The Job Survives. The Worker Doesn’t.
Part II of a series. The first essay asked how the great professional services firms survive the machine. This one turns from the firm to the human inside it, and picks up a thread I have been pulling across Genesis for a year: that the decisive action is not in the technology, which is converging everywhere, but in the behaviour we build around it.
The loudest argument about AI and work is the wrong one. It asks whether the machine takes the job. That question is binary, dramatic and, for now, mostly unanswerable, which is exactly why it dominates the conversation. Meanwhile a quieter change is already underway, and it is the one that matters. Before a single role is cut, AI has altered how the people in that role think, trust, learn, exert effort and earn status. The job may survive untouched on the org chart while the behaviour inside it is rewired completely.
I have argued before that AI does not fix an organisation, it amplifies it: the same tool, in the same quarter, drives one team into burnout and frees another for its finest work. This essay is the mechanism behind that claim. The behavioural shift is not a side effect of the automation story. It is the story. Whether this decade delivers the Star Trek future of governed, shared abundance or the Mad Max future where it is neither will be decided not by what the models can do, but by how we choose to behave around them. The evidence for that is now substantial, and most of it points somewhere uncomfortable.
I. The debate we are having, and the one we are not
Ask most boards about AI and work and they reach immediately for headcount. Will we need fewer people. Which functions go first. It is the question the spreadsheet knows how to answer, so it is the question that gets asked.
And the number they reach for to settle it is the one I have argued we should retire. The headline unemployment rate, the single most trusted figure in economic life, is structurally blind to what AI is actually doing. It can sit calmly at around four per cent while the transformation runs beneath it, because unemployment measures one narrow condition and misses almost everything that matters here. A firm that once hired two hundred graduates a year and now hires eighty, handing the rest to agents, barely moves the rate; the graduates who were never hired do not register as unemployed, they simply never appear. Discouraged workers who stop looking drop out of the count entirely. Participation slips a half-point and no headline notices. The youngest cohort in the most exposed occupations already sits close to a fifth below where it should be, and the aggregate figure stays green throughout. We are watching the wrong dial, and its steadiness is not reassurance. It is latency.
But the real-world data on how AI is actually used tells a different and more interesting story. The best running record we have, drawn from tens of millions of genuine work conversations with a frontier assistant and mapped against occupational task lists, shows that AI today leans toward augmentation rather than wholesale automation: in the early readings, roughly three in five uses enhanced a human doing the task, while two in five handed the task over. Depth is shallow. Only around one occupation in twenty-five uses AI for three quarters or more of its tasks. For most people, for most of the day, the machine is a collaborator, not a replacement.
That is the reassuring headline. Here is the part that should hold your attention. Across successive readings of that same record, the balance has been drifting. The share of fully delegated, directive requests, the “do this for me and I will not look too hard at the result” pattern, climbed from roughly a quarter to nearly two fifths inside a single year. The question is not whether we sit at augmentation or automation today. It is how fast, and how thoughtlessly, we are sliding from one to the other. Augmentation is not a destination. It is a behaviour, and behaviours erode.
This is also why the unit of analysis has to change. As I argued in Directors or Operators, agentic AI quietly moves the unit of work from the task to the loop. We are no longer only doing tasks or handing them over; we are standing inside loops that run partly on their own, and the whole question of competence becomes whether we direct those loops or merely operate them. Almost half of frontline workers now report spending more time managing and directing AI than doing the work itself. The behaviour that matters is no longer how you perform a task. It is how you govern a loop.
II. The levelling of expertise, and the Frozen Workforce
The single most robust finding in the workplace literature is also the most double-edged. When you give a capable AI to a population of workers and measure what happens, the gains are not evenly spread. They concentrate, dramatically, at the bottom.
In a widely cited study of several thousand customer-support agents, access to an AI assistant raised issues resolved per hour by around a seventh on average, but the headline average hides the real result: novices and the least-skilled improved by roughly a third, while the most experienced improved barely at all. A landmark field experiment with several hundred management consultants found the same shape. Inside the tasks AI handled well, everyone got faster and better, but the lowest performers gained around forty-three per cent against seventeen for the top. The mechanism is the same in both cases. The model has absorbed the tacit craft of the best practitioners and now dispenses it to everyone, flattening the experience curve that used to separate a first-year from a tenth-year.
