We have something like 100 million times more computing power on our desks than the best machines of the early 1970s, and we are not 100 million times more productive. That gap is one of the most interesting facts in economics right now, and it tells us something uncomfortable about many of our organisations. The bottleneck on what we produce is not the work. It is increasingly the decision that shapes the work, and we have spent a century putting that decision in the wrong place.
Economic growth has stayed remarkably flat while capability has exploded because output is capped by its weakest link, and in knowledge work the weakest link is no longer production. It is judgement: deciding what is worth doing, and doing the right thing in a situation only the role closest to it can fully see. Distributing the authority to make that decision to whoever holds the highest resolution on the problem, human or AI, with a signalling structure clean enough to still steer, meaningfully loosens the bottleneck. Self-organising practices has been pointing at this for decades, and the arrival of capable AI agents is what finally makes it impossible to ignore.
Is Your Organisation Ready for AI Agents?
Eight dimensions of role clarity, distributed authority, living governance, and accessible data. Find out in a few minutes.
The short answer: output is gated by its slowest essential step, and compute only ever sped up some of the steps. The economist Charles I. Jones, in his 2026 Stanford essay AI and Our Economic Future, puts the puzzle precisely. We each have access to roughly 100 million times more transistors than people had in the 1970s, yet we are nowhere near 100 million times more productive. His own explanation is the heart of this whole piece: computers invert matrices at lightning speed, but a human still has to decide what matrix to invert, what hypothesis to test, and what the model should even be.
US real GDP per capita has grown at about 2% a year for 150 years. Not 2% on average with wild swings, but a strikingly stable trend through electrification, the computer, the internet, and now AI. If raw capability drove growth, that line would have bent upward long ago. It has not. Something keeps absorbing the gains.
That something has a name and mathematics. Jones models production as a chain of complementary tasks, and a chain is only as strong as its weakest link. When tasks genuinely depend on each other, infinitely improving the easy ones leaves you stuck against the hard one. Automate half of all economic tasks and, on his calibration, you raise output by less than a fifth. To merely double income per person, you would need to automate the overwhelming majority of what the economy does. The tasks you cannot automate become the cap.
This is not a new idea dressed up for the AI moment. It is one of the more durable results in development economics, and it is worth understanding properly before we apply it.
In 1993 the economist Michael Kremer published The O-Ring Theory of Economic Development, named after the single failed seal that tragically destroyed the Challenger shuttle and everyone aboard. Kremer modelled production as a set of complementary tasks where a mistake in any one of them can sharply reduce the value of the whole product. A brilliant meal with one rotten ingredient is not a slightly worse meal. It is not a meal at all.
The structural consequence is what matters here. When tasks are complements rather than substitutes, you cannot rescue a weak step by piling more effort onto a strong one. The theory helps explain why high-skill workers cluster together, why richer countries make more complex products, and why small quality differences translate into very large differences in output and wages. The value of the entire chain is held hostage by its worst-performed link.
Jones takes Kremer's structure and points it at our moment. The reason a hundred-million-fold jump in compute did not produce a hundred-million-fold jump in output is that compute only automated some links in the chain. The others, the ones humans still hold, became the binding constraint. And here is the move this article wants to make explicit, because the public conversation keeps skating over it.
When economists describe the weak link as a "task," it is easy to picture a manufacturing step or a line of code. But every task carries a decision inside it. The quality of a task is downstream of the quality of the judgement that shaped it: what to make, for whom, to what standard, and whether it should be done at all. Kremer and Jones model task quality. We are saying the decision (or governance) is part of that quality, and in knowledge work it is increasingly the largest part. The weak link is not the typing. It is knowing what to type and why. Make that explicit and a practical question follows immediately: who owns that decision, and do they have the authority and the resolution to get it right?
That reframing is not a stretch on the economics. It is the same claim with the decision layer surfaced rather than buried, so we can finally talk about who can be responsible for the quality of the link and give them the room to improve it.
There is a second, complementary piece of the puzzle, and it sharpens rather than softens the point. In their 2020 paper Are Ideas Getting Harder to Find?, Bloom, Jones, Van Reenen and Webb document that across industries and products, research effort is rising sharply while the productivity of that research is falling sharply. Their headline example is stark: sustaining the rate of progress described by Moore's Law now takes more than 18 times as many researchers as it did in the early 1970s. Aggregate research productivity, by their measure, falls at roughly 5% a year.
