Ask most executives what AI transformation looks like, and they will still describe an AI assistant at every desk. Each employee receives a licence for an AI ‘copilot’, which drafts their emails, summarises their documents, reviews their spreadsheets and answers their questions.

This, to be sure, is a real benefit, and a company that gives its employees a copilot is better placed than one that affords no AI enablement. But such a company remains, inevitably, the very same company with slightly faster employees.

By ‘the copilot model of AI’, I mean any arrangement in which a person performs each task and an AI helps them to do it. My chief aim in this essay is to suggest to you that this model, however useful for now, rests on a mistaken understanding of what AI has become, and that a company which treats it as the destination will never capture the upside that AI makes possible. Whereas a copilot makes a company’s people faster, real automation changes what kind of company it can be.

A category error is the mistake of placing something in a category to which it does not belong. The copilot model makes this mistake about AI. It treats AI as a tool to be wielded by a human, when instead it is an agent in its own right and can now do valuable cognitive work by itself.

I anticipate the objection that much knowledge work has long been automated (take bookkeeping software, for example). This may be so, but only where the work is reducible to deterministic rules that can be written down in advance. Software automated the rules, but AI is the first technology that can take on what remains, namely handling exceptions, navigating abstract frameworks and exercising judgement. This is the difference between a tool and an agent.

Given a well-defined task and access to the company’s systems, the best current models can plan the work, carry it out, review it and improve it, much as a capable member of staff would. What they largely cannot do (in the current regulatory paradigm) is answer for the work, as they lack legal personality, and therefore bear no responsibility and carry no liability. This is why a human must still remain accountable for results. But owning the work is not the same as doing it. No manager would hire a new rockstar employee and then insist on doing every piece of work personally, yet this is, in effect, how most companies now use copilot-style AI.

The error is understandable, because until recently the models could do useful work but still made too many mistakes, and could not carry out a long, complex task to its end without someone standing over each step. For that generation of AI models, the copilot was indeed the right design. For a large and rapidly growing share of cognitive work, however, it is no longer the correct model.

It is at this juncture that I will put it to you that the existing, generally available models are already capable of doing vast quantities of existing knowledge work better than qualified humans. In a recent study of month-end accounting tasks, a frontier model working alone was about fifteen times faster, and more accurate, than licensed accountants using AI. Only eighteen months earlier, the best models had scored below the accountants’ average on the same tasks, and the models in the study have since been superseded by even more capable ones. One must realise that this is the worst that the models will ever be, but even if development stopped today, the automation of most knowledge work is already feasible using only current models.

The improvements in AI capabilities have occurred so fast (largely within the past year) that most people are still completely unaware of what is now possible. Moreover, organisational change is always slow, and habits always outlast the conditions that formed them. In established companies, the prevalence of a copilot model (to the extent that the company has gotten even that far) is reinforced by the fact that the copilot is the one model of AI adoption that leaves the company’s existing hierarchies, structures and workflows intact, which is to say that it is the path of least resistance. Many companies, having equipped their employees with AI assistants, believe that they have been transformed. This, then, is the copilot fallacy.

Consider as an example a professional firm in which twenty juniors do the client work under a handful of managers and partners. Give each junior a copilot, and the firm has twenty faster juniors. First drafts appear faster, documents are summarised and reviewed more quickly and correspondence with clients is written for them. This is a real efficiency gain.

But every piece of work still passes through a person’s hands, at a person’s pace and within a person’s working day. In the language of the AI field, the person is still ‘in the loop’, which is to say that no piece of work is done until a person has done or approved it, with work being done in turns by human and machine together. However much of a task the machine prepares, the person must still read it, decide on it and approve it, and that sets an inherent limit on how much faster the work can go. If a human’s own portion of a task irreducibly takes ten minutes, no assistant can make the task take less time. As a result, the firm’s productive capacity can grow in only one way, which is through hiring more humans.

Another inherent drawback of the copilot model is that this limitation does not relax as AI models improve. No matter how capable a copilot becomes, whatever part of a workflow remains the task assigned to a human does not gain from advances in model capability.

It is an idea that needs getting used to, but the fact is that where AI is competent to perform a task, a person in the loop adds little but unnecessary time, and even risks forcing error. This is to say nothing of the significant difference between the relative cost and speed of human employees and agentic systems. There is, in other words, a growing collection of tasks for which a human in the loop detracts value rather than adding it, and for which automation is an inevitability.

By ‘automation’ I do not mean simply taking the existing workflow and having a machine perform each of its steps. Elon Musk famously sets out a sequence for improving any process: question every requirement, delete every unnecessary step, simplify what remains, speed it up, and only then automate. The order matters because most existing workflows take their shape from the limits of the people who perform them. Work is batched into periodic cycles because people cannot attend to everything as it arrives, divided into stages and handed between people because no one person can do all of it, and reviewed in layers because people tire and make mistakes. A machine largely does not share these limits, so much of that structure has no purpose once the machine does the work. The question is therefore not how to perform each step faster, but which steps would exist at all in a workflow designed for the machine. Many of them, rather than being made faster, simply disappear.

True automation has features that place it in an entirely different category from the copilot model. I highlight some below which show why the two cannot be treated as faster and slower versions of the same concept.

The first is what factor limits output. In the copilot model, where work passes back and forth between human and machine, output is limited by the attention of the company’s people. Once the machine does the work, however, output is limited by computing power and by the capacity to verify that work, both of which can be scaled cheaply in a way that human attention cannot.

The second is time. A person attends to one task at a time, and only during working hours. A machine can carry out a hundred tasks at once, through the night and without tiring, so that work no longer waits in a queue for someone to attend to it.

The third is cost. In the copilot model, the cost of each additional task tracks a salary. With automation, it falls towards the cost of the computing and the verification that the task requires, which is typically orders of magnitude cheaper than human remuneration.

The fourth is where skill resides. In the copilot model, each person must learn separately how to get good work out of the machine, and many never will. Once the machine does the work, skills are built into the system once and apply to everything it does thereafter.

The fifth is continuous learning. Automation requires the company’s method to be made explicit, and an explicit method can be inspected and corrected. A correction made once applies to every later task of its kind, so the system improves with each error it is shown. A copilot’s help, by contrast, stays private to each user and leaves little behind by way of organisational learning.

I anticipate another objection: if a machine does the work, and a person must review every piece of work in full, little has been gained in total time saved. This, of course, is perfectly true, and it is why verification must itself be automated as far as possible.

It is instructive to consider the recent developments in software engineering practices. Many leading programmers no longer read every line of code that their AI systems write, because doing so would forfeit most of the efficiency gains. Instead, the code is run against thousands of automated tests and scanned meticulously for vulnerabilities by other AI systems during adversarial review, and a person reviews what those tests escalate. The same pattern will eventually extend to every domain in which work can be verified, which is virtually all knowledge work. Human verification then shifts from inspecting the technical minutiae of how the work was done to reviewing and validating outcomes.

Moreover, much of the verification can itself be designed into the operation of the automated system. One AI system can review the work of another, independently and at machine speed, so that first-pass review no longer waits on a human. Other tests can be built once for each type of task and applied to every instance of it, reconciling results against external sources of truth, comparing them with earlier and comparable cases, and testing them for internal consistency. The faster the review, the faster the company turns its work around, which makes the automation of verification among the most valuable things a company can implement.

