Last Thursday evening, at the official launch of Digital Directors Australia in Perth, someone asked me a question I have been thinking about since. We had been talking about AI governance in the boardroom, the gap between the speed of adoption and the readiness of oversight, when one of the guests turned to me and said: "But at some point, don't we just have to trust the system?"
It was a real question, asked without provocation. And I found I could not answer it briefly.
Trust in a system means something different depending on whether the system carries human values at all. Whether there are people in it, and on it. Whether anyone in the room, when the world model makes its recommendation, is positioned to ask what the model is not seeing.
The question has become more pressing since March 31, when Jack Dorsey (founder and former CEO of Twitter) and Sequoia partner Roelof Botha published an essay titled "From Hierarchy to Intelligence."
The Thesis, Taken Seriously
"From Hierarchy to Intelligence" opens with the Roman Army. Two thousand years before the first corporate org chart, the Romans solved the problem every large organisation still faces: how do you coordinate thousands of people across vast distances when communication is slow? Their answer was a nested hierarchy built around one consistent principle. A leader can effectively manage between three and eight people. Build that span of control into every layer, give each layer a clear reporting line, and information moves through the structure. Slowly, but reliably.
Dorsey and Botha trace this logic through the Prussian General Staff (created after Napoleon destroyed the Prussian forces at Jena in 1806), the American railroads of the 1840s, Frederick Taylor's scientific management, and into the modern corporation. Their argument is that hierarchy was always a workaround: a technical solution to a coordination problem that no other available technology could solve. The hierarchy exists to route information. Full stop.
AI, they argue, is that technology. For the first time, a system can maintain a continuously updated model of an entire business, routing information, pre-computing decisions, and maintaining alignment across thousands of people without a human intermediary at every layer. Block is restructuring accordingly. Three roles remain: individual contributors who build; directly responsible individuals who own specific outcomes on 90-day cycles; player-coaches who stay hands-on while developing people. The rest of what management does, the "world model" will do instead.
I take this seriously. The historical grounding is solid; the diagnosis of hierarchy's limitations is accurate. But the thesis rests on one assumption I want to name before testing it: that the intelligence of an organisation is separable from the people who constitute it. That you can capture its coordination functions in a data architecture and route decisions through a machine without losing anything essential. That is the assumption I do not share, and it is what matters most in the Australian context.
Technology or Capability? A Distinction That Changes the Thesis
In February, at the WA CEO Institute summit in Perth, Voleno founder Matt Mueller offered a distinction I wrote about in this series: AI is a capability, not a technology. The framing matters more than it first appears.
When we treat AI as a technology, we focus on deployment and efficiency gains. When we recognise it as a capability, we confront a harder question: what kind of organisation are we becoming? French neuroscientists Gaëtan de Lavilléon and Marie Lacroix, whose research I cited in that February article, found that while companies capture AI's productivity benefits, the social capital built over decades erodes progressively beneath the surface. The brain adapts to its tools. Prefer the machine long enough, and an organisation risks losing the competence for collaboration itself.
The distinction carries real weight when applied to the Dorsey/Botha thesis. If AI is a technology, it performs the coordination function that hierarchy previously required humans to perform. Replace the mechanism; the function continues. Clean logic.
If AI is a capability, the picture changes. A capability is embedded in the people and institutions that exercise it; it cannot be separated from the judgment of whoever deploys it. The "world model" encodes the assumptions and blind spots of the people who built it, defined what it measures, and decided what counts as signal; it is not the neutral coordination mechanism its name implies. Those choices do not disappear when the hierarchy does. They migrate to a smaller, less visible group of engineers and product leads. The coordination function is not eliminated; it is relocated and made less transparent.
The capability framing also introduces a distribution question Dorsey and Botha do not address. A genuine world model of the kind Block is building requires proprietary transaction data at scale and a remote-first organisational environment where all work is already machine-readable. Most companies do not have those conditions. Which means the efficiency gains from eliminating the coordination layer accrue to organisations already at the frontier. This is a reconcentration of intelligence, not its democratisation. And the workforce implications follow directly: who captures the value of the transition, and who absorbs its cost, is not a technical question.
Australia at the Inflection Point
The essay was published on March 31. By February, before it appeared, Australian companies were already acting on its logic.
WiseTech Global, the Australian logistics software company whose CargoWise platform handles a substantial share of global customs transaction data, announced the elimination of approximately 2,000 jobs: nearly a third of its global workforce across 40 countries. CEO Zubin Appoo described it as a "deep AI transformation" toward a "leaner, more efficient AI-led organisation." Teams in product, development, and customer service face reductions of up to 50 percent. Commonwealth Bank confirmed technology role cuts on the same day.
