That book is on a list I’ve just done for BookDNA, and it’s live now. Five books about the invisible rules that actually run your working life. Here it is.

Where it came from is this. I spent my first decade in business assuming the game was chess. Play well, win, everyone agrees on the rules. It took me an embarrassingly long time to notice that most rooms don’t run on the rules in the rulebook. They run on power, on persuasion, on who is quietly measuring themselves against who, and on the thing that makes a team come apart months before anybody says a word about it. Every one of these five named a rule I’d been losing to without knowing it was there.

And the deepest rule of the lot was the ladder itself. Do the junior work, get trained, move up. Nobody questioned it because it had always held. I build AI tools for a living, so I get to watch that one come apart in near real time, which is roughly why I wrote a book about what replaces it.

The Harari is the one that looks forward, and that’s the one I keep handing to people. I read it just before the pandemic, right as I was turning the agency into a consultancy. Timing wasn’t ideal. Sapiens looks backwards, this one looks at what’s coming, and that pulled me in a lot harder. It sits close to the future of work I’d actually want. I don’t think we get there, mind, because people will reject the things that make them feel bad long before they’ll change.

Cialdini’s Influence I’ve used almost daily since 2016. Not to manipulate anyone. To get my perspective across so it actually lands. It changed how I run advisory work more than anything else on the list: liking by being personable and honest, reciprocity by giving away real advice that sits outside the sale, and scarcity, which I now see everywhere and can’t unsee. The deflating part is watching someone’s biases hold so firm that no amount of care shifts them. I’d keep it away from anyone who persuades for manipulative ends, because in the wrong hands it’s a manual. In mine it’s a way of thinking a bit more carefully about my own thinking.

Then the 48 Laws, which I read in 2021 just after launching the advisory, when I started noticing dynamics in rooms I’d never been in before and wanted to understand what I was looking at. I read Greene as describing the world, not recommending it. Most people never consciously learn any of this. They’ve copied behaviours that seem to work, a bit like a cargo cult. Law ten, avoid the unhappy and the unlucky, still makes me deeply uncomfortable, and I think it only holds if you’re a zealot about it.

Five Dysfunctions is on there because I’ve caused at least two of them myself. Early on it was inattention to results, because my ego was convinced I ran the best digital agency going. Later, and more damaging, it was fear of conflict. Not dealing with a problem until it had turned personal. Any team above four or five people, these show up, and the hard bit is that the founder usually can’t fix them, because the founder is one of them. I’ve sat in a board meeting doing 85% of the work for half the value, watching one partner chase results and another dodge accountability for the entire hour.

The Peterson I picked for one rule, really. Compete with who you were rather than the people around you. I’d been watching colleagues as comparison rather than research, which dragged me down and helped none of the businesses. It falls apart in small edge cases, and people love an edge case as a reason to bin a whole book. The one I disagree with most is the personal truth one, because those truths move as your life moves. I doubt there’s a single one you hold forever.

Anyway. The list is live, there are five covers on it and a short bit from me under each one, and it took me longer to choose them than to write about them.

The Next Rung is out 19 January 2027 with Practical Inspiration Publishing. Pre-order at thenextrung.co.

Anthropic’s latest funding round valued the company north of $900 billion. That puts it ahead of OpenAI’s $852 billion valuation. It makes Anthropic worth more than all but a handful of the FTSE 100. More than BP. More than Unilever. More than most of the companies that employ the people who read this newsletter.

I built Orca on Anthropic’s API. Claude is the engine. The company whose technology sits inside my product is now valued at a figure that most people cannot meaningfully visualise.

I am not complaining. The technology works. It is why I chose it. But the concentration of capital in AI infrastructure companies is not a side story. It is the story.

Here’s the thing. Chapter 10 of the book lays out what I call the J-curve. Productivity gains from a new technology follow a pattern: disruption first, then a trough, then a turn. The question is not whether the turn comes. It is who captures the surplus when it does.

Between 1995 and 2025, US labour productivity rose 65 per cent. Median wages rose 8 per cent. The productivity gains went somewhere. They went to capital. The J-curve turned, and the workers whose labour generated the productivity did not see the upside.

AI is accelerating this pattern, not reversing it. Anthropic’s $900 billion valuation is capital pricing in a future where a small number of infrastructure companies capture an enormous share of the productivity surplus that AI generates across the entire economy.

Which brings us to the part that actually matters.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The Big Four tech companies, Alphabet, Microsoft, Amazon, Meta, are spending a combined $650 to $700 billion on AI infrastructure in 2026 alone. That is not investment in the sense that a factory is investment. A factory employs people in the town where it is built. AI infrastructure investment creates value that flows to shareholders, not to the displaced workers whose jobs funded the efficiency gains.

Chapter 17 of the book asks a question I keep coming back to. The most likely outcome if we do not make deliberate choices is not collapse. It is what I call neo-feudalism: a stable, unequal arrangement where the owners of AI infrastructure capture the surplus and everyone else adjusts downward. Not because anyone chooses it. Because it is the path of least resistance.

Neo-feudalism is the default because it does not require anyone to choose it.

Anthropic is not the villain of this story. Dario Amodei has been more transparent about the risks than most. The company publishes its safety research. It mapped 800 occupations and published the exposure data. I use the technology every day and I think the product is good.

But a company can be well-run, ethically serious, and still be part of a capital-concentration pattern that makes the world worse for most people. Those two things are not in conflict. They coexist quite comfortably, which is precisely the problem.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

I write this as someone inside the contradiction. I pay Anthropic for API access. I charge clients for the product I build on top of it. The value chain works. It works for me. The question from the book is whether it works for the 3 to 4 million UK knowledge workers whose roles are being absorbed by the same technology.

The capital is pooling. The gains are concentrating. The productivity surplus is being captured before the trough has even bottomed out.

A $900 billion valuation is not just a number. It is a signal about where the economy thinks the value is going. And it is not going to the people reading this newsletter.

The book spends two chapters on this. Chapter 10 on the J-curve and the political question of surplus capture. Chapter 17 on what happens if nobody makes the choice. I wrote them a year ago. The Anthropic valuation did not exist then. The argument did not need it. But it is a decent illustration of the point.

The UK needs 250,000 new tradespeople to meet its housing and infrastructure targets. 104,000 electricians alone by 2032. Heat pump installations, EV charging, the retrofit programme that keeps getting announced and then quietly re-announced with a lower number. The demand is real, funded, and growing.

Apprenticeship starts are falling.

The NI increases and the rising living wage have made it more expensive for small firms to take on apprentices. A plumbing firm with four employees does not have an HR department to navigate the levy. It has a bloke called Dave who does the books on Thursday evenings. Dave looked at the numbers and decided he could not afford to carry a first-year apprentice at the new rate. So the apprenticeship did not happen.