These changes describe a different kind of company. By an AI-native company, I mean one built from the outset around the machine performing its work, with people placed where judgement, accountability and relationships are required.

Consider an audit partner. An audit partner does not redo the audit, but relies on a structure of review, in which managers review the work of seniors, who review that of juniors, and the partner makes the significant judgements personally and signs off. An AI-native audit firm keeps that structure but largely replaces the layers beneath the partner with machine work and machine review. The partner’s role barely changes, while everything beneath it does.

The economics of the two kinds of company differ accordingly. The traditional company’s leverage is the number of staff each manager can supervise, and its costs rise with its headcount. The AI-native company’s leverage is the number of judgement calls each of its people can make, and its costs rise with computing and verification. The first grows by hiring. The second can serve many times as many clients with the same people, at a fraction of the price, and deliver its work in hours rather than weeks.

It is no use to anyone trying to bury the implications of automation in platitudes and euphemisms. A task that is automated is no longer performed by humans, and most of the tasks that currently fill people’s working days will not survive genuine AI transformation. This is not a by-product of automation but its purpose. The copilot model currently offers the false comfort of pretending otherwise, but sticking to that model for technical work confines a company’s pace and output quality to human standards, which, in most cases, will be a losing game.

This is certainly not to say that companies will have no use for people. As the technical work moves increasingly to the machine, people move from within the loop to above it, where they set direction, decide what matters, make the calls that carry consequence, build relationships of trust with clients and answer for outcomes. Technical competence alone, once the scarce thing that many companies sold, no longer distinguishes them. To transcend the loop is to stand above it, rather than to abandon it completely. A company could, of course, use automation simply to cut costs and reduce its workforce. A visionary leader would instead leverage the people the company already has by placing them above the loop, so that together they serve a far larger client base than was possible before.

The practicalities of how a company should make that transition, and what it means for the professions in particular, I will consider in later essays on the practicalities of reinvention and on the reshaping of professional firms.

The copilot will remain tremendously useful for whatever work still requires a person to do it. But companies that mistake it for the holy grail of their ‘AI transformation’ will in time find themselves trying to compete with companies that leverage their people above the loop, rather than leaving them stuck inside it. This competition will not last long, for the odds are tremendously skewed in favour of the AI-native company. The question before every company is therefore not how much faster AI can make its people in their existing workflows, but what the company could become if its people transcended the loop entirely and abundant intelligence were unleashed. With people above it charting the course, the value-creating loop could be brought to turn at speeds which no copilot model could ever hope to match.

In ‘Transcending the Loop’, I argued that seeing the copilot model of AI as the desired end state is a fallacy, and that the transformative value of AI lies in automating tasks rather than in assisting people in doing it faster. This essay is intended to turn from why to how. There are many roads to reinvention, but the principles that should guide the journey are mostly universal.

Naturally, AI transformation is about putting a transformative technology to work properly. But, perhaps counterintuitively, the more important part of any real AI transformation is organisational redesign around the new technology, more so than the new technology itself. The technological solutions a company requires must be engineered, and engineered well, but the harder part is to redesign the company’s functions around a new reality that none of them was built for, namely that intelligence is now abundant and no longer a limiting factor.

Every business function acquired its current form under conditions in which intelligence was scarce. Attention was expensive, so work was batched and made to wait in queues. No one person could hold the entirety of a complex task in mind at once, so projects were divided into stages and handed from one person to the next. People tire and make mistakes, so their work was reviewed in layers. And because those doing the work could not see (nor had the time to see) what everyone else was doing, a layer of managers grew up to carry information up and instructions down. In ‘From Hierarchy to Intelligence’, for example, Jack Dorsey and Roelof Botha have argued that this routing of information is the principal reason corporate hierarchy exists, and that AI now makes much of it redundant.

None of the above arrangements represent some eternal aspect of organised work. Each was a rational solution to a constraint that is now busy lifting. Fundamentally, this is why AI transformation is as much a matter of organisational design as of technical engineering, and usually more so. A business function designed around scarcity of intelligence simply cannot capture the value of abundant intelligence just by having AI perform its existing steps. It has to be redesigned from the ground up.

This is not the first time such an argument has been made. In 1990, Michael Hammer urged companies to stop using computers to mechanise old ways of working and to redesign their processes from scratch, an approach that became known as business process re-engineering (BPR). Hammer was right that work should be organised around outcomes rather than tasks, and that much of the time a piece of work spends inside a company is spent waiting rather than being worked on. One insurer he cited estimated that an application spent twenty-two days in process but was actually worked on for only seventeen minutes. Re-engineering nonetheless became associated with mass lay-offs and failed programmes, and Hammer later conceded that he had been ‘insufficiently appreciative of the human dimension’.

The lessons from the BPR movement are worth keeping in mind. Redesign imposed in a single upheaval, aimed chiefly at cutting costs and indifferent to the people who do the work, always tends to fail. What is different now, however, is that AI can perform judgement-laden work itself. The redesign therefore reaches beyond the process alone to the roles, authority and verification around it, and a redesigned function can now keep improving as it learns from being corrected by its human overseers.

A company’s work can be divided and understood at several levels of analysis. At the first level of division are departments, such as ‘finance’ or ‘operations’, which are administrative containers of related functions. By a ‘function’, I mean a distinct but coherent piece of work the business must perform to achieve a specific end, such as onboarding a client, invoicing or procuring trading stock. Each function consists of one or more workflows, by which I mean processes aimed at transforming certain inputs into a desired output, and each workflow is carried out through individual tasks. A job, finally, is the bundle of tasks assigned to one person, sometimes drawn from several workflows and functions.

Of these levels, the correct conceptual unit of reinvention is the business function. The concept of a department is itself a product of scarcity. Because no one person could master more than a narrow field, people had to be organised into groups of specialists, so that accountants sat with accountants and marketers with marketers. AI, which draws on many specialisations at once, removes much of the reason for that grouping, and a function redesigned around it need not respect departmental lines. A job is too disparate a bundle of tasks to be the most sensible unit of automation, and automating a single task in isolation leaves the workflow around it as it was.

Nor is the workflow the right unit on its own: a function may consist of a single workflow, or of several that must be reinvented together, and reinventing the function in the face of AI sometimes makes some of the workflows that once formed part of it unnecessary altogether. The relevant distinction is one of means and ends. The function is concerned with the end the business seeks to achieve, and its workflows are merely the current means of achieving it. Reinvention therefore begins by identifying the end, and then reconsiders the means. The function, together with its constituent workflows, is the correct level at which a company’s work can be rebuilt systematically.

Reconsidering the means follows Musk’s sequence set out in my previous essay: one must question every requirement, delete what is unnecessary, simplify what remains, and only then automate. Of each, one should ask who required it and why. Was it the client, an internal check, a regulator, an insurer, or simply a developed habit? One would be surprised how many steps survive only because no one has thought to question their existence.

