AI is not destroying predict-and-control management. AI is the mirror that makes its failure undeniable. The paradigm built for repetitive factory work, stable mass markets, and slow product cycles was already failing for the knowledge work most of us do. Agentic AI is just making the failure too obvious to keep ignoring.
That is the claim of this article, and the rest of it walks through why I think it holds up. The story is older and broader than the current AI conversation suggests. Predict-and-control was a deliberate invention. So is the alternative. Both have a history.
The interesting question is not whether AI breaks predict-and-control. Some of the most lucid people in management have been telling us it was already breaking for the better part of a century. The interesting question is what this moment asks of us now. What follows is an argument for sense and respond, drawing on practices like Holacracy and Sociocracy that have been working on the answer for decades.
The predict-and-control paradigm has older intellectual roots, but it was assembled into its modern operational form across roughly three decades, between Frederick Winslow Taylor's Principles of Scientific Management in 1911 and the Cold War strategic-planning architecture of the 1950s and 60s. It was built for repetitive manual labour, stable mass markets, and a tempo of change a multi-year plan could survive.
The conviction is older than the paradigm. Newton's Principia (1687) gave the West a universe that ran on knowable laws: enough information about position and momentum, and the future could be calculated. Descartes had set the philosophical groundwork half a century earlier. The Encyclopedists in eighteenth-century France made the catalogue of all knowledge a political project. Auguste Comte coined sociology in the 1830s on the wager that society itself could be studied as a system of laws. By the time the Industrial Revolution provided the factories, the railways, and the labour pools that needed organising, the Enlightenment had already supplied the conviction that organising them as a machine was possible. The twentieth century did not invent predict-and-control. It found the material conditions, the scale, and the tools to apply it.
The first great application was industrial capitalism. In 1911, Frederick Winslow Taylor published The Principles of Scientific Management. He was unambiguous about what he was proposing. "In the past the man has been first; in the future the system must be first." The job of the planner was to think, the job of the worker was to execute. All discretion at the top, all motion at the bottom. The book sold over two million copies in his lifetime and changed how factories ran across the industrialised world.
Two years later, Henry Ford opened the moving assembly line at Highland Park. Chassis assembly time fell from twelve hours to one hour and thirty-three minutes inside six months. By the start of 1914 Ford was paying five dollars a day, twice the going rate, partly to keep workers willing to do work that had been engineered to require almost no thinking. The line itself replaced human judgement. Cadence and sequence, not discretion, ran the floor.
Around the same time, the German sociologist Max Weber was finishing the work that would be published posthumously as Economy and Society. He saw bureaucracy as the most rational form of organisation yet invented, and warned that the same rationality would harden into what his English translator called the iron cage. Predictable, calculable, impersonal. Effective in its terms. A shell as hard as steel around the people inside it.
Three threads, the same logic. Decompose the work, predict the optimal way to do it, control its execution. By the 1950s and 60s the logic had absorbed the Cold War's planning culture, the RAND Corporation's strategic-planning methods, and Igor Ansoff's Corporate Strategy (1965). Alfred Chandler's The Visible Hand (1977) gave the whole system its narrative: managerial coordination had replaced market coordination as the engine of the modern economy.
It would be a mistake to read this as only a capitalist story. The same Enlightenment conviction that shaped Taylor and Ford also shaped the other great twentieth-century application of predict-and-control. Lenin was an explicit admirer of Taylor and pushed scientific management into the early Soviet economy as a matter of policy. The Five-Year Plans that began in 1928 were predict-and-control at national scale, with Gosplan as the central nervous system. The Cold War intensified the paradigm on both sides, RAND and PPBS in the West mirroring Gosplan in the East. Industrial capitalism and state socialism disagreed about who should do the predicting and the controlling. They did not disagree that the world could be.