Multiply Dave by tens of thousands of SMEs and you get the paradox: the economy is creating demand for physical work at exactly the moment it is defunding the pipeline for physical workers.

I wrote about this in the book. Chapter 3 lays out the four-tier exposure map, and the bit that surprises people is the bottom tier. Trades requiring physical presence, care work, roles with regulatory human oversight. These are not “safe from AI” in some romantic sense. They are structurally less exposed because the body has to be in the room.

Which brings us to Gareth. He is a composite character in Chapter 11, but the pattern he represents is not composite at all. Plumber, 36 years in, heat pump certified, charges £85 an hour in the Tewkesbury area. Two ex-accountants applied for apprenticeships with him last year.

Read that again. Two people who trained for years to work in professional services looked at the labour market and decided that learning to fit boilers was the better bet.

Here’s the thing. The book’s argument in Chapter 3 is that the plumber who earns £60,000 a year is structurally more secure than the marketing director who earns £75,000. Not because plumbing is better work. Because AI cannot do plumbing. It can do marketing orchestration. The exposure inversion is real: AI hits the educated and well-paid first.

The trades were supposed to be the fallback. The thing you did if university was not for you. That framing was always patronising, but it also had a structural logic to it. There was a professional track and a vocational track, and the professional track paid more.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

That gap is closing. In some cases it has already closed. Gareth’s son is at college doing electrical installation. He will walk into a market with six-figure demand and shrinking supply. The marketing graduate from Bristol UWE who applied to 43 positions and got zero offers would swap places in a heartbeat.

But here is the paradox at its sharpest. The government wants 250,000 new tradespeople. The same government raised the costs that make it uneconomical for small firms to train them. The pipeline for the work that actually survives AI displacement is being squeezed by policies designed to fund other things.

And the people who might retrain into trades cannot afford to. You cannot retrain into a trade when you cannot afford to stop working for a week, let alone six months. That is Ryan’s problem from Chapter 11. Three platform jobs, £18,000 a year, rent at 63 per cent of take-home. He is not retraining into anything.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The apprenticeship paradox is not a policy failure in isolation. It is a symptom of a system that has not yet understood what is happening to its own labour market. The knowledge economy is shedding orchestration roles. The physical economy needs workers it cannot train fast enough. And the bridge between the two, the retraining infrastructure, is neither funded nor designed for the scale required.

I keep coming back to the same line from the book: the mechanism that made it fine last time is gone. Previous transitions had a next rung. Displaced agricultural workers moved into factories. Displaced factory workers moved into offices. This time, the next rung for displaced knowledge workers is trades and care work, and we are actively making it harder to get there.

The demand is real. The supply pipeline is broken. The irony writes itself. Except it is not funny if you are Dave, looking at the numbers on a Thursday evening, deciding he cannot take the risk.

The UK GDP growth forecast for 2026 has been halved. Down from 1.1 per cent to somewhere between 0.4 and 0.7 per cent, depending on whose model you trust. The Iran war’s energy shock did most of the damage. Inflation was 2.8 per cent in April but expected to climb. Mortgage rates are uncertain. The Bank of England is holding at 3.75 per cent and watching.

None of this is news to anyone paying attention. But the GDP cut matters for a reason that has nothing to do with macroeconomics textbooks and everything to do with the people whose jobs are being automated at the same time.

AI displacement hitting a strong economy is manageable. People lose roles but find new ones. The labour market absorbs the shock because there’s demand elsewhere. Retraining programmes have somewhere to retrain people into. Employers are hiring, just for different things. The adjustment is painful but navigable.

AI displacement hitting a weakened economy is a different problem. The new roles don’t materialise because nobody’s hiring. The retraining leads nowhere because the sectors that were supposed to absorb people are contracting too. The safety net stretches thinner because tax receipts fall while welfare claims rise. The whole system that was supposed to cushion the transition is under strain from the same direction.

That’s where we are.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Chapter 7 of the book is about the mortgage problem. Tom and Priya, the composite couple, bought a house in 2024 at 4.2 times their joint income. Priya was a paralegal earning £42,000. Tom worked in financial services compliance at £68,000. Mortgage was £460,000 on a £575,000 property. The numbers worked. They were meant to work. The lending criteria said they worked.

What the lending criteria assumed was that both salaries would persist for the duration of the mortgage. That assumption was already questionable when the book was written. With GDP halved, inflation rising, and the energy shock compressing corporate margins, it’s worse. Priya’s paralegal role is in the blast radius of AI-driven legal automation. Tom’s compliance role is increasingly being handled by automated monitoring systems. The mortgage assumed two salaries and a growing economy. They’re getting neither.

The energy shock is the accelerant. When Brent crude spiked past $100 in March and kept climbing, the downstream effects hit every sector that runs on energy, which is all of them. Corporate margins compressed. Discretionary spending contracted. CFOs who were considering AI adoption as a medium-term efficiency play suddenly had a short-term cost imperative. The restructurings that were planned for 2027 moved to 2026. The contractor renewals that were borderline became clear-cut.

The Bank of England’s hold at 3.75 per cent is a bet that inflation will moderate without further tightening. If they’re right, mortgage rates stabilise and Tom and Priya’s repayments stay where they are. If they’re wrong, and the Iran-driven energy shock pushes inflation above 3.5 per cent for more than two quarters, rates go up and the mortgage gets more expensive at exactly the moment one or both incomes are under threat.

The book’s argument in Chapter 7 is that the professional-class mortgage segment, roughly 1.2 to 1.5 million UK households with £540 to £750 billion in exposure, is structurally correlated. These are not random households with random employers. They are clustered in the same sectors, the same regions, the same income bands, and the same exposure categories. When AI displaces orchestration work, it doesn’t hit one household at a time. It hits a demographic.

A 15 to 20 per cent default rate in that segment would produce £81 to £150 billion in problem mortgages. The 2008 bank recapitalisation cost £137 billion. The numbers are comparable.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Now layer the GDP cut on top. Growth at 0.4 per cent means the economy is essentially flat. It means tax receipts are flat. It means government capacity to respond, whether through retraining, welfare, or mortgage support, is constrained. The fiscal room that a 1.1 per cent growth forecast provided has been halved along with the forecast.

Tom and Priya don’t read GDP forecasts. They feel them. They feel them in the price of the weekly shop, the energy bill, the interest rate on the credit card they used to cover the gap between Priya’s redundancy and Universal Credit arriving six weeks later. The macro number is an abstraction. The lived experience is a household running on one income in an economy that’s barely growing, with a mortgage designed for two incomes and 1.1 per cent growth.

The GDP cut is not the cause. The cause is a confluence of pressures: AI displacement, energy shock, geopolitical instability, and a housing market built on assumptions that no longer hold. The GDP cut is the measurement that tells you the cushion is thinner than anyone budgeted for.

Fifty per cent of UK organisations have eliminated degree requirements from their job postings. That’s up 28 per cent year on year. Eighty-five per cent now say they use skills-based hiring.