Reinventing the workflow in the face of AI therefore involves breaking it down into the units of work, or tasks, that must be performed to transform the inputs into the output. Each unit should be small enough to be verifiable on its own but not so small that the context needed to do it well is lost in passing it from one unit to the next. Each unit is then assigned to the mechanism best suited to it. Work that can be reduced to deterministic rules can be done by conventional software, which is inexpensive and predictable. Work that calls for reading, interpretation and judgement goes to AI. Work that carries serious consequence, or that the law requires a person to do, stays with a human, who receives it with the evidence required for verification already assembled.

Human review is placed where it does the most good. Ideally it comes at the end of a workflow, where a person approves the outcome and answers for it. But where a wrong turn early in a workflow would corrupt everything that follows, review may be required early, before the error spreads. The point at which one workflow feeds into another is also usually a sensible checkpoint for verification. In each case, the principle is that review belongs wherever an error is cheapest to catch, relative to the damage it would otherwise do.

None of this works unless the AI system knows what the company knows. A company’s method, its precedents, its standards and its house style exist mostly in the heads of its senior people. Making that knowledge explicit, in a form the machine can use, is among the most demanding parts of any reinvention, and among the most valuable, because it is what turns a general intelligence into the company’s own individuated solution. For the same reason, that knowledge should remain under the company’s control, held in a form that can move between AI providers rather than being locked into any one of them.

The greatest leverage lies in automating a company’s innermost loop, or core value-creating flywheel, which is to say the work its clients pay for. Consider a law firm. Automating client intake, timekeeping or invoicing is valuable, but is categorically different from automatic its ability to do quality legal work. An AI system that researches the authorities for an argument, drafts the particulars of claim or writes an opinion in the firm’s house style, with a lawyer steering it and reviewing the result, can collapse days of work into hours, and multiply the productive capacity of the firm. While there are still many skeptics of this proposition, I have seen it happen with my own eyes many times over, and would therefore advise you to suspend such skepticism as you may have.

This does not mean that every AI transformation must necessarily begin with the innermost loop. For a company still cautious about AI, the back office is a sensible training ground. A function such as onboarding or timekeeping carries comparatively little risk, its results are easy to measure, and success there builds the confidence and foundations that later work in the innermost loop will need. Beginning in the back office is far better than not beginning at all.

In deciding where to begin, a company should consider both which of its existing functions to reinvent and which functions that once lay outside it can now sensibly be brought inside. Much of what a company buys in from outside, it buys because the work once required specialists it could not afford to employ, and that, too, is an arrangement born of scarcity. A law firm, for example, might now bring much of its marketing in-house and design it around AI from the outset, although higher-stakes work may, for now, still go to specialists.

When functions are redesigned in this way, the organisation around them changes too. Which roles exist, and what each role consists of, are necessarily reshaped. Much of what a junior lawyer does today is performed by the machine, and what remains of the role tilts towards judgement, review and contact with clients. The routing layer of middle management, which existed to carry information between those who decide and those who do, largely falls away, although the work of developing people and holding them to account remains in different forms. Decision rights move closer to the people who answer for outcomes, and even the way work is priced may have to change, since a fee built on billable hours is logically inconsistent with work which takes little human time.

No company is served by pretending that none of this will affect its people. Those who do the work today hold much of the knowledge a redesign needs, and they will share it fully only if they are told candidly what is changing and where they are going. A company that redesigns its functions in order to serve more clients, rather than merely to cut costs, must make it clear to employees that their work is moving above the loop. What this means for the shape of professional firms in particular, I consider in a separate essay.

Reinvention tends to have consistent causes of failure.

The first is when reinvention is treated as a technology purchase. A company buys an off-the-shelf AI product, as it might once have bought accounting software, and expects transformation to follow. Or it starts from a tool and then looks for a use case, rather than starting from a function and asking how the function should work. Or it hands the transformation mandate to its IT department, whose expertise lies in keeping existing systems running rather than in organisationally redesigning how work is done. Existing products have their place, and integrating them is often wiser than building everything from scratch. But they must be fitted to a redesigned function, not the function to them. One must not confuse means with ends.

The second is when reinvention is treated as a single project. Pilots are launched and never scaled, or a single programme is run to an end date, as though reinvention could be done once and for all. Some companies postpone everything until their data has been cleaned up, an exercise that can absorb years without producing anything, even though modern AI systems are far more tolerant of imperfect data than software of old. Reinvention is better understood as a continuing, iterative loop of its own, in which each redesigned function is measured, corrected and improved, and in which the foundations laid for one function make the next one faster to rebuild.

The third is when it is left to chance, or a wait-and-see approach is followed. Bottom-up adoption, in which each employee experiments with whatever tools they choose, produces scattered gains and no organisational redesign, because no individual employee has the authority to rebuild a function that crosses several departments or even jobs. Reinvention needs an owner, someone with the authority to change how a function works from beginning to end. It only works if it comes from the top.

Could a company carry out such a reinvention by itself? In principle it could, if it had the right people and expertise, but in practice, very few companies do. Reinvention calls for two disciplines that are rarely found together. The first is organisational design, the industrial engineer’s understanding of how work flows through a company and how it ought to. The second is AI engineering, which is not the same as conventional software development and which most IT departments do not possess. Neither can succeed alone. Organisational design without engineering produces fanciful plans that are never built, and engineering without organisational design produces tools grafted onto workflows that should not exist.

This suggests three tests for evaluating who you should turn to for help. You should be skeptical of AI tools and platforms sold in the way software has always been sold, as products to be installed. You should be skeptical of developers who propose solutions before they have spent considerable time understanding the business. And you should be skeptical of consultants who can describe a transformation but lack the hardcore engineering ability to build it.

Reinvention has no final destination at which a company arrives and stops. Each function reinvented lays down knowledge, skills and infrastructure on which the next can be rebuilt more quickly, so that the roads of reinvention are laid as they are travelled. In time, the question of ends and means even reaches beyond the company’s constituent functions to the company as a whole. In the final analysis, its existing departments, hierarchies and jobs were also means, adopted when intelligence was scarce because they were the best way of achieving its overall ends. Reinvention that begins with a single function culminates, if pursued far enough, with the reinvention of the company itself, rebuilt to maximally realise the ends it exists to achieve, now liberated from the scarcity of intelligence.

Most public argument and uncertainty about artificial intelligence (AI) still concerns its future, and in particular whether the models will continue to improve and how quickly they will do so. For the purposes of this essay, however, I propose to set that argument aside and to assume that all progress in AI capability stops today. That assumption would, to be clear, be ridiculous, and I make it only for the sake of argument. My contention is that even on that assumption, the existing models are already capable of performing most of the tasks of which knowledge work on a computer consists. Disruption does not depend on some unproven hypothesis about what AI may one day become, but is already latent due to the capabilities of publicly available AI models today. And if disruption follows on the most conservative assumption about capability, it follows all the more on any realistic one.

In ‘Transcending the Loop’, I argued that the transformative value of AI lies in automating business functions rather than in helping people to perform parts of them faster, and in ‘The Roads of Reinvention’, I considered how a company might go about reinventing itself around that fact. This essay concerns the phenomenon that will make such reinvention necessary, namely competitive selection, and why I expect it to act sooner than the history of earlier technological revolutions would suggest.

My first premise concerns what the models can already do. Most knowledge work on a computer consists of discrete tasks, such as reading a document, extracting what matters, drafting, analysing, and checking the result against a rule or a standard. Each task takes text or data as its input and produces text or data as its output, which is precisely the medium in which the models work. On benchmarks of well-specified professional tasks, the leading models already beat or tie industry professionals in most comparisons, and anyone who uses the latest models for doing real work will understand mighty well what I am talking about.