Read optimistically, and one influential camp of labour economists does read it this way, this is the great re-democratisation of expertise. The judgement that once took a decade to accumulate becomes available on demand, extending real competence to people and places that never had access to it. That is a genuine and hopeful prospect, and I do not want to dismiss it.
But look at what it does to the climb, because this is where I have to insist on the harder reading. The entire apprenticeship model of skilled work rests on a bargain: juniors do the unglamorous foundational work, slowly, badly, at first, and in doing it they build the judgement that later makes them seniors. Here is the paradox at the centre of what I have called the Frozen Workforce: judgement is the scarcest and fastest-appreciating input in the AI economy, and the apprenticeship that historically produced it was a free by-product of exactly the drudgery AI now absorbs. Remove the drudgery and you do not just save time. You quietly switch off the machine that manufactured judgement.
So the Frozen Workforce is not a story about mass unemployment, and this is the part most commentary gets wrong. A frozen worker is not unemployed. The early labour-market signature is not redundancy but a leadership pipeline quietly starved of the repetitions it needs, years before anyone notices the gap. The youngest cohort, those in their first years in the most AI-exposed occupations, already shows employment falling by double digits relative to their older colleagues, by one estimate as much as a fifth, while hiring freezes and attrition absorb the shock long before it surfaces as a headline redundancy figure. We are not firing our way into the new economy. We are failing to hire our way into it, and certifying a generation for tasks that are being automated while never teaching them to direct the systems doing the automating. Operators, where we needed directors.
III. The jagged frontier and the competence illusion
The second finding every leader should internalise is that AI’s competence is not a smooth hill, higher in easy places and lower in hard ones. It is jagged. Two tasks that look equally difficult to a human can sit on opposite sides of a cliff edge, one handled brilliantly, the other failed completely, and nothing in the output tells you which side you are on.
The same consulting experiment that showed large gains inside the frontier showed the cost of the edge. On a task deliberately chosen to sit just outside the model’s competence, the AI-assisted workers were nineteen percentage points less likely to reach the right answer than colleagues using no AI at all. Not merely unhelped. Actively worse, because the machine produced something fluent and confident and wrong, and fluent confident wrongness is the hardest kind of error to catch.
Two patterns of behaviour emerged among the people who navigated this well. Some worked as centaurs, dividing the labour cleanly, giving the machine what it was good at and keeping for themselves what it was not. Others worked as cyborgs, interleaving with the AI at every step. The cyborg style produced the most, but it also carried the most risk of the behaviour that quietly undoes all the gains: over-reliance, the slow slide from using the tool to trusting the tool, from checking its work to assuming it. The literature has a vivid phrase for what happens next. People fall asleep at the wheel. The output looks finished, so the mind that should be auditing it disengages, precisely at the moments the jagged frontier makes auditing most necessary. This is the loop, ungoverned: the director quietly demoted to passenger without noticing the demotion.
IV. Thinking atrophy and the irony of automation
Here is where the behavioural change becomes physical, in the sense that it reshapes the mind doing the work.
A recent survey of several hundred knowledge workers found a clean and troubling relationship: the more a person trusted the AI, the less critical thinking they reported doing; the more they distrusted it, the harder they scrutinised its output. The effort did not vanish. It moved. The locus of work shifted, in the researchers’ phrase, from task execution to oversight, from thinking by doing to choosing among things the machine had already thought. That sounds like a promotion. It is not necessarily one. Choosing well requires the very judgement that doing used to build.
And judgement, like a muscle, weakens when it is not loaded. The uncomfortable anecdote that recurs across this research is the senior professional who, asked to reason through a problem at a whiteboard with no laptop and no assistant, finds they cannot, not because they have forgotten the knowledge but because they have not exercised the act of unassisted first-draft thinking in months. Cognitive offloading is reliable in both directions. It reliably improves the task in front of you, and it reliably degrades the capacity you handed over.