The degree is a proxy. It always was. It told employers that a candidate could complete a three-year programme, write to deadline, absorb structured information, and function within an institutional framework. Those are useful signals. But they are signals about the ability to do orchestration work: synthesise, format, coordinate, report. When the work the exam selects for is the work the technology replaces first, the proxy breaks.

That line is from Chapter 13 of the book. It is, I think, the cleanest version of the argument I’ve managed to put on paper. The degree was never a measure of capability. It was a measure of the ability to do the kind of work that AI now does faster, cheaper, and more consistently than a graduate with a 2:1.

The 85 per cent figure sounds like progress. And in one sense it is. Skills-based hiring, done properly, opens doors that the degree requirement kept shut. It lets people demonstrate what they can actually do rather than where they sat for three years. The research backs this up: 94 per cent of employers using skills-based hiring report that those hires outperform degree-based ones. One case study showed a 340 per cent increase in the qualified applicant pool.

But there’s a catch. “Skills-based hiring” is a broad label that covers a range of practices from rigorous capability assessment to “we removed the degree box on the application form and didn’t replace it with anything.” The latter is not skills-based hiring. It’s credential removal without an alternative signal. And without an alternative signal, hiring defaults to network, familiarity, and gut feel. Which tends to reproduce the same outcomes the degree was supposed to prevent.

The replacement signal matters. This is where I have to declare an interest, because I co-founded a platform that measures it.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Enterprise Skills assesses eight workplace capabilities: commercial awareness, decision-making, problem solving, financial literacy, adaptability, data analysis, team collaboration, and leadership. These are underpinned by four traits that the book identifies as what survives AI automation: judgement, context, relationships, and ethics. The platform currently works in UK secondary schools and colleges, 50-plus schools, 630-plus students. The idea is that if the degree proxy is breaking, something needs to replace it, and that something should measure the capabilities that actually predict performance in a world where the orchestration layer has been automated.

I’m not pretending this is neutral observation. I built one of the tools I’m describing. The builder’s dilemma runs through everything I do: I’m writing a book about displacement and co-founded a company that measures the capabilities that displacement makes more valuable. The honesty is the position.

But the trend is real regardless of my involvement. Half of UK employers didn’t drop the degree requirement because of Enterprise Skills. They dropped it because degrees stopped predicting what they needed to predict. The signal degraded. A 2:1 in business studies tells you roughly as much about someone’s ability to solve a novel commercial problem as their A-level results did. Which is to say: not much.

The alternative credentials market is projected to reach $300 billion globally by 2030. Micro-credentials, digital badges, portfolio assessments, skills passports. The infrastructure is being built. The question is whether it measures the right things or just creates a new set of proxies that will break in their own way.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Chapter 13 of the book goes into the design choices. Recency-weighted scoring, because a capability demonstrated last month matters more than one demonstrated two years ago. Zeros count, because not attempting something is information. Scope qualifier always present, because “leadership” at a school project level and “leadership” at a department level are different claims. These aren’t academic distinctions. They’re the difference between a credentialing system that works and one that becomes the next degree requirement: a box to tick rather than a capability to demonstrate.

Fifty per cent is a tipping point. When half the market has moved, the other half follows or gets left with a shrinking talent pool. The degree isn’t dead. It’s just no longer sufficient. What replaces it will determine whether the skills-based hiring movement opens doors or just changes which doors are shut.

That’s the promise. Universal high income, not the basic sort, the high sort. Work becomes optional. Money becomes, and this is his word, irrelevant. Ten to twenty years and nobody has to earn a living ever again, because the robots earn it for us and we all just, what, collect.

It’s the nicest thing anyone has ever said to a tired person. And most people are tired. Mortgage that went up twice, a job that got heavier while the pay stood still, a pension that’s a rounding error, kids who’ll rent forever. Then a man worth more than some countries leans in and says don’t worry, there’s about to be so much of everything that you’ll never have to do any of this again. Of course it lands. It’s built to land. You’d have to be made of stone for some part of you not to want it to be true.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

I want to sit with that for a second, because I don’t think the people cheering are stupid. I think they’re exhausted. And the pitch is aimed at exactly them, on purpose, because they’ve got the least reason left to defend a system that’s already treated them badly, and the most reason to want the whole thing torn up. Which makes them the perfect crowd to wave through the one thing that finishes the job.

Here’s the part the promise is built to stop you reaching.

Abundance doesn’t fall out of the sky. Somebody owns the factories that make the robots. Somebody owns the compute, the energy, the land it all sits on. “Universal high income” isn’t wealth appearing in your account by magic, it’s a transfer. And a transfer is a gift. A gift has a giver. And a giver, any giver, can stop.

So the only question that actually matters in the whole shiny picture is this: who owns the robots? And notice, the man telling you not to worry about the future is the man who’d own most of them. That isn’t a coincidence you can shrug off. That’s the tell.

Now, I’m not even saying he’s lying about the technology. The robots might well come. The curve is real, I’ve built on it, I’m not here to tell you the machines can’t do the work. They can, and they will, and a big chunk of the coordinating-and-organising jobs that hold the middle class together, three to four million of them in this country alone by the mid-thirties on my own reckoning, are going to thin out whether Elon promises you a penthouse or not.

The lie isn’t the tech. The lie is the word “everyone”.

Because even if you make everything, you can’t make the things people actually fight over. You can print a bigger car. You cannot print a car nobody else is allowed. The good address, the front of the queue, the table by the window, the quiet sense that you’re doing better than the bloke next door: none of that scales, none of it, because its entire value is that someone else hasn’t got it. Total abundance doesn’t kill status. It removes every other way of getting it, so the fight for position gets nastier, not kinder. “Plenty for everyone” is false at the exact level where humans decide whether they feel rich.

And then there’s the bit that should keep you up.

Right now the ordinary person holds precisely one card. One. The right to withhold their labour. Strike, quit, go slow, bargain. It isn’t much, but it’s the lever that built the weekend, the pension, the eight-hour day, every scrap of dignity that wasn’t handed down out of kindness. Take the work away entirely and you don’t free that person. You disarm them. No labour, no strike, no bargaining, no vote that money has to listen to. Just a stipend, set by people they’ll never meet, defended by nobody, switched off by anyone who owns the machine that pays it.

You’ve heard the dystopian version of this. “You’ll own nothing and be happy.” Everyone hated that one, quite rightly. But look at it properly and it’s the same building as the penthouse dream, just photographed from the other side. In both, you own none of the thing that makes the money. The only difference between the paradise and the collar is who’s holding the dial, and how generous they feel that decade. That’s a thin thing to stake a civilisation on. The mood of the people who own the robots.