What the models cannot yet do reliably is most professional work as it is presently organised. A benchmark task arrives packaged, in that someone has already framed the question, assembled the relevant material and stated the standard that a good answer must meet. Professional work arrives in a firm without that packaging. The client’s request is usually underspecified, the facts sit in emails, files and systems to which the model has no access, what distinguishes a good outcome from a bad one has not been specified in a form that the model can take into account, and the request is seldom a single task but a sequence of them, each building on the last.

Firms largely left their work ‘unpackaged’ for a good reason, which is that their people never needed the packaging. An experienced professional carries the context of a matter, knows where its documents are kept, asks a colleague when something is unclear and recognises when an answer is good enough. Because people supplied all of this implicitly, no firm has had cause to make it explicit, and every process in the firm was designed on the assumption that a person would perform each step. A model given the same context, access and standard can perform most of those steps. What it lacks is not the intelligence to use them but the packaging that would deliver them. The gap between performing tasks and performing professional work is therefore a failure of packaging rather than of capability, and packaging, being a matter of organisation, is precisely what a firm can supply.

It follows that the use of AI will lag behind its capability, and that the lag will be shortest for individuals and longest for firms. An individual can take up AI in an afternoon, because the only way of working that must change is the individual’s own. A firm cannot, because its work is distributed across roles, handovers and controls that were built around people. Where work is already divided into discrete tasks, as in freelance writing and coding, there is little to redesign, and substitution follows quickly. Elsewhere, the capability is present, and what remains outstanding is the redesign of the organisations within which most knowledge work is done. This, in my view, is why the disruption is inevitable, and already latent in the capability itself. Packaging is a problem that firms can solve, and it is only a matter of time before they do.

My second premise is historical. A general-purpose technology yields its value only once the processes and organisations around it have been redesigned, and redesign takes time. Electric power was commercially available from the early 1880s, yet as late as 1899 electric motors supplied less than 5% of the mechanical drive in American factories, and manufacturing productivity rose markedly only in the early 1920s. The economic historian Paul David attributed the delay in large part to ‘the unprofitability of replacing still serviceable manufacturing plants’. A factory built around a steam engine and its line shafts could not simply be rewired, and was rebuilt only once the old building had served its time.

Most people seem still to think that something similar will be true of AI. Arvind Narayanan and Sayash Kapoor have argued, in ‘AI as Normal Technology’, that its diffusion will take decades, because ‘the speed of diffusion is inherently limited by the speed at which not only individuals, but also organizations and institutions, can adapt’. This contention must be taken seriously, and as a description of how most organisations will behave, it is probably largely correct.

The principal cause of the delay in electrification, however, has no real counterpart in AI. At the point of consumption, AI requires no new factories. It arrives over infrastructure that already connects the world and sits on every desk, and what it renders obsolete are intangible processes and habits rather than expensive buildings and machinery. Moreover, knowledge work has already been conducted on computers for a generation, so that its inputs and outputs, such as documents, correspondence, spreadsheets and records, exist in digital form and in the formats that the models read and write. Very little needs to be digitised, wired or installed before a firm can begin reinventing itself around AI. The models are reached over the internet and paid for as they are used, so that a firm need commit no capital in advance. An improved model reaches every user on the day of its release, without the replacement of any equipment. And the interface is ordinary language, which every professional already speaks. Of the conditions that held electricity back, only one remains, and it is the one that no provider can supply, namely the reorganisation of the firm around the new technology.

Narayanan and Kapoor may well be right that most organisations will take their time to adapt, but it does not follow that the disruption will be slow. It requires only one firm in a given market to reinvent itself around AI for every other firm in that market to face extraordinary competitive pressure, and from that point the pace of change is no longer theirs to choose. In other words, many firms may well take their time, but will soon find that they are no longer able to compete, and will either join the race or die.

My third premise therefore concerns the logic by which a free market determines which firms survive and which fail. It is often assumed that the forces which shape biological evolution are confined to the biological realm, which is manifestly not the case. Wherever entities vary, compete for scarce resources and pass on what succeeds, selection operates, and an economy of competing firms is no exception. Herbert Spencer’s phrase ‘survival of the fittest’ is, moreover, commonly misread. The fittest are those best adapted to their environment, which is not the same as the strongest, and when the environment itself changes, those who were the fittest before may swiftly find that they are maladapted to the new conditions.

In 1950, the economist Armen Alchian, one of the first to apply evolutionary reasoning to economics, argued that markets select firms in much the same way, so that firms which fail to make profits disappear, whatever their intentions. Later evolutionary economists explained why firms so seldom adapt by design. Richard Nelson and Sidney Winter treated a firm’s routines as the counterpart of genes, the stable patterns of behaviour through which a firm reproduces its work from one year to the next. Michael Hannan and John Freeman went further. In their account, markets reward established organisations for reliability and accountability, both of which depend on routines that can be reproduced faithfully, so that the very quality which made a firm successful also makes it resistant to change. They predicted, moreover, that reorganisation is itself so disruptive that it raises the risk of failure, which may explain why incumbents hesitate even when they see the need to change. It follows that industries change chiefly by replacement, as new firms with new routines displace old firms that cannot acquire them.

The implications of this logic for AI follow directly. A firm that has redesigned itself around abundant intelligence is able to produce work of the same or better quality at a fraction of the cost and in a fraction of the time. An advantage of that magnitude cannot be offset by working harder or charging less. Firms that do not redesign themselves will lose clients and margin to those that have, and many will in time be driven out of the market altogether. A firm that waits to see how matters unfold is therefore leaving its fate to the market, and the market is unlikely to decide in its favour. To paraphrase Dostoyevsky, nature does not ask for our permission and is indifferent to our wishes, and so we must accept her laws whether we like them or not. The same is true for the free market.

The same mechanism of selection operates not only among firms that offer the same service but also among entirely different ways of solving the same problem, and it sometimes acts on a whole category of service at once. Some tutoring businesses, to take but one example, have already begun to close, telling parents, in effect, that AI models now offer their children a far cheaper way to learn. No AI-native tutoring competitor was needed to bring this about. A cheaper and more accessible solution to the same problem comes along, and the market moves to adopt it.

Selection among firms has a feature that the debate about the pace of AI adoption tends to overlook. In any given market, the step from zero to one matters far more than the step from one to two. The incumbents in a market come under pressure as soon as one firm in it adopts AI, long before the rest of the market has followed.

Consider a market for professional services in which a dozen firms of broadly similar nature compete for the same clients. Suppose that a participant in that market, whether a new entrant, one of the incumbents or a client’s own in-house team, rebuilds its operations around AI, so that it can deliver work of the same or better quality in a day rather than a fortnight, and at a fraction of the fee. It does not matter which of them it is; from that moment, every other firm in the market is suddenly uncompetitive.

The selective pressure then travels through two channels. The first is market share, as clients move their work to the firm that is faster and cheaper. The second is overall price pressure, and it operates even if no client moves at all, because every client of every other firm now knows what the same service can cost. A client need not move its work in order to use that knowledge, but need only cite the lower price in negotiation, and its present adviser must then either match it or lose the client. A firm whose costs have not fallen can match the price only by surrendering its margin, so that the pressure reaches every firm in the market, whether or not any of its clients has left yet.