There is an old principle in the study of automation, articulated decades before this wave, that we are now rediscovering at scale. The better the automation, the more critical and the more difficult the human’s residual role becomes. You are left to supervise a system that is right almost all the time, which means your attention is least engaged exactly when the rare failure arrives and your intervention matters most. We are industrialising that irony across every desk in the knowledge economy, and calling it productivity.
V. The trust tax, and why the frozen are not lazy
The most under-counted cost of all is not cognitive. It is social, and it is where two of my long-running threads meet.
When AI makes it cheap to produce something that looks like work, people produce it. The research gives this a name: the polished, confident, substance-light output that is passed to a colleague as if it were finished, when in fact the real work of making sense of it has simply been shifted onto the recipient. In one survey, two in five workers reported receiving this in a single month, each instance costing nearly two hours to untangle. Scaled across a large organisation, the hidden rework runs to millions a year. That is the measurable cost.
The unmeasurable cost is worse, and it is the one that connects to the Frozen Workforce in a way I did not fully appreciate until the data arrived. When colleagues receive this kind of output, they do not just lose the time. They revise their opinion of the sender. Around two in five trusted that colleague less afterwards; roughly half judged them less capable, less creative, less reliable. Now hold that beside a second finding: in one large study, the majority of workers who held back from using AI at all were not frozen by fear of the technology. They were making an accurate reading of how their peers would judge them. I have argued before that the paralysis behind the Frozen Workforce is rarely located in the workers themselves, and here is the proof in behavioural form. People are caught in a genuine bind: use the tool and risk being seen as the person who sends slop and cuts corners, or withhold it and fall behind on output. That is not laziness. That is a rational response to a trust environment no one has bothered to design.
This is also the Measurement Gap I described at the level of the economy, now relocated to the office. When you cannot see effort, and output is suddenly infinite and cheap, the proxies you used to manage by stop working. You can no longer read reliability off a finished document, because the document tells you nothing about who, or what, produced it, or whether anyone checked. I have long argued that human experience is not one thing but two, that HX equals CX plus EX, that you cannot deliver a good experience to the customer out of an organisation quietly degrading the experience of its own people. The trust tax is EX decay in its purest form, compounded by a layer of concealment, because many now hide their AI use, which means a colleague cannot even calibrate how much to trust what they have been handed. An organisation cannot run on outputs whose provenance and effort are deliberately obscured. That is not a tooling problem. It is a trust problem, and trust is behavioural.
And concealment is not a character flaw, it is a design outcome. I have argued elsewhere that when you make honesty impossible you should not be surprised when people build something else in its place. Punish honest disclosure of AI use, whether through peer judgement or through targets that assume the old speeds, and you do not get honesty; you manufacture the concealment, the quiet cutting of corners, the shadow workflow nobody will admit to. The behaviour follows the incentive every time. If you want a workforce that is candid about how the work was done, you have to make candour the safe choice, which is a thing you build, not a thing you exhort.
VI. Faster, busier, worse: the productivity that never reaches home
We were promised time. The efficiency gains, the story went, would hand hours back to people, to think, to rest, to do the deeper work. Look closely at the behaviour and that is not what happens.
When everyone has the machine, the machine stops being an advantage and becomes the baseline. The bar does not stay still while you do the same work faster; it rises to meet the new speed. More decks, more analyses, more drafts, produced faster, expected sooner, by everyone. The effort that was supposed to be saved is competed away, and what is left in its place is not leisure but a subtler intensification: the anxiety of keeping pace, the comparison against colleagues who seem to be producing more, the low hum of never being sure your skills will still matter next quarter. The efficiency is real. The relief is not. Faster, busier, and, often, worse.