This is how you get neo-feudalism, and you don’t need a villain or a coup to do it. You just need everyone to agree it’s for the best while it happens. A lord, a stipend, and a lot of people genuinely grateful for the roof, having quietly handed over the one thing that ever made them worth negotiating with. Nobody fights it, because fighting it feels like turning down free money. That’s the engine. The consent is the fuel.

So when you see the abundance post, and you feel that warm little lift, that yes please, that finally, do the thing the post is built to stop you doing. Ask the follow-up. Who owns it. What do I trade to keep my slice. What happens to me on the day the giver decides I’ve had enough.

The tech might be real. The plenty might even be real. Watch one number and ignore the rest. Not GDP, not the robot count, not the price of a fridge. Who owns the robots. Everything else in the picture is set dressing, painted very prettily, by someone who’d quite like you clapping instead of asking.

There’s someone on a team I know who is very good at their job. Ten-odd years in, knows the technical side better than anyone else in the building, the sort of person who will happily spend a day and a half building a spreadsheet that answers a question nobody else knew how to ask. Genuinely good. I want that on the record before I say the rest of it.

Their employer bought the whole team access to one of the AI tools a few weeks back. They haven’t logged in. Every time it comes up, they say they haven’t got the time to learn it.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Then last week they posted in the team chat asking if anyone could help with a piece of work a customer wanted answered by end of day. And they’d laid the whole thing out. Here’s what we suggested, here’s what they came back with, here’s what they’re actually asking for. Fully structured, all the context, nothing missing.

Which is a prompt. That message was already the prompt. Their manager copied it, pasted it into the tool, and had the answer inside a minute.

Their response to that was: right, but now I have to check all of it is correct.

Which is true, and fair, and they should. But notice that it arrived second. The first objection was time. When time got disproved in front of them, the objection became verification. Which reads less like someone worried about accuracy and more like someone quietly moving the fence.

The easy read here is technophobe, or lazy. They’re neither. Their manager can see the usage stats and they’re mid-table for the team, so they’re poking at it. And this is someone who, handed a genuinely clever new thing, would normally be all over it. It’s a sweet shop for a person like them. They should be delighted that the monster spreadsheet doesn’t need building any more.

They aren’t delighted. Which means that something else is going on.

Here’s what I think it is, and it travels a long way outside this particular job. Most people’s work is two things bolted together. There’s the judgement, which is the part you’d describe if someone asked you what you do. Then there’s the production, which is where the hours actually go: pulling the numbers together, formatting them, chasing the right version, making the deck, writing it up, sending it round, and so on. The judgement is real. The judgement is maybe a fifth of the day.

If you sit down with one of these tools for a full afternoon, you find out what that fraction is. Not roughly. Actually. And the reason people don’t sit down with it isn’t the afternoon. They can already feel roughly what the answer is, and a feeling is survivable in a way that a number isn’t.

I’ve seen the louder version of this too. Someone on LinkedIn told me the content from my book was AI slop, which was decent going given the book isn’t out until January. Their profile headline said they cracked PageRank in 2008. PageRank stopped being a thing anyone could crack around 2010. Fifteen years of a professional identity built on a signal that isn’t there any more, and what comes out is anger at whoever points at the ground.

Same fear. Two exits. One shouts, and one is very, very busy.

The busy one is harder to help, because busyness is respectable. Nobody in the history of work has been pulled aside for being too diligent. You can go years flat out on things that quietly stopped mattering, and everyone around you will describe you as hard-working right up until the restructure.

So if any of that landed uncomfortably, the bit worth sitting with is this. The audit isn’t the threat. The audit is the thing that buys you time. Run it in a normal week, still employed, salary landing on the 28th, nothing on fire, and you get to make choices about what happens next. Everyone runs the same audit eventually. Most people run it the week after somebody else has made the decision for them, which is a considerably worse week to be finding things out.

I don’t know what they’ll do. My guess is they’ll be the busiest person on that team for another six or nine months, and then something will shift around them and it’ll be the tool’s fault, or the manager’s, or the market. It usually is.

The question, if you’re going to ask yourself one: last Tuesday, how many of your hours went on the judgement and how many went on producing things. Not a guess. Go and count.

UK job adverts in roles most exposed to AI automation have dropped 38 per cent since 2022. Not layoffs. Not redundancies. Not restructuring announcements. The job that was never posted.

That number doesn’t show up in the unemployment figures. It doesn’t trigger a news cycle. Nobody protests outside an office because a listing didn’t appear on Indeed. But the aggregate effect is structural. Fewer entry points. Fewer lateral moves. Fewer of the roles that used to absorb graduates, career-changers, and people returning from breaks.

The CIPD’s 2026 workforce data makes it worse. One in six UK employers are now planning AI-related layoffs this year. Not considering. Planning. The ones who already acted didn’t call it AI-related. They called it restructuring, optimisation, or nothing at all. They just didn’t post the role.

In Chapter 4 of the book, I introduce Tom. Bristol UWE graduate, 2:1 in marketing, applied to 43 positions over seven months. Six interviews. No offers. What Tom didn’t know, and couldn’t have known, is that one of those agencies had cancelled the role he would have got. The brief went up internally, got discussed, and was pulled when the MD realised the work could be covered by the AI tools they’d adopted three months earlier. Nobody told Tom. There was nothing to tell. The job simply stopped existing between the moment it was conceived and the moment it would have been posted.

Tom is a composite. But the mechanism is not. It is playing out across every sector where orchestration work dominates. Marketing, consulting, financial services, legal support, project management, business analysis. The roles that are fundamentally about gathering information, synthesising it, formatting it, and routing it to someone who makes a decision.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

A 38 per cent drop means roughly four in ten of those roles have quietly disappeared from the market in four years. Some were absorbed into existing headcount. Some were replaced by AI tools. Some were simply deemed unnecessary once the work was reorganised. The net effect is the same: the pipeline that used to move people from education into employment, and from one role to the next, has narrowed by more than a third.

Here’s what makes this different from a normal economic downturn. In a recession, job adverts drop and then recover. The demand is suppressed temporarily. The roles come back when conditions improve. What we’re seeing in the high-exposure categories is not cyclical. It’s structural. The roles are not coming back because the work they described is being done differently. An AI tool that replaces the first-pass analysis doesn’t get unfired when the economy picks up. A workflow that eliminated the coordinator role doesn’t reinstate it because GDP grew half a point.

The 38 per cent is the quiet squeeze measured in aggregate. Individual stories, the Toms and Deepas and Marcuses of Chapter 4, don’t make headlines. But stack 38 per cent fewer entry points on top of one in six employers planning AI cuts, and the picture is clear. The labour market is contracting in the categories that matter most to the professional middle class: the roles that require a degree, pay between £28,000 and £55,000, and involve primarily knowledge-based orchestration work.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The official unemployment rate does not capture this. It counts people who are actively looking for work and not finding it. It does not count the job that was never created. It does not count the graduate who took a retail position and stopped looking. It does not count the contractor whose renewal lapsed. It does not count the career-changer who couldn’t find the lateral entry point that existed two years ago.