In the past, an incumbent that lost ground to a more efficient rival had time on its side. A professional services firm can grow only as fast as it can recruit and train qualified people, so that even the most efficient rival could take on new clients only gradually, and the incumbent had years in which to respond. An AI-native firm faces no such limit. Once its AI system is in proper operation, its capacity grows with the computing power it buys rather than the people it hires, so that it can take on new clients almost as quickly as they arrive, and the incumbent may have months, rather than years, in which to respond.

The resulting position is a trap from which no firm can escape. Collectively, the incumbents might well prefer that none of them adopted AI, so that fees and margins stayed where they were. Once one firm has adopted AI, however, abstaining is no longer a viable choice for any of the others, and any understanding among them to hold the line would be game-theoretically unstable, because each would gain by being the first to break it. The protections peculiar to the professions, such as licensing, regulation and long relationships of trust, delay this process but do not prevent it.

The pace at which a market is disrupted is set not by its average firm, but by whoever first brings the capabilities of the technological frontier to bear on it.

The competition that follows resembles the race in Lewis Carroll’s Through the Looking-Glass, in which the Red Queen explains to Alice that ‘it takes all the running you can do, to keep in the same place’. Firms that run after the first mover will not necessarily grow richer, but firms that stand still will fall behind. Much of the gain will ultimately pass instead to consumers, who will receive better professional services at lower prices.

At this point the analogy with biology reaches its limit. An organism cannot choose its genes, but a firm can choose its routines, and so need not wait for the environment to adopt or reject it. It can deliberately acquire the routines that the new environment rewards, reorganising one function at a time, as I suggested in ‘The Roads of Reinvention’, so that the risks of reorganisation are taken in measured steps. It can also choose to move its people above the loop rather than remove them from it. The firms that survive will be those best adapted to an environment in which intelligence is abundant, and for a firm, unlike an organism, adaptation is a matter of choice.

It remains to lift the assumption with which I began. Capability has not stopped improving, and there is little reason to think that it will. Every step of the argument above holds on the most conservative assumption about what AI can do, and holds with greater force on any realistic one, since each improvement in the models further compounds the advantage of the firm that has adapted.

None of this is a counsel of despair. The same pressure that will dismantle firms built around scarce intelligence will reward those that rebuild around its abundance, and their clients will be better served for it. I would nevertheless expect the change to come more swiftly and more radically than most firms now anticipate, and in each market to begin with the first firm that decides to adapt.

For more than a century, most professional firms have been organised as pyramids, with a few partners at the top, a larger body of managers beneath them and a broad base of juniors at the bottom. The arrangement, naturally, was no accident, because for as long as cognitive work required human minds, it was the most economical way to produce that work, as it allowed the professional judgement of a few seniors to be leveraged over the cognitive labour of many juniors.

AI fundamentally removes the premise on which that structure rests. My argument in this essay is that the pyramid of old will first narrow into something closer to a diamond, that the diamond will then collapse into something closer to a ladder, and that this will happen faster in the professions than in almost any other part of the economy.

The argument rests on two premises. The first is that AI now performs the very work that justified the broad base of the pyramid. The second is that competition will force every firm to act on that fact, and to act quickly. Three consequences follow. The first is that the pyramid will narrow into a diamond, as firms stop hiring juniors for work that machines can do. The second is that the diamond cannot last, because a firm’s middle is drawn from its own base, and nothing will replenish it while other forces press on it from the sides. The third is that the few juniors who remain must be trained in a manner unlike anything the professions have known, because the groundwork through which juniors once learnt is gone. Taken together, these propositions lead to my conclusion that professional services, as we understand them today, will in short order be altered beyond recognition.

By ‘leverage’ in this essay, I mean the ratio of junior professionals to partners in a firm, which determines how far each partner’s judgement can be extended across the firm’s work.

The pyramid is often explained as an apprenticeship of sorts, in which the young learn the ins and outs of the profession from the old. Apprenticeship explains why a firm has juniors at all, because every firm must train the people who will one day take its partners’ places. It does not, however, explain why a firm typically has many juniors for every partner. Obviously, a firm that hired juniors only to train its successors would need few more of them than it has partners to replace. Yet most of the juniors a firm hires will never become partners, and the firm does not expect them to. They are hired for their labour, and the number hired is a matter of leverage, which follows from the nature of professional work.

Much of the day-to-day work of a professional firm is a kind of drudgery. Before a partner can advise a client, someone must research the authorities, read the documents, assemble the figures and produce the memoranda, schedules, drafts, working papers and other artefacts in which the advice is ultimately delivered. The bulk of this work has always been too low in value to justify a partner’s expensive time, and so partners have overseen and steered it while juniors performed it, with leverage further increased by interposing middle management. The partner’s judgement was the scarcest and most expensive resource in the firm, and the most economical way to deploy it was to spread it over the labour of managers and many juniors, each billed to clients at far more than they cost. It is, in effect, a form of labour arbitrage, in which the firm buys junior time at the price of a salary and sells it at the price of professional advice.

The arrangement had a further advantage, which is that the juniors’ billable work essentially paid for their training. Juniors learnt the profession by doing the groundwork for which clients paid, so that a firm’s investment in its future partners was largely a by-product of its present revenue.

Leverage is highest where professional work is procedural, as in audit fieldwork, document review and compliance, and lowest in bespoke advisory work, which has always depended on a few experienced practitioners. The argument that follows applies with most force to the former, but in some measure to all. Anyone who supposes, however, that bespoke advisory work lies beyond the reach of AI is much mistaken, for current models, given the proper context, can already perform much of it at a very high level.

The research, reading, assembly and drafting that fill a junior’s working day are precisely the kinds of work that AI now does faster, more cheaply and, in many cases, better than any human. Most of it is codified work, the application of knowledge that can be taught, written down and verified, and codified work is what AI does best. A model can now research a point of law, assemble a set of working papers or produce the first draft of a memorandum in a small fraction of the time a junior would need, and at an even smaller fraction of the cost.

Once the machine reliably performs that work, nothing remains to justify the headcount at the base. A firm employs juniors for their labour only for as long as their labour is the cheapest way of producing that groundwork, and that is no longer the case.

The labour market appears to have begun to register the change. In ‘Canaries in the Coal Mine?’, researchers at the Stanford Digital Economy Lab use payroll data covering millions of American workers to show that employment of 22- to 25-year-olds in the occupations most exposed to AI now stands 19% below where it would have been had it kept pace with that of their peers in less exposed occupations. The shortfall thus far has arisen primarily through reduced hiring rather than through outright dismissals, a distinction that becomes important when we turn to the middle of the pyramid below.

The savings that AI brings will not remain with the firms. As I argued in ‘Survival of the Fittest’, competition compels each firm to pass them on to its clients, on pain of being undercut by a rival that does. The clients themselves, who can now perform much of this research and drafting in-house with the same tools, will not pay much more than it would cost them to do so. Ultimately, fees will fall towards the cost of compute, and of the human judgement, that a given matter still requires.