The share market's response was predictable. WiseTech closed up 11 percent on the day. The structural logic of that enthusiasm is straightforward: fewer employees, flat or growing revenue, better margins. What is absent from that calculus is harder to price.
KPMG's most recent survey of Australian business leaders names new technologies, including AI, and the use cases and ethics that arise when implementing them as the number one challenge in 2026, cited by 63 percent of respondents. The sequence in that phrase is worth noting: technology, yes, but ethics folded in immediately behind it. Deloitte's 2026 State of AI in the Enterprise report adds the scale of the gap: only 12 percent of Australian organisations say generative AI is already transforming their business, against a global average of 25 percent. Structural changes are arriving ahead of organisational readiness. That gap is where the governance risk accumulates.
What the Algorithm Cannot Carry
Here is the assumption named at the close of Section 1, now tested against practice.
The correct diagnosis is that hierarchy was always an information routing protocol, and that AI can now perform that routing function with greater speed and completeness than any management layer could. That part holds. What follows from it, however, is that the primary value of people in an organisation is routing information. If that is true, the logic follows cleanly: replace the human routers with the world model and push people to the edges to exercise judgment.
But people in organisations are not only routing information. They are also generating it. Consider what a project manager knows about a client relationship that has never appeared in a status report, or the judgment a middle manager exercises when meeting a 90-day performance target and supporting a person who is struggling pull in opposite directions. Neither exists in any artifact record the world model would consume. Remove the layer, and along with the bottleneck you remove some of the signal the model depends on.
In my own experience at the intersection of technology and labour relations, the same people who appear to be underperforming in a broken structure often become strategic thinkers in a well-designed one. The intelligence does not disappear; the structure traps it.
The WiseTech restructure has illustrated the practical stakes. More than 590 employees, reportedly a majority of the company's Australian technical workforce, signed a petition to CEO Zubin Appoo calling for fair redundancy packages, transparency, and genuine consultation. The company says it has sought to engage "in a structured and phased manner." That phrase reads differently depending on which side of the process you are on.
APRA's (Australian Prudential Regulation Authority) April 2026 letter to all regulated entities documented findings from a targeted engagement across large banks, insurers, and superannuation trustees. What it found was a governance problem. Many boards are still developing the technical literacy required to challenge AI-related risks with any real force; they are accepting vendor briefings at face value. A formal letter from a prudential regulator carries the force of notice, not suggestion.
I am conscious of the French comparison here, because I have seen it from the inside. The Comité Social et Économique (National Works Council) carries mandatory consultation rights before any significant organisational change: a structural obligation that forces an employer to articulate the human impact of a restructure before it is finalised, not as an afterthought. Australia's Fair Work framework operates differently; but the obligations that attach to large-scale redundancies still apply when the stated rationale is "AI-led transformation" rather than "cost reduction." The label does not retire the obligation.
There is also a problem the essay does not surface. A model of a company is, by definition, a model: a set of choices about what to measure, what to weight, and what to treat as signal. Those choices encode assumptions about value, about which work is visible and which is not, and whose contribution actually registers in a machine-readable record. A board that cannot interrogate those assumptions has moved from governing to ratifying. And as Dorsey and Botha are careful to point out, the compounding advantage of speed means ratifying quickly.
Leadership in the Intelligence Era
Which raises the question of what leadership in this transition actually looks like, and whether we are building the governance architecture to support it.
Digital Directors Australia launched in Perth last week, dedicated to improving digital literacy and governance capability in Australian boardrooms, with a particular focus on directors who have come up through non-digital backgrounds. In 2024, only 7 to 8 percent of ASX300 board directors had a technical background. Gartner projected that 20 percent of enterprises would use AI to flatten their hierarchies by 2026. A structural transformation is moving through organisations whose governing bodies, by and large, do not yet have the vocabulary to interrogate its assumptions.
But the leadership question runs deeper than technical literacy in the boardroom. The Dorsey/Botha model implies a specific style: high individual accountability, short 90-day cycles, no buffer roles, player-coaches who remain technically hands-on. Speed is the compounding advantage. This works for clear missions and measurable outcomes. What it does not account for is any of the functions that become more important precisely when information routing is automated.