Which means the official picture is rosier than the reality by a margin that grows every quarter. The gap between what the statistics show and what people experience is widening. That gap is where political frustration builds. Not because anyone lied, but because the measurement was designed for a labour market that no longer exists.

For anyone currently looking at the job market in a high-exposure field: the competition has not increased. The supply has decreased. Those are different problems with different solutions. If you’re applying to roles that increasingly don’t exist, the answer is not to apply harder. It’s to look at where the openings are actually appearing, which is below the orchestration layer (trades, hands-on, physical presence) or above it (genuine strategy, creative direction, relationship-led work). The middle is where the 38 per cent came from.

Nobody announced this drop. Nobody will. That’s the whole point of a quiet squeeze.

Microsoft offered voluntary buyouts to roughly 7 per cent of its workforce in spring 2026. The programme targeted longer-tenured employees in operations and infrastructure roles. The AI and Copilot teams were explicitly exempt.

Read that again. The company that writes 30 per cent of its code with AI offered packages to the people who keep the lights on, while the people building the thing that replaces the people who keep the lights on were told they’re essential.

This is the two-speed company. It arrived without a press conference.

In Chapter 4 of the book, I describe something I call the quiet squeeze. The mechanism by which displacement happens without anyone saying the word “redundancy.” The vignettes in that chapter are composites, but they all share the same structure: the job doesn’t disappear, it gets restructured, reclassified, made voluntary, offered as a package. The language is soft. The outcome is the same.

Microsoft didn’t fire anyone. They offered a package. Totally different. Except it isn’t.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

A voluntary buyout is a signal dressed as a choice. It says: we think your role is worth less than the cost of keeping you, but we’d rather you make the decision than force us to make it. The people who take the package leave with money and a story about choosing to move on. The people who stay do so knowing the company valued their role at “worth offering money to leave.” Neither outcome is comfortable.

The exemption is the part that matters most. When you exempt the AI teams from a buyout programme, you are drawing a line through the organisation. On one side: the people building what comes next. On the other: everyone else. The operations staff, the infrastructure teams, the support functions. The people whose work is increasingly being absorbed by the tools the exempt teams are building.

Thirty per cent of Microsoft’s code is now written by AI. That number was reported by the company itself. It’s not a projection. It’s a current figure. The developers writing the other 70 per cent are not safe. They’re just not yet at the threshold where the buyout makes financial sense. Give it eighteen months.

Here’s the thing. Microsoft employs roughly 228,000 people. Seven per cent is about 16,000 roles targeted. The company spent $80 billion on AI infrastructure in the fiscal year ending June 2025. The maths is not complicated. When you spend that much building AI capability, the return on investment has to come from somewhere. It comes from headcount. Not all at once. Not with a single announcement. Through a series of voluntary programmes, quiet restructurings, and strategic exemptions that gradually reshape the workforce into something leaner, more automated, and more polarised between the people who build the AI and the people the AI replaces.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The pattern is already visible across the tech sector. The 2024-25 layoffs hit roughly 276,000 across the industry. The hiring that followed was concentrated in AI, machine learning, and infrastructure roles. The net effect is a workforce that’s smaller in total, more expensive per head in the exempt categories, and structurally different from what existed three years ago.

What this means for anyone outside Microsoft is simple enough. The two-speed company is not a Microsoft-specific phenomenon. It’s a template. Every large employer is going to face the same decision: which teams are building the future, and which teams are servicing a structure that the future makes redundant. The exempt list will be short. The package list will be long. And the language will always be voluntary.

Tom, the composite graduate in Chapter 4 of the book, applied to 43 positions and didn’t know the Bristol agency had cancelled the role he’d have got. Nobody told Tom the job was gone because nobody needed to. The job simply wasn’t posted. Microsoft’s buyout is the corporate-scale version of the same mechanism. The roles aren’t being eliminated. They’re being offered a way out. The distinction matters to HR departments and employment lawyers. It does not matter to the person holding the envelope.

The question I keep coming back to: if 30 per cent of the code is already written by AI, and the AI teams are exempt from buyouts, what happens when it’s 50 per cent? What happens when it’s 70? At some point the exempt list and the essential list become the same thing, and everyone else is on the other list.

Microsoft is not the villain here. They’re the company honest enough to put the structure on paper. The rest will do the same thing with less transparency.

In May, KPMG and Anthropic announced a global alliance. Claude, the same model I use to run Orca, is now being deployed across KPMG’s entire 276,000-person workforce. Not a pilot. Not a sandbox. Not a carefully ring-fenced experiment with 200 people and a change management consultant. The whole firm.

This is the Big Four doing it, not talking about it.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The details matter less than the signal. When a professional services firm with 276,000 employees commits to deploying an AI assistant across the organisation, it tells you something about where we are on the adoption curve. We are past the point where enterprise AI is a conference-stage talking point. We are at the point where it is a line item in a global services contract.

Here’s the thing. The 276,000 people inside KPMG will be fine. They’ll get training. They’ll get internal support. They’ll get new workflows. The firm has every incentive to make this work because the productivity gains flow directly to the partnership. The people who should be worried are the ones who aren’t inside KPMG. The mid-level accountants at the next firm down. The auditors at a regional practice that can’t afford an Anthropic alliance. The analysts whose work product is now being generated by the same AI that a quarter of a million of their competitors just got handed.

In the book, Chapter 6 covers what I call critical mass. The point where AI goes from something bolted on to the existing structure to something the structure gets redesigned around. The analogy I use is the gym: adopting AI is buying the membership, redesigning around AI is actually going. KPMG just started going to the gym. Every firm that competes with them now has to decide whether to follow or to pretend the membership card in the drawer is enough.

The competitive dynamics are brutal. A firm running Claude across 276,000 people can reprice engagements. The audit that took a team of six and three weeks can now be scoped for a team of three and ten days. Not because the AI replaces the auditors. Because it replaces the orchestration work that padded the engagement: the data gathering, the cross-referencing, the first-pass analysis, the formatting, the status reporting. The senior judgement stays. The billable hours around it compress.

Which means the firms that don’t follow are competing on a cost structure that no longer reflects the market. That’s not a slow squeeze. That’s a repricing event.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

I built Orca on the same Claude API that KPMG just rolled out. Same underlying model. Same capabilities. The difference is what it’s pointed at. KPMG is deploying Claude to replace tasks. I built Orca to replace roles. The executive assistant, the meeting prepper, the email drafter, the calendar manager, the contact intelligence layer. Not a tool that helps someone do their job. A system that does the job.

That distinction matters because it tells you what’s coming next. Task replacement is where the enterprise starts. Role replacement is where it ends up. KPMG’s 276,000 people will spend the next twelve months discovering that Claude doesn’t just help them do their work faster. It makes certain categories of work unnecessary. The task layer compresses first. Then the role layer follows.