I anticipate the objection that cheaper work will call forth more demand for it. On this view, as AI lowers the price of professional services, clients will want to buy more of them, including those new clients who could not previously afford them, and the additional work will sustain the base of the pyramid even as each piece of work requires less human time.

Demand may well expand. It is my personal view that it will. But the new demand that a fall in price unlocks is, by its nature, demand for low-price work. It comes from clients who could not afford professional advice at its former price, or concerns work that was never worth doing at that price. A firm can serve that demand only at the new, lower price, and it can meet that price profitably only by having machines do the research, analysis and drafting, since expensive human hours cannot be sold at a price set by the cost of computing.

A boom in such demand creates work for machines and for the few people who direct them. To the extent that any human base survives on this work, it will form a separate labour market of people who operate AI systems on low-price matters, and not the first rung of a career that leads to partnership. The belief that rising demand will preserve the pyramid is, by and large, a false comfort, and not a defensible view.

If AI merely made the base unnecessary, firms might still take their time in dismantling it. Competition removes that option. As I argued in ‘Survival of the Fittest’, a market economy is a competitive environment in which only the firms best adapted to its conditions survive, and AI has changed those conditions fundamentally.

The mechanism is rather simple. A firm whose base has been replaced by machines while partners still assure quality can offer the same work at a fraction of the price. Its competitors must then match that price or lose the client, and they can match it only by doing the same. The result is a race to the bottom on price, in which each firm is compelled to extract the greatest possible leverage from AI, and that means replacing whatever a machine can do for next to nothing. No firm can choose to keep its pyramid once its competitors have dismantled theirs. Clients accelerate the race. A general counsel or finance director who uses AI in their own work can see what a piece of work should now cost, and will not pay for hours that a machine has made unnecessary.

The economist Enrique Ide has shown, in a formal model, that firms competing in this way will automate more early-career work than is good for the economy as a whole, because the cost falls on the skills of future workers rather than on the firm that automates. No single firm can afford to correct for this, since a firm that keeps paying juniors for work a machine can do will be undercut by those that do not.

There are, of course, forces that slow the change. Certain work is reserved by law for admitted or registered professionals, and many deliverables require the signature of a qualified person. Professional liability makes firms cautious, and partners near the end of their careers may prefer to harvest the old model rather than rebuild it. These forces are real, but each protects the partner, not the pyramid. The law requires a qualified person to sign the opinion or the audit report and answer for it, but it says nothing about how many juniors must prepare it. A partner close to retirement may succeed in postponing change inside the firm, but cannot postpone it in the market. Clients will take their matters to competitors that never built a pyramid and charge a fraction of the fee, and the partners who remain will inherit a firm that can no longer compete.

For a time, as the base of the pyramid collapses, the professional firm will look more like a diamond, with a narrow base beneath a broad middle of experienced people who direct and review what machines produce. This middle is not the routing layer of management which, as I argued in ‘The Roads of Reinvention’, largely falls away once information no longer needs to be carried up and down a hierarchy. It consists of practitioners whose role has moved from producing analysis and drafts to directing and reviewing the machine’s output.

The middle persists for a time for three reasons. The first is responsibility. The machine cannot answer for its work, and someone with the standing to do so must. The second is tacit knowledge. AI substitutes most readily for the codified knowledge that juniors bring from their studies, but it complements the tacit knowledge that practitioners acquire through years of practice, and the Stanford researchers find faster employment growth among experienced workers in occupations that rely heavily on such knowledge. The third is that the review of machine work still requires an experienced eye, at least until much of that review is itself automated.

There is also a less flattering reason. Part of the middle is broad only because its members were hired before the base narrowed. The diamond is, to that extent, what remains of the pyramid that preceded it, and it is a transitional structure that may well prove short-lived.

Three interlinking phenomena will, to my mind, collapse the diamond into a ladder.

The first is the (lack of) pipeline. A firm cannot promote to manager the people it never hired, and the middle of every firm is drawn from its own base of a few years earlier. Nor can the problem be solved by hiring experienced people from other firms, which works for any one firm but not for a profession in which every firm will reduce its intake. The effects are only starting to emerge, and the latest models have barely begun to be deployed in professional services, but such a reduction is all but inevitable. Notably, this hypothesis does not depend on any further improvement in AI. Even if the models stopped advancing today, their deployment alone would reduce the intake of juniors, and that reduced intake would work its way up through every firm as each cohort aged.

The second is pressure from the ‘sides’ of the diamond, as it were. As fees fall, the middle tier of a professional career becomes less lucrative, and fewer partnerships remain to compete for. Many in the middle will leave for other pursuits, and others will be forced out as machine capabilities advance even further. Some of the most capable will leave to found practices of their own, because, with AI performing the groundwork, they no longer need the expensive juniors that such a practice once required. The middle therefore narrows not only because of the effluxion of time due to the reduced pipeline, but also because of active departures.

The third is a falling demand for seniors. Many now ask how firms will find senior practitioners in future if they do not hire juniors today. The question assumes that tomorrow’s senior will do today’s senior’s work. As the models improve, the logic that removes the junior applies to much of what the senior now does as well, so that firms will need far fewer seniors, and those they need will be the people who direct the machine’s work and answer for it.

What remains is something closer to a ladder, with few people at each level and each senior followed by a successor rather than by a cohort. The leverage that juniors once supplied is supplied instead by machines, and, as I suggested in ‘Transcending the Loop’, it comes to be measured in the judgement calls a firm’s people can make rather than in the number of people each partner supervises. How quickly this happens, no one can say with confidence, but if I were a betting man, my money would be on ‘a few years at most’.

To the extent that firms continue to exist, juniors will still have a place in them, though in far smaller numbers, and will be valued for future judgement rather than present labour. In that sense, the professions will return to the apprenticeship model on which they were founded, but in a wholly new form.

The traditional apprenticeship model worked by slow absorption of skills. A junior spent years doing groundwork under supervision and acquired judgement incidentally, through the accumulation of countless small tasks and corrections. That model depended on the groundwork, but now the groundwork is going. Nor can universities fill the gap. They teach codified knowledge, which is the very knowledge that AI now supplies most readily, and they do not provide the practical experience that has become scarce and valuable. Something must stand between the university (to the extent that university education in the professional industries survives in the first place, which is highly doubtful) and the new working environment, and that is what I will call the practical academy.

The practical academy must overcome a difficulty first identified in industrial automation. In ‘Ironies of Automation’, published in 1983, the cognitive psychologist Lisanne Bainbridge observed that automating most of a task leaves its operators without the practice that their ongoing work once supplied, so that, far from needing less training, they need more. A junior who has only ever reviewed machine work may lack the eye to tell when it is wrong.

AI itself partially supplies the answer, for the same machine that deprives juniors of the groundwork through which they once learnt can supply them with practice instead. It can generate realistic matters without limit, plant errors of known kinds for the trainee to find, mark each attempt at once and set each exercise at the edge of the trainee’s competence. A firm’s method, once made explicit for its AI systems, also becomes the practical academy’s curriculum. Training can combine the oversight of AI on live matters, in which a senior reviews the junior’s review, with short and deliberate practice by hand and with simulated matters, much as pilots rehearse in simulators the emergencies they may never meet in the air. What once took years of low-involvement learning must now be compressed into months of intense practice. A junior who cannot achieve that leap in competence in so short a time will have nothing to offer that the machine does not already provide. This will select for highly motivated, highly competent and highly intelligent individuals.