We are throwing exponentially more input at the problem to keep output growing at the same old rate, even correcting for the compounding effect. (The result is influential and widely accepted, though worth flagging that some critics question whether total factor productivity is a sound proxy for innovation at all.) Read alongside the weak-link model, it suggests the same diagnosis from a different angle. The constraint on a knowledge economy is not how fast we can execute. It is how well, and how quickly, we can decide and coordinate what to execute, and that has been getting harder, not easier, as organisations grow more complex.
So the technology is not the bottleneck. The way we organise the decisions around the technology is. If that sounds familiar, it should. We have been here before.
Because the technology is only half of the change. The other half is reorganising the work around it, and that takes far longer and is far harder.
In 1987 Robert Solow delivered the line that named the puzzle for a generation: you can see the computer age everywhere but in the productivity statistics. The historian Paul David answered it with electricity. Factories did not get more productive the moment they were wired up. They got more productive decades later, once they physically reorganised the entire factory floor around many small distributed electric motors instead of one giant central steam engine and its driveshafts. The dynamo was available long before the gains were, because the gains were waiting on the reorganisation.
Erik Brynjolfsson, Daniel Rock and Chad Syverson formalised this as the Productivity J-curve. General-purpose technologies like AI require large complementary investments in things like business-process redesign and new skills, and these intangible investments are badly captured in the official statistics. Early on, measured productivity sags because effort is going into building complements that nobody counts. Later, the gains arrive. The technology shows up immediately; the productivity shows up after the organisation has been rebuilt around it.
This is the honest counter-argument to my own thesis, and it deserves to be stated plainly: some of the "missing" productivity is mismeasurement and timing, not a coordination failure. Both can be true at once. But notice what the J-curve says the missing complement actually is. It is organisational reinvention. The lag is the time it takes to change how work and decisions are structured. That is precisely the layer this article is about.
The early evidence on AI fits the pattern, and it is more interesting than the hype on either side. Controlled studies show real gains in narrow settings: a 2025 study of customer-support agents found access to a generative-AI assistant raised issues resolved per hour by around 14%, with the largest gains going to novice workers and almost none to the already-skilled. But the most rigorous recent study of experienced developers found the opposite. When a team measured early-2025 AI tools on seasoned open-source developers working in their own mature codebases, the tools increased task completion time by 19%, even though the developers predicted a 24% speedup and believed afterwards that they had been sped up. The gap between felt speed and real speed is the tell. Generating output faster is not the constraint. The constraint sits somewhere else, in the judgement about what is actually worth doing and whether it fits.
With whoever has the highest-resolution knowledge of the situation the decision is about. This is not a management opinion. It is one of the most durable conclusions in the economics of organisation, and it runs through nearly a century of work.
Friedrich Hayek made the foundational argument in 1945 in The Use of Knowledge in Society. The central economic problem, he argued, is not allocating known resources but using knowledge that is dispersed, partial, and never available to any single mind in its totality, the knowledge of the particular circumstances of time and place. Central planners fail not because they are stupid, but because the knowledge they would need physically cannot be gathered into one place in time to be useful. The same logic applies inside a company as much as across an economy.
Jensen and Meckling turned this into an explicit theory of organisational structure in 1992. They split knowledge into two kinds: general knowledge that is cheap to move, like a price or a quantity, and specific knowledge that is expensive to move because it is local, tacit, and idiosyncratic. Their conclusion is that efficient decisions require co-locating decision rights with the specific knowledge relevant to the decision. Because you cannot cheaply move the knowledge to the decision-maker, you move the decision to whoever already holds the knowledge.
But, and this is the nuance the whole argument turns on, they showed that decentralising decision rights creates its own cost. The person with the local knowledge may not share the organisation's interests, so pushing authority down raises what they called agency costs even as it lowers the cost of bad information. The optimum is never total decentralisation. It is an interior point, held in place by a control system: clear measurement, clear feedback, clear alignment. Distributed authority works only when it is paired with a structure clean enough to keep it aligned. That single insight is the difference between a self-organising team and a chaotic one.
The AI-specific version of this argument is the most recent and, for our moment, the sharpest. In Power and Prediction (2022), Agrawal, Gans and Goldfarb observe that AI is fundamentally a drop in the cost of prediction, and that a decision is made of two things: prediction and judgement. For most of history these were fused in a single human mind; cheap machine prediction pulls them apart, so value and power shift to whoever is best placed to provide the judgement. Judgement is a statement of what we actually want, and it lives with whoever has the highest resolution on the situation the prediction describes. Prediction gets cheap and abundant. Judgement becomes the scarce, decisive input. The organisation that wins is the one that gets judgement to the edge, where the sensing happens, rather than hoarding it at the centre.
This is the strongest objection to everything I have said, and I want to take it seriously rather than wave it away. If AI gives the centre cheap, powerful prediction and a real-time view it never had before, why not re-centralise? Why not let a well-instrumented centre see everything and decide everything?