The quiet squeeze that Chapter 4 of the book describes, where displacement happens without anyone using the word “redundancy”, is about to play out inside the world’s largest professional services firms simultaneously. Nobody will get fired because of this announcement. Hiring plans will get revised. Contractor renewals will lapse. Graduate intake numbers will quietly drop. The headcount reduction happens through absence, not action.

For the 276,000, the question is whether they’re the ones using Claude or the ones whose output Claude is replacing. For everyone outside the Big Four, the question is simpler: your competitors just got faster, cheaper, and more consistent. What’s your move?

The adoption curve just kinked upward. The firms that treat this as someone else’s problem will find out it’s theirs about eighteen months from now, when the repriced engagements start winning the work they used to get.

Another one of these tests crossed my feed this week. Will AI Replace You? Ten questions, two minutes, a score at the end. No jargon, no doom, and to be fair to whoever built it, no dishonesty either. The score most people get back is some version of: you’re fine for now, here’s where AI could help you level up.

Which is true. That’s the annoying part. Almost every one of these tests gives an honest answer to the question it was asked.

The problem is the question.

“Will AI replace me” sounds like a yes or no question, so people treat the answer like a yes or no answer. You take the test, you get a decent score, you relax. Job done, quite literally. And nowhere in the process does anyone say the one thing that would actually be useful: come back next week and ask again.

A score like this is a photograph of a moving object. It tells you where the thing was when the shutter clicked. It says nothing about the speed.

And the speed is the story. In March, Anthropic published data on how their AI is actually being used across 800 occupations. Not projections, not expert panels guessing. Observed usage. Computer programmers came out with roughly three quarters of their core tasks already being handled by AI in the wild. Customer service, about 70 per cent. Two years ago those figures would have read as science fiction. Now they’re a dataset on HuggingFace that anyone can download.

The same data holds a detail that should bother anyone with a good score and a degree on the wall. Workers in the most exposed occupations earn 47 per cent more than workers in the least exposed ones. The exposure runs up the ladder, not down it. Which means that the people most likely to take a two-minute test, get a reassuring answer and move on are, on average, exactly the people whose ground is moving fastest.

None of this makes the tests bad, I should say. Any structured look at your own work beats no look at all, and most people have never done even the rough version. My issue is with what happens after the score. Nothing happens after the score. Nobody tells you it expires.

Because it does expire. Any assessment of AI exposure has a shelf life, mine included, because the capability underneath it keeps moving. The models that produced March’s numbers are already not the models people are using in July. A test built in January was measuring a different machine. So the test you took, whichever one it was, wasn’t wrong. It was dated. Those aren’t the same thing, but they end the same way if you treat one score as settled.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

There’s also a better question hiding underneath the quiz question, and it’s less comfortable. Not “could AI do my job” but “will the economics of my employer make it rational to replace what I do within my working life.” That one doesn’t fit in a two-minute test, because it depends on things no test can see: your firm’s margins, your sector’s adoption curve, what your board read on the flight back from the offsite. In the book I spend a chapter on the fuller version of this, the honest audit. The short version of the argument is that knowing is a feeling and an audit is a number, and the distance between the two is where most people hide.

But even the fuller audit is still a photograph. What turns any of this into something useful is repetition. So here’s the discipline, and it costs nothing.

Take whichever test you like. This one, another one, a task list on the back of an envelope, whatever. Write the score down with the date. Put a reminder in your diary for six months from now to take the same test again. When it fires, take it, and compare.

The first score tells you almost nothing. The gap between the first score and the second tells you nearly everything. That gap is your personal velocity reading: not where the technology is, but how fast it’s closing on the specific things you do all day. And then again six months after that, and so on, for as long as you’re working.

If the score doesn’t move, decent. Genuinely. Some work is less exposed for structural reasons, for now, and it’s worth knowing that with dates attached rather than assuming it. If the score does move, you’ve learned something no single test could ever tell you, and you’ve learned it while you still have runway to act on it. Retraining from stability is a fundamentally different exercise from retraining after redundancy. The people who get caught by this shift will mostly be people who got a reassuring score once and framed it.

So take the test. Honestly, take all of them. Just read the result the way you’d read a weather report from March in July.

Not this week is a fine answer. It just has a shelf life, and nobody prints the date on it.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

In May 2023, BT announced that it would cut up to 55,000 jobs by the end of the decade, around 42 per cent of its workforce. Around 10,000 of those cuts were directly attributed to AI and automation. The then-CEO Philip Jansen said the company would emerge “leaner” and better positioned for the future. The share price bounced. The announcement got praised as bold.

Two and a half years later, his successor Allison Kirkby has warned that the number could go higher. EE, the BT-owned mobile operator, runs a virtual assistant called Aimee that handles around 60,000 customer conversations a week. That’s the labour of somewhere between 200 and 400 full-time customer service agents, depending on how you count. Those agents are not being hired. The conversations happen, the customers get answers, the productivity shows up on the operational dashboard, and the headcount that used to be there isn’t.

BT is not a special case. It’s the case, done loudly and in public. The quieter version is happening everywhere, in every sector that has a knowledge work component.

Every time one of these deployments gets covered, the coverage focuses on the three human supervisors still running the system, or the small number of new “AI strategist” or “prompt engineer” roles being created. The implicit argument is that the jobs aren’t really disappearing, they’re just changing. The coverage never gets around to the forty-seven people who used to do the work and now don’t.

This is the book in one paragraph. The deployment maths doesn’t close. The absorption mechanism doesn’t exist. The conversation doesn’t acknowledge either fact.

The cases that are already in the open

Klarna, 2023: roughly 700 customer service roles replaced by AI. Public case study. By 2025, the company is quietly rehiring, but at a smaller number and in less secure roles. Net displacement: probably several hundred.

Duolingo, late 2023: contractor headcount reduced by around 10 per cent, functions absorbed by AI. Translators, copy reviewers, content developers. The learners haven’t noticed any quality change that affects engagement metrics. The displaced contractors didn’t get a press release.

IBM, 2023-2024: Arvind Krishna announced a back-office hiring freeze, with an estimated 7,800 roles targeted for reduction or replacement. HR, finance, administrative functions. By 2025, parts of this were quietly walked back, but the direction is clear: net reduction in administrative knowledge work.

BT, 2023 onwards: 55,000 announced, a meaningful share already executed, more signalled for 2026 and beyond. Customer service, operations, network management.

Amazon, 2024 onwards: multiple rounds of corporate layoffs totalling tens of thousands, some of which the company has explicitly linked to AI-driven efficiency programmes.

Microsoft, 2025: major rounds of corporate cuts alongside aggressive Copilot rollouts internally. Not always framed as AI-caused, but the timing isn’t coincidence.