With so few places available in the ladder, the tournament for partnership, once played out over many years inside the firm, now takes place at the gate, and in a sense before it, among students competing for places that most of them will never be offered. Firms should select for exceptional general intelligence, unusual industry and a command of the field deep enough to recognise when the machine is wrong. Mid-tier ability will not suffice in the slightest.

The practical academy must also be paid for. In the pyramid, juniors paid for their own training through the billable groundwork they produced. As Luis Garicano and Luis Rayo put it, ‘Apprenticeships let juniors pay for training by doing menial work. AI now performs an increasing share of that work, putting the bargain at risk.’Because firms will train only a few apprentices, the total cost of training them is limited. That cost now falls on the firm itself, however, paid in the time of its most expensive people instead of being recovered from clients, and a firm will bear it only if it deliberately chooses to invest in its future partners.

On these premises, and barring some black swan event, only a small minority of those currently starting their training to become lawyers, accountants, analysts and consultants (in my estimation, no more than one in twenty), will find gainful employment in these professions at all by the time they complete their studies and apprenticeships. That is a prediction, not a certainty, but in my opinion it would be dishonest for those of us who hold this view to suggest otherwise, or to paper over our concerns with platitudes. Anyone who aspires to one of these places, which will soon be fiercely contested, should ask themselves candidly whether they have an unusual devotion to their field and ability of the first rank, because, as I have argued above, those are the qualities that selection at the gate will require. The naked facts will not bend to anyone’s hopes, and those who confront them early will be best placed to act on them.

The professions are likely to be among the first and most heavily affected by AI, for reasons that compound. The first is that their product is cognitive work, which is the medium in which AI itself operates, so that there is no factory to retool and no physical constraint to slow adoption. The second is that the groundwork at the base of the pyramid is the most codified work in the firm, and codified work is what AI does best. The third is that labour is the largest cost of a professional firm, so that the savings are large and immediately visible. The fourth is that clients are sophisticated buyers who use AI themselves and can reasonably estimate what a given matter should cost. The fifth is that a competitor needs little capital to enter the market, so that a firm which keeps its pyramid will face rivals that never built one to begin with. And the forces that slow the change protect the partner, not the pyramid.

Each of these reasons would be enough on its own to unsettle the professions. Together, they make it likely that the pyramid will give way far sooner, and far more completely, than most of those who lead professional firms currently expect.

From the beginning, attentive intelligence, that faculty necessary for creating cosmos out of chaos, has been scarce. Since ancient times, mankind across cultures has understood the unparalleled transformative power of paying attention. Moses attends to the glimmer of the burning bush and there encounters the ground of Being; the watchful eye of Horus rescues the state from tyranny; and even the mighty Odin sacrifices his eye at Mimir’s Well to see more deeply into reality.

Attention, knowledge, and thinking have always been expensive treasures, slow to come by and difficult to scale. Consequently, all our institutions were erected in light of that scarcity, and few have been designed more particularly around it than the modern corporation. Hierarchies, reporting lines, approval chains, departmentalisation, specialisation and the division of labour are not ‘eternal forms’ of organised work, but phenomena largely caused by the scarcity of intelligence, a condition so old and elemental that we have ceased even to think about it.

Attentive intelligence remains foundational, but in a stunning inversion, many of its component capabilities are now becoming abundant, inexpensive, general, and available on demand. It is difficult to overstate the profundity of this fact. This increase in capability has happened, and continues to happen, at such a pace that the primordial constraint of scarce intelligence has begun to lift for the first time. Now, organisations are becoming limited less by scarce cognitive capability than by their inability to effectively harness and channel the unfolding intelligence explosion.

Artificial intelligence (AI) can already perform substantial portions of knowledge work at a remarkably useful level and at an economically compelling price. This is not a claim about a hypothetical future breakthrough but an empirical observation regarding the world as it now is. The future has arrived in raw capability much faster than we have been able to adapt our organisational forms to capture it.

But this lag will not endure long. AI is not the spectre of some hype cycle, nor an incoherent absurdity prophesied by futurists. Far from being an incremental technological advance, it is a real and proliferating general-purpose technology, and it represents a paradigm shift unprecedented in its implications.

The coming onto the scene of a general-purpose technology does not transform the world by making old methods of working more efficient. At first, people tend to try to insert such new technologies into existing processes, but soon find that these processes constrain the potential of the new technology. Only when those processes and organisational forms are rebuilt around what the new technology makes possible does it become possible to unleash its true potential.

Consider, for example, electricity, which was initially threaded through factories still arranged around steam-powered shafts and belts. Its real industrial power was unleashed only when machinery, production flows, and the factory floor were redesigned with electricity in mind. The advantages afforded by new general-purpose technologies can only be realised when institutions reorganise around such new technologies.

Today, most organisations experimenting with AI still make the mistake of grafting it onto existing processes without a thought to how the processes themselves must change. People try to fasten an AI assistant to their inboxes and documents inside existing workflows while preserving the queues, hand-offs, approvals, roles, and partitions embedded in the old organisational form. This approach may afford marginal gains in efficiency, but can never unleash the transformative productive power of AI.

The true vector of transformation is not the task, tool, individual job or department, but the institution surrounding all of these: its workflows and methods, proprietary knowledge, constituent roles, authority, goals, evaluations, verification function, and integration of feedback. To capture the full value of abundant intelligence, these elements must all be made to cohere around it, which can only be done by reorganising them (and therefore the organisation) around its presence. Organisations that reorganise around this new source of leverage are able to compound the advantage it offers, whereas those that refuse to adapt will find it increasingly difficult to remain competitive in this new paradigm.

An organisation can be understood as a system for solving one particular kind of problem (or a related set of problems) better, faster, or cheaper than the surrounding economy. In a competitive market, the organisation perpetuates its existence by coordinating people and resources to solve this problem more effectively and efficiently than a collection of disconnected actors or transactions could.

Stripped of all the paraphernalia which attends running a business, there lies at the heart of an organisation its innermost loop: the flywheel of actions in which the organisation actually produces the outcomes for which the market rewards it. All other business functions support, govern, finance, protect, or represent this loop, but ultimately, this loop is the reason for the organisation’s existence, its economic justification and raison d’être.

In their efforts to keep pace with technological advancements, many organisations fail to see the importance or nature of their innermost loop clearly. Some confuse visible activity with value creation and mistake the back office for the engine. They often automate processes that are easy to observe but ultimately have no bearing on the success of the business or the furtherance of its aims. In thinking about reorganisation around abundant intelligence, the first step is therefore diagnostic: where, precisely, does this organisation turn attention, effort, resources, and judgement into value? And once this value-creating flywheel has been identified, the next question is how AI can improve or accelerate it.

The proper aim of introducing AI into the organisation’s systems is not mere expedience or economy, but the multiplication of productive capacity in service of the organisation’s purpose. The question is emphatically not how to hold revenue constant while using AI to force costs down. The real question is how to multiply the value an organisation can produce, and what larger or more capable manifestation it could achieve if it were willing and fully able to pursue the opportunities available to it.