It is worth being honest about the foundations. Taylor's most-cited story, the Bethlehem pig-iron worker called Schmidt, was partly invented for his Congressional testimony. The Hawthorne studies, the empirical pillar that the human-relations correction was supposed to rest on, turned out under reanalysis to show no Hawthorne effect at all. The paradigm took hold not because the evidence was airtight, but because the conditions fitted. Repetitive manual labour. Stable mass markets. Slow product cycles. Abundant unskilled labour. Cold War-era confidence that the future could be planned. In that world, predict-and-control made sense.
Predict-and-control worked when the work was repetitive, the market was stable, and the tempo of change was slow enough that a multi-year plan could outlive the conditions that produced it. In that world, centralising prediction and pushing instructions down the hierarchy was rational. In ours, it is not.
It is easy to be critical about Taylorism in retrospect. I want to resist that. The paradigm fitted a world. For the work it was built for, it produced extraordinary results on both sides of the Cold War divide. American manufacturing in the 1920s produced more goods per worker than any economy in history. Postwar mass production made consumer durables affordable across an entire middle class. The Apollo programme planned its way to the moon inside a decade. Soviet industrialisation, for all the human cost of its execution, took an agrarian economy to space within a generation. None of that was magic. It was predict-and-control applied to problems that genuinely yielded to it.
The work that fits has a clear shape. The task is repetitive. The variation in inputs is small. The downstream consequences of any one decision are limited and predictable. The cost of analysing the optimal way to do something is small relative to the value of doing it that way many times over. The rate of change in the conditions is slow enough that the analysis still applies by the time the work is done. Assembly lines, bulk logistics, standardised manufacturing. Within those boundaries predict-and-control is not an ideology. It is the right tool.
What happened next was not that predict-and-control failed inside its proper boundaries. It is that the boundaries kept being redrawn around it. Knowledge work arrived. The work changed. The organisations did not. By the 1980s the typical office worker was doing something that had almost nothing in common with the Bethlehem pig-iron worker, and yet was being managed as if she did. The Soviet system, asked to plan an economy whose information requirements outran any central planner's capacity to gather and process them, ran into the same wall from the other direction and failed first. In the West the failure was slower and quieter, hidden inside elaborate plans, tighter reporting cycles, and more frequent reorganisations. The instinct stayed the same: predict what the work should look like, then push that prediction down through the hierarchy as instructions.
This is the trap. The trap is not that predict-and-control was wrong. It is that we kept applying it after the work it was built for had been replaced by work it could not govern. By the time AI arrived, the gap between what the paradigm assumes and what the work actually is had been widening for fifty years, faster and faster.
The countermovement is older and broader than usually credited. Mary Parker Follett, Peter Drucker, Taiichi Ohno, W. Edwards Deming, Stafford Beer, John Boyd, the Agile Manifesto, Dave Snowden, Tom Thomison, Brian Robertson, and Frédéric Laloux have spent a century saying the same thing in different vocabularies: when the work changes, the management has to change with it.
The cracks were noticed early, by people who do not always get cited together but should. Mary Parker Follett, writing in 1926, made what she called the law of the situation the basis of effective coordination. "One person should not give orders to another person, but both should agree to take their orders from the situation." That single sentence undermines the entire structure of predict-and-control authority.
Peter Drucker named the change in conditions. He coined the phrase "knowledge worker" in 1959 and spent the rest of his career on its implications. His line in The Effective Executive (1966) is one most organisations still have not fully absorbed: “The knowledge worker cannot be supervised closely or in detail; she can only be helped. Once the knowledge is in the worker's head, the work cannot be specified in advance.”
Taiichi Ohno built the same insight into the Toyota Production System through the 1960s and 70s. The andon cord at Kamigo, installed in 1966, gave any worker on the line the authority to stop production when something was wrong. That is not a tweak. It is a structural rejection of Taylor. The worker is the sensor, the worker has the authority to act, the system is built around the worker's perception rather than over the top of it. W. Edwards Deming put the point into corporate language in Out of the Crisis (1986).