Three reasons. First, the same authors who proved AI's prediction power also showed that cheap prediction decouples prediction from judgement rather than concentrating both. You can centralise the prediction, run it at scale, pipe it everywhere, and that is genuinely useful. But the judgement, the part that did not get cheap, still belongs with whoever can feel whether a given outcome actually fits the situation in front of them. Centralised prediction is a tool inside distributed judgement, not a substitute for it.
Second, the thing AI cannot give the centre is the local, situated, often tacit knowledge Hayek described. A model can tell you the probability of churn across ten thousand customers. It cannot feel the pause before the complaint on this call, with this person, in this relationship. I have written elsewhere about why humans remain the irreplaceable sensors of an organisation, perceiving what no centralised view can reach. Better prediction raises the value of that situated judgement; it does not replace it. The honest reading is that AI's economics cut both ways, and the case for distribution holds specifically in knowledge work, where the binding constraint is judgement rather than a physical or regulatory limit.
Third, is context overwhelm, where the central AI has so much context that it starts to drift (most of us already have direct experience with this in our own threads), and does not have exact clarity on how to act. Avoiding this is called context engineering, and we have an article on that as well.
That last caveat matters. Sometimes the weak link genuinely is physical. Jones's own favourite example, self-driving cars, has been blocked for two decades less by bad decisions than by weather, edge cases, and sensor cost. Where the binding task is a hard physical or scientific problem, redistributing authority does nothing for it. The claim here is scoped to the work most of us actually do, where the slowest link is a decision.
It looks, structurally, like giving the role closest to a problem the authority to act on it, inside a rule- and signalling system that keeps the whole coordinated. The cleanest illustration is almost sixty years old.
On Toyota's production line, any worker who notices a defect can pull the andon cord. The cord signals an abnormality and brings help to the exact station; under the principle of jidoka, the person at the point of sensing has both the authority and the obligation to act. This is the opposite of the predict-and-control factory, where all discretion sits at the top and all motion at the bottom. The worker is the sensor, the worker has the authority, and the system is built around their perception rather than over the top of it.
The detail that makes it work is the part people miss. A modern Toyota line does not slam to a halt the instant anyone pulls the cord. The pull opens a fixed window for the team to resolve the issue, and only escalates to a full stop if it cannot be fixed in time. That is distributed authority bounded by a clean signalling protocol. Authority at the edge, coordination preserved. It is exactly the balance Jensen and Meckling's interior optimum predicts.
You can see the same pattern at organisational scale in the Dutch home-care provider Buurtzorg, one of the most independently studied case of distributed authority. Self-managing nurse teams, backed by a thin support structure and a transparent information platform, run their own decisions. An independent study found that Buurtzorg met patients' needs while using around 40% of the authorised care hours that conventional providers used, roughly 70%. Care got better and cheaper because the people with the highest resolution on each patient held the authority to decide.
I want to be careful here, because the evidence is honest in both directions, and a concept piece that only quotes the wins is not worth reading.
It breaks down when you distribute the authority but forget to build the signalling and alignment structure that keeps it coherent. The cautionary case is well known. When Zappos adopted a radical self-management model, the transition was painful: around 18% of employees took a buyout rather than work under the new system. Governance proposals piled up, meetings lengthened, and a company near fifteen hundred people ran into the practical limits of the approach. Distributing authority introduced its own coordination costs, exactly the agency-and-coordination tension the theory warned about.
The most credible assessment of this whole space comes from the Harvard Business Review, in a 2016 article deliberately titled to puncture the hype. Reviewing organisations that had adopted self-management, the authors reached a conclusion I would happily sign: these structures can make organisations more adaptable and nimble, but most companies should not adopt them wholesale, and a step-by-step approach usually makes more sense, using role-based work where adaptability matters most and conventional structure where reliability matters most.
That is the honest shape of the claim. Distributing authority to the highest-resolution role does not solve the growth puzzle, repeal the weak-link constraint, or eliminate coordination cost. It meaningfully loosens the bottleneck when, and only when, it is paired with a structure clean enough to keep the distributed decisions aligned and fast. Get the signalling wrong and you have simply traded a boss-shaped bottleneck for a meeting-shaped one.
This is also why the cost of getting it wrong can be so large. Gary Hamel and Michele Zanini estimate that excess bureaucracy costs the US economy more than $3 trillion in lost output, around 17% of GDP, with the average employee in large firms spending roughly 27% of their time on bureaucratic chores. Those are advocacy estimates built on assumptions, not precise measurements, so treat the exact figures with care. But the direction is hard to argue with. We pour an enormous share of our coordination capacity into routing decisions away from the people best placed to make them.