Implemented correctly, AI yields compounding returns. Each completed cycle of the flywheel, coupled with the capacity to evaluate its outputs and learn from correction, produces better memory, richer context, clearer methodology, and greater capacity for the next turn of the wheel. And as the institution works faster, it also learns faster. Its intelligence, and the institutional scaffolding around it, becomes increasingly adept at solving those particular problems it exists to solve. Given the right harness, general intelligence is able to specialise. The efficiencies created by this compounding learning loop are why AI can produce a multiplication of capacity rather than only marginal gains in productivity.

In the age of abundant intelligence, a firm need not abandon the innermost loop through which it has always created value. Instead, it must discover how intelligence can enter that loop and multiply its power, just as a fire given oxygen burns at far greater intensity. Once the flywheel begins to turn at a much greater speed, however, the old organisational form, built around its former pace, itself begins to constrain the newfound productive potential. The old procedures must therefore be reconceived around the accelerated centre so as not to negate the gains achieved by the accelerated flywheel.

An organisation cannot indefinitely preserve a form built around scarce intelligence while its competitors reorganise around abundance. A rival that has successfully brought intelligence into its innermost loop and rebuilt around the resulting capacity can create more value and respond more quickly, serving the market at prices and on terms the older form cannot hope to match. The choice is therefore not a neutral, preference-based decision between transformation and stability, but increasingly an existential one between (i) achieving forward progress by remaking the organisation around an accelerated flywheel and (ii) gradually becoming incapable of competing with those that have done so.

By its nature, such reorganisation will ultimately require substantial creative destruction inside organisations. Old processes and organisational structures, and sometimes even whole business models, will need to be dismantled over time. This, however, is not a process of aimless destruction, but the enabling of a much-improved system to take the place of what came before.

The goal of a reorganisation around abundant intelligence is not simply to assemble a useful collection of tools. The goal is the attainment of institutional intelligence: intelligence conjoined to an organisation’s methods, knowledge, rules, people, infrastructure, legal authority, purpose, and obligations. Particular models and tools may change, but these institutional foundations and the scaffolding surrounding the intelligence must remain the organisation’s own. Although most people today will still find it a strange metaphor, it is instructive to conceive of institutional intelligence as being the equivalent of an organisation being given a ‘mind’ of its own, turning the organisation into an entity that is a persistent, governed, self-organising, problem-solving whole with continuity across time.

In that sense, the legal fiction of the juristic person begins to acquire an operational reality. The organisation becomes more coherent and more able than any one individual to bring the relevant context into view. It also becomes more adept at resolving incoherence and disagreement within its systems. This is not to say that humans surrender moral agency, nor that such an organisational ‘mind’ represents anything like human consciousness. Institutional intelligence’s authority remains derivative, with human beings remaining responsible for imposing a robust governance and verification framework. An organisation’s people remain the source of its direction and the final authority. They determine its ends, delineate its limits, set the standards by which its work is validated, and intervene where necessary.

But neither is an intelligent institution a passive instrument awaiting continuous human direction. Once intelligence is conjoined to memory and methods, and to the mechanisms through which it can act, it can perceive conditions across the organisation, recognise patterns, deliberate and form preliminary assessments, initiate routine action, and carry matters forward without requiring a human hand at every stage. The institution is no longer a loosely organised collection of people, processes, and tools, coordinated through persistent managerial efforts. It becomes a coherent self-organised system, optimised around unleashing the innermost loop.

Conceptually, the organisation itself is therefore the appropriate unit of transformation. Left to bottom-up experimentation and adoption alone, AI fastens itself to existing tasks, scattering capability across personal accounts and disconnected tools while processes remain unchanged and risks obscured. The institution’s proprietary advantages are left exposed to systems it neither governs nor understands. Genuine AI adoption must therefore be institution-led, for only the organisation can unite intelligence with common context, shared infrastructure, redesigned processes, clear authority, and systematic verification. Direction must come from the whole. Only then do isolated gains become compounding institutional capacity.

As raw cognitive capability becomes abundant, competitive advantage no longer follows simply from access to clever people or technical knowledge. It moves towards what intelligence alone cannot supply: data, unique context, trust and relationships, legitimacy, authority, physical presence, coordination, judgement, taste, and the willingness to assume responsibility. These are not decorative supplements to intelligence. They determine how intelligence becomes operative in practice, and remain decisive sources of institutional differentiation.

One should be clear-eyed that the societal transformation necessitated by the advent of AI will be disruptive. Existing roles will change, and many will not persist in their present form. It would be dishonest to pretend that the transition will be painless or that every person will move naturally into more consequential work. But it would be equally mistaken to judge the future only by the tasks and occupations it displaces. The proper aim of transformation is the multiplication of productive capacity, not merely the diminution of labour costs. Wherever possible, organisations should use AI as a source of perpetual promotion for their people: using machine capability to enlarge the scope of human agency, elevating them towards roles which leverage their insight, judgement, taste, relationships, accountability, and leadership.

This is a deliberate commitment, not an automatic consequence of organisational transformation. Institutions must choose to elevate their people rather than reduce them and must protect the paths by which expertise and judgement are acquired. Two firms with equal access to abundant intelligence may therefore achieve very different outcomes: one may use it only to compress the cost of its current routines, while another may bind AI to its purpose and free its people to undertake more ambitious and consequential work. As cognitive capability expands, human agency, responsibility, and ambition do not become obsolete, but more important sources of leverage than ever before.

Our purpose with hegemonikon is to help organisations navigate rapid technological change, remain abreast of the frontier, and transmute possibility into capability. Presently, that purpose finds its chief expression in helping organisations convert abundant general intelligence into institutional intelligence of their own. In each engagement, our goal is to help discover the organisation’s innermost loop, infuse it with intelligence and redesign the surrounding system so that greater capability becomes an enduring institutional property.

Our partnership with an organisation therefore begins with a process of discovery in which we develop a deep understanding of the organisation and explore what the latest technology makes possible for it, before agreeing on a desired destination. We then design and engineer the particular solutions required to achieve our partners’ goals, followed by hands-on implementation and training. We do not leave a partner with a collection of tools, but with the ability to multiply its productive capacity, with iterative improvements and support over time to ensure that our partners continue to gain maximal leverage as new capabilities emerge.

To do this, hegemonikon combines strategic judgement and commercial acumen with technical engineering capability, grounded in continuous research at the technological frontier. Our Engineering Lab, Deployment Corps and Futures Bureau work in lockstep to ensure that our partners are always able to benefit from the latest innovations in pursuing their goals.

The Stoics of old gave the name hegemonikon to mankind’s ruling faculty: the seat of reason, perception, attention and judgement by which the whole is directed. We adopted this name because it expresses our conviction that attentive intelligence is a foundational force in the world which must be turned towards worthy ends. By making intelligence institutional, we help our partners actualise that faculty in ways that have never before been possible, systematically giving an organisation as a whole a coherent capacity to sense, deliberate, remember, decide, act and learn.

We set out on this mission because we believe that organisations will require competent, industrious, and trustworthy partners in navigating the significant disruptions resulting from rapid technological changes that are now unfolding at an ever-accelerating pace.

And ultimately, we believe that abundant intelligence, harnessed productively by institutions, can transform and ennoble the world.