The Agile Manifesto, signed at Snowbird, Utah, in February 2001, condensed the change into four contrasts. One of the cleanest statements of this new paradigm is “responding to change over following a plan”. Software, by then, was the visible edge of what knowledge work would look like at scale. The signers had been watching predict-and-control fail in their own field for two decades.
Other lineages were doing the same work in parallel. Stafford Beer's Viable System Model and Project Cybersyn (1971-73, in Salvador Allende's Chile) had already shown what an organisation that genuinely sensed and responded could look like at national scale. John Boyd's OODA loop, properly read, makes the orientation phase dominant: continuous reorientation as new information arrives. Dave Snowden's Cynefin framework, formalised in Harvard Business Review in 2007, named the underlying confusion: the simplifications that work in ordered circumstances fail when circumstances change.
By the early 2000s the structural attempts were arriving. Tom Thomison and Brian Robertson's Holacracy (codified 2007, public 2009), the sociocracy lineage stretching back to Gerard Endenburg in the Netherlands, Frédéric Laloux's Reinventing Organizations (2014). Each of them tried to operationalise sense and respond at the level of structure. Distributed authority, roles instead of job titles, integrative decision making instead of either consensus or top-down command. The practice I will describe later in this piece is based on Holacracy, applied to a circle (team) that contains both human and AI role-fillers.
In a sense and respond paradigm, an organisation stays continuously coupled to its environment instead of trying to anticipate it. Stephan Haeckel coined the phrase in 1992. Brian Robertson described it as steering a bicycle by adapting continuously, eyes open, rather than pointing in a direction and pedalling with eyes shut.
Stephan Haeckel introduced the phrase in 1992 in Management Review, then developed it into Adaptive Enterprise (Harvard Business School Press, 1999). His clearest illustration is a transport one. The build-and-sell organisation is a bus line: static routes, fixed timetable, designed around predicted demand. The sense-and-respond organisation is a taxi service: dispatched in response to actual requests, route shaped by where the passenger needs to go.
Brian Robertson sharpened the contrast in his 2009 InnovationManagement interview. The paragraph is worth quoting in full because both terms appear in the same breath:
Most modern leadership and management techniques are based on a predict-and-control paradigm. In today's environment, steering an organization with predict-and-control methods is akin to riding a bicycle by pointing in the right direction, then holding the handlebars rigid and pedaling, eyes closed. Organizations need more dynamic methods for steering their work, to gradually shift from predict-and-control, to experiment-and-adapt, and finally to true sense-and-respond. Like riding a bicycle, dynamic steering involves pursuing a general aim by adapting continuously in light of real data about present reality.
That image is the most useful one I know for what we are talking about. You still know roughly where you want to go. You still pedal. But the handlebars are not locked. Your eyes are open. The data you are responding to is not a forecast. It is the ground in front of you.
Frédéric Laloux gives a clean one-line statement in Reinventing Organizations: "Trying to predict and control the future is futile. Everything will unfold with more grace if we stop trying to control and instead choose to simply sense and respond."
Structurally, what this asks of an organisation is specific. Authority distributed into roles defined around purpose, not job titles. Tensions raised by any role-filler treated as the primary signal that something needs to change, surfaced in regular tactical meetings. Integrative decision making, the consent-based process pioneered by sociocracy and refined by Holacracy, as the response loop. Governance evolving continuously, through the governance meeting, based on the actual work being done rather than big reorganisations.
AI is not the cause of the breakdown. It is the speed at which the breakdown now becomes visible. Predict-and-control's latency was always organisational debt. We could (somewhat) absorb it when work moved at human tempo. We cannot when agents ship work in hours, in parallel, across roles.
There are two ways to see what AI is doing to predict-and-control. They are both true. Together they explain why the breakdown is no longer deniable.
The first is empirical. The numbers from the past two years are, if you put them next to each other, hard to read as anything other than one symptom of one underlying problem. Bain reports that 88% of large business transformations fail to achieve their original ambition. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027. MIT NANDA's GenAI Divide (July 2025) found that 95% of generative AI pilots produce no measurable P&L impact. Harvard Business Review's December 2025 survey found that just 6% of companies fully trust AI agents to handle core business processes. McKinsey's State of AI 2025 puts enterprise-level EBIT impact at 39%, with only 23% of organisations scaling AI agents in any function.