None of these are controversial or disputed. They’re in the public record. Add them up and you’re already in the low hundreds of thousands of named, announced, AI-attributable displacements in the UK and US, with more in train.

The optimistic framing is that these are offset by new jobs being created elsewhere in the economy. The optimistic framing has one very large problem. Nobody has produced a credible list of the new jobs. Every attempt to do so collapses into “AI will create jobs we can’t yet imagine”, which is a statement of faith, not an argument.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The graduate intake signal

A better leading indicator than named layoffs is what happens at the entry level. If AI is eliminating middle-skill knowledge work, the career ladder into that work gets compressed. You can see this in professional services graduate hiring.

Major UK accounting firms reduced their graduate intake for 2025 entry compared to 2023. Magic circle law firms held intake roughly flat while increasing reliance on AI-assisted document review. Management consulting firms reduced intake substantially, particularly at the strategy end, where AI-augmented partners can cover ground that previously required teams of analysts.

The story for 2026 is still forming, but the early signals suggest the compression is continuing. Law firms are tightening trainee retention rates. Accounting firms are redesigning graduate programmes around AI-literate specialist tracks rather than broad-based training. Consulting firms are quietly restructuring the analyst pyramid into something with a narrower base.

What this means, practically, is that the 22-year-old graduating this summer has fewer places to start than the same graduate would have had five years ago. Not zero places. Fewer. The effect compounds across cohorts. A smaller cohort starting now produces a smaller cohort at mid-career in ten years, because the intake has shrunk permanently rather than temporarily.

This is the mechanism by which the middle layer hollows out. It doesn’t happen through mass layoffs of existing middle managers. It happens through the gradual under-recruitment of the next generation, while the existing generation works with AI assistance until retirement.

The outcome is fewer middle-skill professionals, concentrated among older cohorts, trained in a previous era, doing their jobs with AI tools. The next generation never gets trained in depth, because there aren’t enough entry roles to train them.

What the forty-seven do

If the absorption mechanism is broken, what happens to the people displaced or never hired?

Some retrain successfully into genuinely AI-resistant work. Trades, care roles, specialised professional work that requires physical presence. The pool that makes it through this door is real but small, bounded by physical training capacity and the job market in each destination.

Some accept lower-quality substitute work. Gig economy, part-time, flexible-but-precarious. This is the Klarna hybrid model’s demographic: students, parents, rural workers taking flexible AI-augmented customer service roles for less than they’d have earned in a traditional call centre a decade ago. The roles exist. They don’t substitute for what was lost.

Some exit the workforce. Early retirement if they can afford it. Caring responsibilities taken on formally. Long-term disability that was previously masked by low-demand work. Informal economy participation.

Some stay unemployed or underemployed for extended periods, burning through savings, drawing down equity in homes, borrowing against retirement. This is the category nobody likes to talk about because it’s politically uncomfortable, demographically concentrated among middle-aged middle-class workers, and fiscally expensive.

A small number become loud. This is the category that shows up in political movements, public protest, and the kind of electoral shifts that governments fail to predict. If the Next Rung thesis is right, the loud minority gets larger over the next decade, and the politics get sharper.

The forty-seven don’t evaporate. They distribute across these destinations, unevenly, with consequences that accumulate quietly until they don’t.

The NHS version

Rowan’s conversations about rework in the NHS are the sharpest example of the hidden-cost problem. If rework runs at 30 to 60 per cent of NHS activity, the hidden cost runs into tens of billions a year. This is separate from the Next Rung thesis but connects to it in a specific way.

AI is being aggressively piloted in the NHS for diagnostics, summarisation, administration, and workflow optimisation. The productivity case is being made, loudly. If the deployment follows the pattern this series has traced, gross productivity gains will be reported, rework costs will be absorbed invisibly, and the net benefit will be smaller than claimed. Meanwhile, the 60 per cent rework rate sits there, uncounted, consuming resources that could pay for defence spending twice over.

The combination is worth staring at. We have a public healthcare system running on a 30-60 per cent error rate that nobody measures. We’re deploying AI into that system with productivity claims based on gross output. We’re simultaneously cutting administrative roles on the basis of AI-driven efficiency gains. The same mechanism that gives you the efficiency gains hides the rework cost. The rework cost, if it were measured, would dwarf the efficiency gain.

This is not an AI problem. It’s a measurement problem that AI is making worse. And the people who pay the cost are patients, taxpayers, and the displaced administrators who used to catch errors before they propagated.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

What the argument needs to be

The Next Rung thesis is usually summarised as “AI is eliminating knowledge jobs and the economy can’t absorb the displaced workers”. That’s true, but it’s incomplete. The fuller version includes the rework dimension.

AI is eliminating knowledge jobs on the basis of gross productivity claims that don’t net out. The displaced workers can’t move up the ladder because the ladder has been compressed. The economy can’t absorb them because the mechanisms that used to do the absorbing aren’t operating. The productivity gains that would have made displacement more tolerable are smaller than advertised. The rework cost of AI-augmented workflows will, over time, offset a meaningful share of the announced efficiency gains. The displaced workers will neither get new jobs nor cheaper goods to buy as consumers.

The only group that wins, unambiguously, is the owners of the AI systems and the capital that funds them. Labour loses jobs. Consumers lose quality. Governments lose tax revenue from the displaced workers, and inherit the fiscal cost of supporting them.

This is not a story that resolves itself. It’s a story that compounds, slowly, until the politics catches up with the economics.

The decade ahead

I put the quiet squeeze period through 2028. The breaking point comes between 2028 and 2032. The reshaping follows through the 2030s. These are directional estimates, not forecasts. The specific year depends on the interaction of labour market statistics, political response, and the pace of AI capability gains. All three are uncertain.

What’s not uncertain is the direction. The displaced worker count keeps rising. The offsetting new-job count does not catch up. The productivity gains get reported gross, and the rework costs accumulate quietly. The QC rung doesn’t open at the scale required. The graduate intake compresses. The middle layer hollows out.

At some point, probably before the end of the decade, the politics of this becomes unmanageable through the usual mechanisms. Tax the AI. Regulate the deployment speed. Redistribute the gains. Break up the platforms. Something. The specific policy response depends on which party is in power when the pressure breaks, and which country breaks first. The pressure itself is already building.

The book I’m writing is about the period between now and then. What the mechanism looks like from inside. Why it’s hard to see while it’s happening. What the people who are paying attention should do, both for themselves and for the wider economy.

Rowan’s contribution to this project, which started with a response to the original Next Rung post, has sharpened the argument in a specific way. The rework dimension matters because it breaks the productivity consolation. If the gross gains were real and net, the displacement might still be worth it on utilitarian grounds. If the gains are gross and the net is smaller, the displacement is worse than advertised, and the whole policy calculation changes.

