In brief: Block is rebuilding itself so that an AI system, rather than middle management, carries information and coordination across the company. Whatever its commercial merits, that changes where accountability sits: it moves upstream, into how the system was designed, trained and monitored. ISO 42001 and the EU AI Act ask the right questions but were not written for AI as the organisation’s structure. Before moving in this direction, a board needs independent assurance of the system, defined escalation to humans, written accountability for its outputs, and monitoring that does not depend on the system itself.
At the end of March 2026, Jack Dorsey and Sequoia partner Roelof Botha published an essay setting out how Block, the financial technology company behind Square and Cash App, is restructuring itself around artificial intelligence. The essay describes AI as the organisation’s coordinating intelligence, going well beyond a productivity tool or a copilot for staff. It is serious, carefully argued and worth reading in full. It also raises a question most boards have not yet been asked to answer.
Block’s model may prove to be what its authors claim: a new form of organisation that gains a lasting advantage through speed. This article makes a narrower point. Moving AI from support tool to coordination layer changes the kind of governance an organisation needs. Existing frameworks only partly cover that change, and boards need to understand it before it reaches their own organisations.
Copilot or coordinator
Most organisations using AI today are, in Dorsey’s terms, giving everyone a copilot. The underlying structure stays the same. Humans make decisions, and AI helps them decide faster or with better information. The accountability chain stays intact: a person approved this, a manager reviewed it, a board set the direction.
Block’s model works differently. Its “world model”, a continuously updated picture of the company’s operations, priorities and performance, takes over what middle management used to do. Instead of information passing up and down a human chain of command, the system holds the context and distributes it. Three roles remain: individual contributors who build and operate the system, Directly Responsible Individuals who own specific outcomes for set periods, and player-coaches who focus on craft and developing people. The essay is explicit that there is no permanent middle management layer.
The essay presents this as an old information-routing problem finally solved. For two thousand years, from the Roman contubernium to the modern corporation, hierarchy existed because people were the only way to move information and decisions through a large organisation. The argument is that AI can now do that job better and faster, without the distortion that builds up as information passes through human layers.
The argument deserves serious attention. It rests on a real analysis of why organisations are structured as they are, and it identifies a real weakness of conventional hierarchy. The governance question is what it means for accountability, and whether boards are ready for that.
Where accountability goes
In a conventional organisation, accountability can be traced. A person made a decision, within a process, with the information available at the time. When something goes wrong, a board can ask who approved it, what they knew and what process they followed. The answers may be uncomfortable, but they exist.
When AI becomes the coordination layer, accountability moves upstream, into the design and training decisions that shaped the system. The world model produces an output because an engineering team defined its objectives, a data science team trained it on the data available, and a governance team, if there was one, set limits on what it could do.
That is harder for a board to question. Few boards are yet set up to ask what the system’s objective function is, what data it was trained on, how edge cases were handled, and when it was last independently validated. None of these needs technical expertise to ask. They are governance questions, but they need a different vocabulary, and different assurance arrangements, from those boards have built around human decision chains.
The risk of compounding error is particularly sharp in financial services, where Block operates. Research on AI in financial services warns that algorithms can pick up discriminatory proxies that people had not previously considered. Scale is what turns that into a board-level risk. A biased human manager makes biased decisions within a limited span of control. A biased world model makes them continuously across the whole organisation. If an AI layer handling lending develops a subtle error in its customer model, it can produce thousands of flawed outcomes before anyone notices. EY’s analysis of AI discrimination in financial services confirms that regulators now expect firms to demonstrate accountability for AI-driven decisions, and removing the human from the decision chain does not reduce that obligation.
What current frameworks offer
ISO 42001, the international management system standard for AI governance published in December 2023, is a useful starting point. It requires organisations to identify their AI systems, assess their risks, set policies and controls, and show ongoing monitoring. Applied rigorously to Block’s model, it would raise the right questions: who owns the world model as a governed AI system, what is the risk assessment for removing the human coordination layer, and how, and by whom, is performance monitored?
The EU AI Act, which entered into force in August 2024, adds obligations, particularly for high-risk AI. Under Annex III of the Act, AI systems used to make decisions affecting work-related relationships, allocate tasks based on individual behaviour, or monitor and evaluate performance are explicitly classified as high-risk. Financial services organisations adopting anything like Block’s model will need to work out where their systems fall in the Act’s risk classification, and what that means for transparency, human oversight and conformity assessment. Rules for high-risk systems were originally due to apply from August 2026, but the EU’s Digital Omnibus on AI, which entered into force on 27 July 2026, has moved that deadline to 2 December 2027 for standalone high-risk systems. That gives boards more time, but the obligations themselves are unchanged.
Both frameworks assume a particular structure: people using AI inside a managed organisation, with identifiable decision points and accountable individuals at each stage. They were not written for AI as the structure itself. They still help, because they prompt the right questions and their obligations are real and enforceable. But they do not yet give complete answers, and organisations adopting these models now cannot wait for standards and regulation to catch up.
What responsible adoption requires
Before a board considers moving in Block’s direction, four things need to be in place.
First, independent assurance of the world model itself. If the AI system is doing the coordination work that management used to do, it should be treated and assured as a critical system. That means independent review of its objectives, data, outputs and failure modes, by people other than the team that built it. Whoever monitors the AI must be separate from the part of the organisation that depends on it for operational intelligence.
Second, defined escalation to people. Block’s essay says humans make the calls the model should not make on its own, “especially ethical decisions, novel situations, and high-stakes moments.” That is a principle, and a principle needs a mechanism before it becomes a control. A board needs to know how the system recognises that a situation needs human judgement, who receives the escalation, how quickly, and what authority they have to act.
Third, accountability that survives the removal of the human chain. Without a middle management layer, the accountability structure has to be rebuilt around the system. Who is responsible for the world model’s outputs? Who can intervene when it produces an unexpected result? Who reports to the board on its performance, and on what basis? The answers need to be written down rather than assumed.
Fourth, monitoring that is independent of the system being monitored. A world model that updates itself continuously can drift by design. Objectives can shift a little at a time, biases can compound, and performance against the original intent can decline in ways the system itself cannot see. Independent, human-led monitoring with the authority to intervene is what keeps such a model trustworthy over time.
The question boards should be asking now
Whether or not Block’s model becomes the norm, AI is taking on more of the work that human management layers do, at Block and elsewhere, and the pace appears to be increasing.
The governance task is to have the right questions ready when it reaches your organisation. The organisations that handle this well are likely to be those that recognised early that adopting AI as a coordination layer is a governance decision before it is an operational one.
A board that asks “what does our assurance look like if we remove the human chain?” before the change is far better placed than one that asks afterwards. Answering it needs no technical expertise. It needs clarity about where accountability sits, and the discipline to insist that the answer can be audited.
Read alongside The Risk Detection Gap: AI, Restructuring and Board Accountability, which looks at the early warning function that middle management performs, and what happens when it is removed.
This article was researched and drafted with the assistance of AI tools. All claims have been verified, all sources checked, and editorial judgement exercised throughout by the author.
Updated 23 September 2026: publication date of the Block essay corrected to 31 March 2026; the Digital Omnibus entry-into-force date (27 July 2026) added; the Brookings reference now reflects what that research actually says.