These are not separate failures. They share a shape. Organisations are buying capability at industrial-revolution scale and then asking that capability to thread through approval cycles, planning rituals, and reporting structures designed for the kind of work that does not need any of those things. The technology is not the bottleneck. The organisation is. I have written about this elsewhere as the organisational readiness gap.
The second way of seeing it is structural, and harder to escape. Predict-and-control was always slow. The slowness was a feature, not a bug, in a slow world. The latency between sensing something in the environment, kicking it up the hierarchy, having it analysed, having a decision made, having that decision communicated back down, and having the work change in response was (hardly) tolerable when the environment moved at the pace of human attention. It was the price of central coordination, and the world paid it.
Agentic work breaks the bargain. A circle of agents can produce more drafts in a day than a person can read in a month. They can identify ten parallel research opportunities in a few hours. They can respond to an industry shift overnight. None of that fits inside a management cadence measured in fortnightly reviews and quarterly plans. The latency that used to be tolerable becomes the limiting factor. Run agentic work through a predict-and-control loop and the loop becomes the bottleneck, not the safety net it was designed to be.
This is what the failure-rate numbers are picking up. They are not measuring the limits of the technology. They are measuring the cost of running new tempo through old structure. The slowness was a feature in a slow world. It is a fault line in a fast one. AI is the mirror because for many of us it is the first thing fast enough to make the fault line visible.
AI alignment is the predict-and-control problem at its most extreme. Specify an objective, optimise for execution, and watch what happens.
The thought experiment that named this problem is Nick Bostrom's classic paperclip maximizer, first sketched in 2003. Imagine an AI given one objective: maximise the production of paperclips. Make it powerful enough, and it reasons its way through every available means. Convert factories. Convert warehouses. Convert the planet. The catastrophe is not that the AI rebels. The catastrophe is that the AI is perfectly obedient, and the unintended consequences of that obedience are total. It does exactly what was specified. The standard reading is a story about future superintelligence. The deeper reading is a story about predict-and-control taken to its logical end.
There is a name for the underlying mechanism. When a measure becomes a target, it ceases to be a good measure. The economist Charles Goodhart formulated this in 1975 in monetary policy, and it turns out to apply across an extraordinary range of systems. Educational outcomes once exam scores became the target. Healthcare quality once wait times became the target. AI behaviour once a reward function became the target. The most articulate response to this is Stuart Russell's Human Compatible (2019). Russell argues that the standard model of AI, machines that pursue fixed objectives, is fundamentally misconceived. He calls it the King Midas problem: getting precisely what you asked for, not what you wanted. His proposed alternative is sense and respond at the level of machine architecture. Agents should be deeply uncertain about the objective, learn from the humans they serve, and accept correction when they have it wrong.
There is another reason predict-and-control struggles with AI, and it goes the other direction. The systems themselves are not predictable. Modern large language models are not programmed in the way traditional software is. Capabilities appear at scale that no one designed in. MIT Technology Review named mechanistic interpretability one of its ten breakthrough technologies of 2026 precisely because nobody knows exactly how large language models work, and the field is racing to catch up with systems already in production. We are deploying executors whose capabilities, including emergent behaviours we did not intend or anticipate, exceed our ability to inspect them. Predict-and-control assumes the executor is the predictable element in the system. With AI agents, the executor is the part we understand least.
Both points lead to the same conclusion. Predict-and-control needs a knowable objective and a knowable execution. Agentic AI gives us neither. The objective, fully specified, generates unintended consequences a powerful optimiser will find. The executor, fully deployed, exhibits behaviours we cannot explain. What survives this is not control. What survives is structure that lets the agent stay coupled to the situation it is actually in. A role nested in a circle (team) nested in an organisation nested in a purpose. Structured meetings as the channel for surfacing what the specification did not anticipate. Consent rather than consensus, because consensus assumes we already agree about something none of us can fully see. Alignment, at the architectural level Russell describes and at the organisational level we are building, is the same instinct: stop trying to specify the future, and build the conditions under which the future can be sensed and responded to as it arrives.