This is one reason the grand macroeconomic payoff keeps failing to arrive in the figures. One sober estimate puts AI’s boost to total factor productivity at well under one per cent over a decade, a fraction of the revolution the rhetoric implies. Some of that is timing, the usual lag before a general technology reorganises the work around it. But some of it is behavioural leakage, and it is the micro-level source of what I have called Ghost GDP. At the macro scale, Ghost GDP is output that registers in the national accounts yet never completes the circuit into households and wages. At the desk, it is the same leakage one layer down: productivity created and then lost before it becomes value, competed away into higher expectations, offloaded into slop that others must repair, handed to clients as lower prices, or burned supervising a jagged machine. The gap between what AI can do and what organisations and societies actually capture is not, in the main, a gap in capability. It is a gap in behaviour, in measurement and in design, and the money that falls into it does not vanish so much as fail to reach the people it was supposed to reach.
This is why the two blind metrics belong together. The unemployment rate stays green because the jobs were never cut, only never created; the output statistics stay green because the value was produced, only never circulated. One dial misses the frozen entrant, the other misses the household that never receives the gain. Manage by either and you will conclude that everything is fine right up until it very obviously is not. The Measurement Gap is not an accounting footnote. It is the reason a transformation of this size can run for years while every official instrument reports calm.
VII. The Fork runs through the org chart
Everything above can be read two ways, and that is the entire point. The same technology, placed in front of the same worker, produces radically different outcomes depending on how the human behaves around it. This is the choice I keep returning to, The Fork: one road toward a Star Trek future of governed, shared abundance, the other toward Mad Max, where abundance is neither governed nor shared. We tend to debate that fork at the altitude of nations and economies, where it feels suitably grand and conveniently distant. It is not distant. The fork runs through every organisation, every team and every workflow, and it is chosen not at summits but on ordinary Tuesdays.
| Dimension | The Star Trek path (augmenting) | The Mad Max path (hollowing) |
|---|---|---|
| Mode of use | Human directs the loop, delegates narrowly | Human hands over the loop, checks rarely |
| Effect on skill | Expertise extended and deepened | Expertise offloaded and atrophied |
| The junior | Learns faster with a tutor at the elbow | Frozen out before judgement can form |
| Output to colleagues | Work with real substance behind it | Slop that shifts effort downstream |
| Effect on trust | Reliability compounds | The trust tax; everything re-verified |
| Where the gains go | Captured as real value and capacity | Lost to Ghost GDP before reaching home |
Notice that the fork is not decided by the model’s intelligence. A more capable model makes both paths more extreme: a better tutor on the augmenting road, a more convincing source of confident error on the hollowing one. The decisive variable is human and organisational behaviour. This is why I argue so insistently for the Long-AND, not the Short-OR. The lazy framing is replace-or-cut, a short-term optimisation that takes the hollowing road by default because it is the one the spreadsheet can see. It is the Turing trap: using the machine only to mimic and displace human labour, which is precisely the configuration that does not compound. The durable framing moves through disruption to genuine augmentation and then to the creation of new work, which is harder, slower and the only sequence that pays off. Abundance and disruption arrive on the same curve. The ordering is set by institutional choice, not by the technology.
VIII. Governing the loop before it governs us
If behaviour is the variable, then behaviour is the thing to govern. Not with posters about responsible AI, but with the hard architecture of how work is designed, measured and rewarded. This is the move from prompts to loops to loop governance: once the unit of work is the loop, the unit of governance has to be the loop as well. Five commitments, in rough order of urgency.
I. Design for direction, not delegation. The drift toward full hand-off is the default, and defaults win unless you fight them. Build workflows that keep a human genuinely in the loop at the points of judgement, not as a rubber stamp but as a load-bearing step, and make the tool surface its reasoning so the human has something real to audit. The goal is directors, not operators.
II. Protect the climb. The apprenticeship ladder will not survive on nostalgia. It has to be rebuilt deliberately: preserve some unassisted repetitions for juniors even when the machine could do them faster, because the point of those reps was never the output, it was the judgement they grew. Deliberate difficulty is now a training strategy, not an inefficiency to be optimised away. A government or a firm that treats workforce infrastructure as AI infrastructure, funded and sequenced rather than left to HR after the technology has landed, is choosing the Star Trek road on purpose.