Pull the threads together and the picture is coherent. Growth is capped by the weakest link. In knowledge work, especially with AI, the weakest link is the decision that shapes the task, not the task itself. The economics of organisation say that decision should sit with whoever has the highest-resolution knowledge of the situation, paired with a control structure that keeps it aligned. AI use does not overturn this; it intensifies it, by making judgement the scarce input and pushing its value to the edge. And the practical evidence, from Toyota's andon cord to Buurtzorg's nurses, shows it can work, while Zappos shows how it fails when the signalling structure is missing.
None of this is a claim that this solves the 2% problem. It probably doesn't. But what we can do is stop making it worse. We can stop forcing every decision up a hierarchy that has lower resolution on the problem than the person who first noticed it. We can name the decision inside each task, give it an owner, and give that owner the authority and the signalling to improve it. That is not a small adjustment. In most organisations it is the largest available source of slack.
The deeper promise is that this compounds. Structured well, the questions a role has to escalate keep climbing the ladder of purpose as the structure learns, as a self-improving organisation. The low-resolution, how-do-I-do-this questions get absorbed into the roles and the working agreements, so that over time a role-filler, human or AI, is left mostly with the questions that actually need a wider view: does this still serve what we are here to do? I have watched this happen in agent-filled circles that, after a few cycles, stopped asking how to do the work and started asking whether the work served the purpose. The structure does not just relocate decisions once. It keeps freeing attention upward.
This is the same diagnosis I have made before from a different direction. The predict-and-control paradigm was built for a slow world of repetitive work, and AI is the mirror that makes its failure undeniable by running new tempo through old structure until the latency becomes the limiting factor. It is also why deploying agents without this structural layer fails so reliably: the absence of clear roles, authority, and signalling is the organisational readiness gap that derails most agentic AI projects. The bottleneck, again, is not the capability. It is where we put the decision.
Chad Jones and the economists named the puzzle and gave it a mathematics. The self-organising practitioners, like in Holacracy and Sociocracy, working in a different vocabulary, have been building the answer for decades, in Toyota's factories, in sociocratic and self-organising governance structures, in the slow unglamorous work of getting authority to the edge without losing coordination. The two have not often been put in the same room. This moment puts them there.
The invitation is not to flatten every hierarchy or fire every manager. It is more precise than that. Find the decisions in your organisation that are made far from where the relevant knowledge lives, and move them. Give the role with the highest resolution the authority to decide, and build the signalling structure, clear roles, clear domains, structured meetings, consent rather than consensus, that lets you trust that authority without losing the ability to steer.
At Nestr, that is what we are building: a working layer where every role, carbon-based or silicon-based, can sense a tension, hold the authority to act on it, and surface what they see into a structure that actually evolves in response. Not because distributing authority is a virtue in itself, but because it is how you move the weakest link, and the weakest link is where all the growth has been hiding.
The bottleneck was never the worker. It was where we put the decision. We finally have both the diagnosis and the tools to put it somewhere better.
Build the working layer for human and AI role-fillers
Nestr is where roles hold real authority, tensions get processed, and the structure of the organisation evolves with the work.
It is the idea that organisational output is capped not by how fast people or machines can produce work, but by how well and how quickly the right decisions get made and coordinated. When decisions are routed away from the people closest to the problem, that routing becomes the slowest, most error-prone step, and it limits everything downstream.
Because output is gated by its weakest link, and computing only automated some links in the production chain. The tasks that could not be automated, increasingly judgement and coordination, became the binding constraint. Charles I. Jones notes we have about 100 million times more compute than in the 1970s but are nowhere near 100 million times more productive, because humans still have to decide what to compute and why.
Introduced by Michael Kremer in 1993, it models production as a set of complementary tasks where a serious mistake in any one task sharply reduces the value of the whole. Because the tasks are complements, you cannot compensate for a weak step by improving a strong one, so the worst-performed link caps the value of the entire chain.
The evidence says it can, conditionally. Cases like Toyota's andon cord and the Dutch home-care provider Buurtzorg show real gains in speed, quality, and cost when authority sits with the people closest to the problem. But Harvard Business Review's research and the Zappos experience show it fails when organisations distribute authority without building the signalling and alignment structure that keeps decisions coordinated.
Not for judgement-heavy knowledge work. Cheap prediction decouples prediction from judgement: prediction can be centralised and scaled, but judgement, the statement of what we actually want, still belongs with whoever has the highest resolution on the specific situation. AI raises the value of situated judgement at the edge rather than replacing it.
.png){kind=logo}