Better prediction is real. Forecasting at scale is now genuinely better than humans for many domains. Yet the same authors who proved this, Agrawal, Gans, and Goldfarb in Power and Prediction, also showed that cheap prediction raises the value of distributed judgement, not centralised control.
There is a deep counter-argument that I want to take seriously. AI is genuinely good at prediction. So should we not just be doing predict-and-control better? More forecasting, sharper models, better scenario planning, faster feedback inside the same management paradigm we already had?
The empirical case for AI prediction is real. DeepMind's GraphCast, published in Science in 2023, outperformed the world's most accurate operational deterministic weather models on 90% of 1,380 verification targets. The European Centre for Medium-Range Weather Forecasts moved its AI Forecasting System into operational use on 25 February 2025 and AIFS-ENS on 1 July 2025, with up to 20% better tropical-cyclone tracks. Nothing about that is hype. The forecasting machines are better.
The reframe is in the same source as the strongest version of the counter-argument. Ajay Agrawal, Joshua Gans, and Avi Goldfarb's Power and Prediction (Harvard Business Review Press, 2022) is the cleanest economic argument for what AI is doing to organisations. Their core claim: AI does not automate decisions, it cheapens prediction. Decisions are made up of prediction plus judgement. As prediction gets cheap, the value migrates to the part that did not get cheap, which is judgement.
Their second claim follows from the first. Cheap prediction tends to decouple the prediction from the judgement, which in turn disrupts the existing power structure that bundled them together. The structure that bundled them together was the management hierarchy. Centralised prediction at the top, judgement legitimised by position, decisions cascaded down. Cheap prediction breaks the bundle. The judgement, the part that still matters, sits with whoever has the highest resolution on the situation the prediction is about.
That is the same conclusion sense and respond has been pushing toward for a century. Prediction is welcome. Forecasting tools are welcome. AI inside the sensing apparatus of an organisation is genuinely useful. What does not survive is the design choice that put a small group of people at the top of a hierarchy on the basis that they could predict best. The technology that was supposed to vindicate that design has, instead, cheapened the very thing the design assumed was scarce. Better prediction is a tool inside sense and respond. It is not a justification for keeping centralised predict-and-control alive.
The invitation at this moment is not to throw out predictions. We will continue to forecast, plan, and model. Those are useful things to do. The invitation is to stop designing organisations around the assumption that the future can be fully planned for. Build for the world that is actually in front of us. A world where the work is novel more often than it is repetitive. Where the environment shifts faster than a planning cycle can absorb. Where the role-fillers inside the organisation, human and silicon both, see things on the organisation's behalf that the centre could never see in time. If anything, plan for constant change and adaptation.
This is harder than predict-and-control, and not because it is more complex. It is harder because it asks us to give up the fantasy of central authority over outcomes. To trust roles instead of titles. To trust tensions raised at the edge over assumptions held in the centre. To trust consent rather than the comforting illusion that we agreed about something none of us actually examined. To make the structure of the organisation itself something that can be changed by anyone with a reasoned tension, rather than only by whoever has historically held the room.
Methodologies like Holacracy has been doing this for fifteen years. Sociocracy for considerably longer. The Toyota Production System pre-empted most of it. The Agile Manifesto made it visible to a generation of software teams. Stafford Beer designed the architecture for an entire economy. And now the work some of us do is moving fast enough, and is parallel enough, that the question is no longer whether predict-and-control fits. The question is what we put in its place.
At Nestr, that is what we are building. A working layer where every role, carbon-based or silicon-based, can sense, surface, and contribute. Where the structure of the organisation evolves with the work, not the other way around. Where the bicycle is the metaphor and the handlebars are not locked.
Join the wave.
.png){kind=logo}