III. Make AI use legible. Concealment is corrosive and, as the trust data shows, rational under the wrong norms. Establish clear expectations for when AI is appropriate and when it is not, expect provenance rather than punishing honesty, and model discerning use from the top. The organisations that preach “use it everywhere, all the time” manufacture slop at scale and teach their people that withholding is safer than admitting; the ones that treat it as a tool for specific jobs, used on purpose, close the trust tax at source.
IV. Measure outcomes and judgement, not output volume. The billable hour is dying for the same reason the keystroke count never worked: when output is cheap, counting it rewards exactly the wrong behaviour. Close the Measurement Gap by measuring the quality of decisions, the soundness of the thinking and the value delivered. If you reward volume in an age of infinite volume, you will drown in slop.
V. Govern the loop, and make the governance enforceable. Safety and quality are not only properties of the system; they are properties of the human-machine loop and the behaviours inside it. The test I apply to any governance claim is simple: is it enforceable, or is it theatre. Oversight that no one has the time, skill or incentive to actually perform is not oversight. Two failure modes are worth naming because both pass audit on paper. The first is the line in the risk register that reads “human review.” I have argued that this is a promise about a person, not about a model version: it tells you a human glanced at the output, not that anyone with the competence, the time or the authority actually caught what the jagged frontier slipped past. Written down, it looks like control. Performed on a busy Tuesday by someone measured on throughput, it is a signature. The second is the quiet conflict at the heart of most assurance, the pattern where the auditor is the vendor, where the party promising the independent check also sells the thing being checked, or depends on the relationship that the check exists to question. An oversight function that cannot say no, or whose incentives all point toward yes, is decoration. Loop governance means making the human role real, resourced, independent and accountable, which is the only version of a human-in-the-loop that survives contact with the real pressures of the work.
IX. Short-term turbulence for long-term abundance
I remain, stubbornly, an optimist about where this ends, and readers of Genesis know I have staked that position from the start. The re-democratisation of expertise is real. The prospect of extending genuine capability to people and places long priced out of it is one of the most hopeful things technology has offered in a generation. The abundance is available.
But it is not automatic, and the evidence is clear that the default path does not lead there. Left to drift, the behaviours compound the wrong way: delegation deepens, judgement atrophies, the ladder freezes, trust decays, and the gains leak into Ghost GDP while the people doing the work feel busier, warier and less sure of their worth. That is the turbulence, and it is already measurable.
The machine is not deciding this. We are, in a thousand small behavioural choices that no one is currently governing. The job debate asks whether the work survives. The real question is whether the worker does: whether the human on the other side of the tool ends the decade sharper, more trusted and more capable, or faster, hollower and quietly diminished. That outcome is not a forecast. It is a design decision, taken on ordinary Tuesdays, and the time to make it deliberately is now, while the behaviour is still forming and the fork is still open.
I want to be precise about the stakes, because my own tagline is easy to misread as a reassurance. Short-term turbulence for long-term abundance is a wager, not a guarantee. The honest version of the phrase is that the turbulence leads to abundance or to its opposite, and the branch is chosen, not given. The window in which we still hold the wheel is not open indefinitely; the clock I started counting is already ringing, and the habits, norms and defaults we are laying down right now in how people behave around these tools are hardening into the structure we will be stuck with. Behaviour set thoughtlessly today becomes the architecture nobody can change tomorrow.
The turbulence is the near term. The abundance is the long one, and only one of the two futures on the far side of the fork. What stands between them is not better models. It is better behaviour, built on purpose, governed at the level of the loop, and chosen while choosing is still ours to do.
This essay extends arguments developed across Genesis, including The Frozen Workforce, Ghost GDP, the Measurement Gap, The Fork, the Long-AND not the Short-OR, the shift from the task to the loop, Retire the Unemployment Rate, The Auditor Is the Vendor, the reading of human review as a promise about a person rather than a model version, and the counting of the window in which we still hold the wheel. Its external evidence draws on frontier-lab usage research, the largest randomised workplace field experiments conducted to date, landmark field studies in professional-services and customer-service settings, recent survey work on critical thinking, cognitive offloading and collaborative trust, and the labour-economics literature on automation, augmentation and productivity, spanning 2023 to 2026. Figures are reported as published by their respective studies.
Leave a comment