claude Code and Codex, and forward deployed engineers, a bridge over what models can’t yet do, and expects that gap to close. What stays valuable in his view is a person with a strong point of view, taste, and the nerve to put it out.
World models
Dorsey says the industry’s focus on language models is one path to intelligence among several. The one he thinks fits Block is the world model, because Block isn’t predicting the next word. It’s predicting the next thing a customer might do.
The signal is money. He calls transactions a very honest signal, and says they let Block model how a customer or a seller spends their day, then predict what they’ll do next and how to help. The same idea turned on the company gives a model that suggests what the company should do next, which people then accept or reject.
He also argues that language models are inefficient and that useful models don’t need huge compute. His example is Dessa, Block’s voice model, which Block announced in October 2026 and says was trained on conversation audio with no text. It runs on a high-end laptop, and Dorsey wants it on a phone. It isn’t public yet: Block has said a small group of AI safety researchers gets access first.
His own habit for keeping up is three hours each morning reading two or three research papers on arXiv, and building with the tools to find where their limits are.
Open source as a signal
Block has open-sourced two of its internal AI tools. Goose is a coding agent it released in January 2025. Buzz, released in July 2026, is a workspace where people and AI agents share the same channels.
Dorsey’s test for whether one of these can be a business is whether other companies fork it. Stripe did: its engineering blog says its internal coding agents run on a fork of goose that it made early on. If people want a tool enough to take it and maintain their own copy, some of them will pay not to.
What they pay for, as he lists it:
- Hosting, so they don’t run it themselves.
- Inference, so they don’t connect all the models themselves.
- Connectors into their own systems, with the compliance, security and risk work done.
Each one gives the customer time back, which he calls the easiest thing to charge for. Open source also works as distribution, and as a reading on whether the thing is valued at all.
His measure of success is use, and especially use he didn’t plan for. He credits Clayton Christensen’s Jobs to Be Done and gives Twitter as the example: people started addressing each other with the @ sign to hold conversations, Twitter made that a feature, and it became a conversational platform nobody had planned. The method is to watch what many customers are trying to do and failing at, then build for that.
Depth in one discipline
One question was whether disciplines still matter when a designer can produce code and an engineer can produce an interface. Dorsey’s answer is that a discipline is a way of approaching problems, not a set of tasks. An engineer can ask a model for an interface and like how it looks, without the experience to tune how it feels or make it obvious at a glance. He wants people very deep in one discipline and able to get by in the rest.
He thinks younger people can still build that depth, and faster. He pictures a problem as the roots of a tree. He used to pick one root, commit to it for a long time, and start over if it was wrong. Now he can look a little way down ten or twenty of them before choosing, then go deeper or back out. Taste and judgment come from experiencing a lot and forming an opinion on it, and the cost of that experience has dropped.
The leader’s three jobs
Dorsey says he worked these out about three years into running Twitter, and that a flatter company doesn’t change them.
- Set up the environment: a clear purpose and mission, a culture people can describe, and how they treat each other.
- Make sure decisions get made, and made in context. Making every decision himself would be a failure, since it makes him a single point of failure.
- Raise the bar on what people think is possible, which includes questioning or deleting whatever made the company successful before.
The second gets easier with a legible record. He can see how a decision was set up, whether it worked, how fast a wrong one was noticed and what corrected it. He notes that most organisations only run retrospectives after something goes wrong. The healthy and rare version reviews good outcomes and everyday decisions too.
On his own decisions, he says he’s had to be comfortable looking less credible for a while when he’s sure of a change. He collects opposing views first, then keeps explaining why, and accepts that proving it is on him.
How to start
His advice for leaders outside technology is to start with yourself. Build a system, not a set of questions: something that gives you time back or teaches you something, that you can show other people. If you have children, build with them and watch how they learn.
For other people, he says you can’t ask them to change their minds, only show them what’s possible. Put the tool in front of them and ask for the simplest thing they can think of. Then it’s theirs to take further or not, and forcing it only adds friction.
The habit rule comes from James Clear’s Atomic Habits, which he credits: one minute a day for two weeks, and no more than a minute even if you want to. That gets the behaviour in place at no cost. After two weeks, extend it to five minutes, then longer.
His one worry about AI is power over it concentrating in a few labs. He says he doesn’t believe it can be contained, and that open models and many eyes on a system are what make it secure.
Key terms
- Directly responsible individual (DRI)
- The one person who owns a problem or outcome, sets the plan for it and answers for the result.
- Coach
- Dorsey's name in the interview for what used to be a manager: someone assigned to help a person master their craft. The essay calls the role player-coach.
- World model
- A model that predicts what happens next in some part of the world, such as a customer's next action, instead of the next word in a sentence.
- Fork
- A copy of an open-source project that another team takes and develops for its own use.
- Harness
- The software around a model that lets it use tools and carry out tasks, such as Claude Code or Codex.
- Jobs to Be Done
- Clayton Christensen's idea that people use a product to get a specific job done, which you learn by watching what they do with it.
Quotes from the talk
A company itself is a technology.