When the coordination infrastructure moves to the machine, sensemaking rises in value. The world model optimises within a defined problem space; it cannot reframe the problem. The leader who asks whether we are solving the right problem, who can read ambiguous signals without reducing them to a metric, provides something the algorithm cannot. This is not the "gut feel" of the old leadership mythology; it is pattern recognition shaped by years of contextual exposure the model has not been trained on.
Ethical holding rises with it. When decisions are made at machine speed, the pause between "we can do this" and "we should do this" becomes structurally scarce. The leader who can hold that boundary is not performing a compliance function; they are performing the most distinctly human function in the building.
Relational credibility follows a different logic but matters just as much. Trust is built through consistency and shared experience over time; none of this is machine-readable. The leader who can maintain the human fabric of an organisation through rapid restructuring, who people follow not because of positional authority but because of relational history, has a capacity the world model cannot replicate. Perth's professional culture, as I wrote in February, is still relationship-intensive. Deals close because of trust built over years. That is an asset, and it is under pressure.
Beyond these, what I would call governance intelligence becomes the defining quality. The discipline to ask: what is this model not seeing? Whose signal is absent from its training data? What happens to accountability when a decision is attributed to the algorithm rather than the person who chose to use it? These are the questions that separate oversight from endorsement.
What does not survive: leadership whose value came from controlling information flows; authority derived from title or tenure rather than contribution. These are exactly the functions the world model absorbs, and the leaders most exposed are the ones whose position rested on them.
The reinvention, then, is less about inventing new leadership qualities than about reordering which ones matter. What was treated as soft and supplementary (sensemaking, ethical judgment, relational intelligence, the capacity to challenge a system's assumptions) becomes the primary contribution. What was treated as the core management function moves to the machine. That is a profound shift in what organisations need to hire for, develop, and protect.
The Transition We Owe Each Other
This is the question Dorsey and Botha do not ask, and it is the one that will determine whether the transition they describe is legitimate.
The efficiency gains from eliminating the coordination layer accrue somewhere. In the Block model and in the WiseTech restructure, they accrue upward in the organisation and outward to shareholders. The market priced the WiseTech announcement enthusiastically; it did not price the 590 employees who signed a petition asking for transparency and fair treatment.
The workforce question extends beyond job displacement to the career pathways and social mobility that management hierarchies (however imperfect) made possible. A graduate entering an organisation without a middle management layer has no visible path to develop judgment and earn the standing to lead. The world model can track performance metrics. It cannot replicate what happens when a person spends a decade learning how to navigate a complex human organisation.
Australia's own data suggests the picture is more mixed than either the optimistic or catastrophist forecasts. Jobs and Skills Australia found that generative AI has greater capacity to augment work than to automate it. PwC's AI Jobs Barometer shows wages rising faster in AI-exposed sectors. But those are averages. The organisations capturing the efficiency gains are not the same as the workers absorbing the transition costs, and averages do not describe the distribution.
The social licence question follows. An organisation that captures the benefits of AI-driven restructuring while externalising the human cost is borrowing against its community's goodwill. Australia's National AI Plan explicitly prioritises shared prosperity, and Privacy Act amendments coming into effect in December 2026 will require transparency on automated decision-making. The government has also convened its first tripartite meeting on AI regulation. The room to borrow is narrowing.
Humans need to be in the AI loop and on it: inside the decision process, contributing to what the system learns, and outside it with the authority and the vocabulary to challenge what the model recommends. "In the loop" and "on the loop" are not the same thing. The second is harder to sustain and more important to protect. It is what ensures the world model remains accountable to something beyond its own optimisation criteria.
I do not write this as a defence of the old hierarchy. The hierarchy was imperfect, frequently unjust, and built for an information environment that no longer exists. But the case for the transition to intelligence requires more than a proof of concept from one well-capitalised technology company in San Francisco. It requires an honest account of who bears the cost and what governance architecture ensures the benefits are shared rather than concentrated.
The Answer in the Room
The question asked at the Digital Directors launch deserves an honest answer: "At some point, don't we just have to trust the system?"
Trust in any system is earned through accountability and the capacity for correction. The hierarchy earned its trust, where it did, because there were named people behind its decisions: people who could be questioned and held responsible. The world model earns trust through a different mechanism, through the quality of its governance and the presence of people who have both the technical literacy and the ethical standing to challenge it when it is wrong.
The boardroom conversation about AI governance cannot wait for the technology to stabilise; it will not stabilise. It will compound. And in any organisation now facing this transition, the more important question is who decided what counts as signal, and whether anyone in the room had the standing to challenge it.
That capacity is human, and it is becoming the most important thing an organisation can have.