Seven pieces of writing later, including this one, the argument is harder to dismiss than it was when I started. That’s the point of writing a book on something rather than a blog post. You get to see the argument from more angles, against more evidence, with more pressure on the weak points.

The weak points that have survived this process are the ones worth defending. The rework-adjusted displacement thesis has survived. It’s the load-bearing wall for what comes next.

With thanks to Rowan Jackson, whose feedback on the original Next Rung post catalysed this series, and whose ERW paper, published in December 2024, supplied the analytical framework that runs through all six pieces. The orthopaedic consultant’s 60 per cent NHS rework estimate in particular is worth sitting with. It may be the most consequential back-of-envelope figure I’ve encountered since starting this project.

Primary sources: BT strategic update (2023) and subsequent AGM commentary (2024-2025); Klarna public disclosures and press coverage, 2023-2025; IBM back-office automation disclosures, 2023-2024; UK Office for National Statistics labour market data through Q4 2025; law and accounting firm graduate intake reporting, 2024-2025; Rowan Jackson, ERW paper, December 2024; NHS England financial directions 2025-26; H.J. Harrington, Business Process Improvement, 1991; HBR, Beyond Toyota: How to Root Out Waste and Pursue Perfection (1996); HBR case study on Danaher Business System.

On 1 August 2012, Knight Capital Group deployed a software update to its trading system. The update contained a bug. The system began executing trades at a rate of several thousand per second, in patterns that had no connection to the strategy it was meant to run. By the time someone pulled the plug, 45 minutes later, the firm had lost around $440 million. Knight Capital, which had been a top-tier market maker on the New York Stock Exchange, had to be rescued within days. It never recovered its independence.

The bug wasn’t catastrophic on its own. It was catastrophic because it executed at speed. A slower system would have flagged anomalous behaviour before the losses compounded. A human trader making bad decisions would have run out of positions or capital within minutes. The software, operating at machine speed with machine confidence, did enough damage to bankrupt the firm before the control mechanisms could respond.

This is what “faster means worse” looks like in a real system. Speed without correction is not a productivity gain. It’s a risk multiplier. Every AI deployment that operates faster than its verification layer is running the same experiment, usually without realising it.

The conditions that make speed dangerous

Speed is helpful in systems that have three properties. Errors are local and bounded. Feedback is fast. Correction is cheap. A calculator is a good example. It runs at speed, any mistake affects one computation, you see the result immediately, you can redo it in seconds if it’s wrong.

Speed is dangerous in systems where any of those three conditions fails. In a trading system, errors propagate across positions and counterparties. In a content system, errors reach audiences and get repeated. In a clinical system, errors affect patients and may not become visible for weeks. In a legal system, errors get signed into contracts and only surface during disputes. The faster the production, the more error is produced before anyone notices.

Rowan Jackson’s framework on errors, rework, and waste makes the point directly. “Technology does not replace human error. It just makes it faster. In some cases more consequential.” The second sentence is the important one. Most of the damaging deployments of technology aren’t failing because the error rate went up. They’re failing because the error rate stayed roughly the same while the production rate went up an order of magnitude. The absolute volume of error rises with volume. Beyond a certain threshold, the verification layer can’t keep up.

AI sits exactly at the point where this threshold is being crossed in domain after domain.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The 737 MAX, a different kind of speed

Boeing’s 737 MAX is a different shape of the same problem. It wasn’t a trading algorithm running too fast. It was a development programme running too fast. Boeing was under commercial pressure to match Airbus’s A320neo, and it compressed the development and certification timeline on an aircraft that, to avoid expensive re-engineering, was fitted with a flight control system called MCAS that pushed the nose down in certain high-angle conditions.

MCAS relied on a single angle-of-attack sensor. Pilots weren’t told it existed in most training materials. When the sensor failed, the system activated without warning, and the pilots didn’t know what was happening or how to disable it. Two crashes, 346 deaths, the aircraft grounded globally for almost two years.

The speed that caused the 737 MAX disaster wasn’t operational speed. It was decision speed. Boeing moved faster than its safety culture could verify. The regulator moved faster than its own inspection protocols. The airlines moved faster than their training programmes. Every individual step looked reasonable in isolation. The system as a whole outran its own verification capacity.

This is the shape of the AI deployment risk right now. Not that any individual AI system is catastrophically broken. That the collective deployment pace exceeds the capacity of regulators, auditors, firms, and users to verify what the systems are actually doing. The analogous failure is not a single chatbot telling a single customer the wrong thing. It’s the distributed, cumulative effect of thousands of AI-driven decisions that each look defensible and collectively produce systemic outcomes nobody chose.

The flash crash class

Financial markets have given us a library of “faster means worse” failures. The 6 May 2010 flash crash in US equities wiped almost a trillion dollars off US stock values in minutes before the market recovered. The 15 January 2015 Swiss franc move, when the Swiss National Bank removed its peg, destroyed several retail forex brokers and caused tens of millions in losses for their clients before anyone could react. The 23 September 2022 sterling crash during the Truss mini-budget was partly driven by automated deleveraging in pension funds’ liability-driven investment strategies, which sold gilts into a falling market and triggered further falls.

Each of these events shares the same structure. Automated systems operating at speed, making decisions that are individually rational but collectively destabilising, running faster than the humans who designed the systems could intervene. The systems don’t fail because they’re broken. They fail because they’re working exactly as designed in a market state the designers didn’t anticipate.

AI is now being deployed into an economy full of systems with this property. Supply chains. Pricing engines. Underwriting. Content recommendation. Fraud detection. Each system individually looks fine. Each system makes decisions at speed. Each system is supervised by humans who were told the AI would help them make better decisions faster.

The question nobody is asking is what happens when several of these systems interact in ways their designers didn’t model. The history of financial markets suggests the answer is: occasional catastrophic failures that nobody planned for and nobody can stop in time.

Normal accident theory, briefly

Charles Perrow published Normal Accidents in 1984, analysing why complex, tightly-coupled systems produce catastrophic failures with a regularity that surprises their designers. The argument is that in systems with many interacting components, sooner or later some combination of events will occur that nobody modelled. The failure looks bizarre after the fact, but the probability of some failure of some shape is close to one, given enough time.

Three Mile Island, Chernobyl, the Challenger explosion, the Deepwater Horizon spill. These are the textbook cases. Perrow’s point is that you can’t fully prevent them through better engineering. The complexity itself generates failure modes that outpace the ability to anticipate them.

AI deployment is rapidly producing systems that match Perrow’s criteria. Many components. Tight coupling between them. Operators who don’t fully understand the internal logic. Feedback loops that aren’t visible until they trigger. Decision speeds that exceed human intervention capacity. The conditions for a normal accident are in place. The question is what the specific form looks like.

It probably won’t be a Skynet scenario. It’s more likely to be a sequence of compounding errors across an industry, at a time nobody anticipated, producing a systemic outcome that looks obvious in hindsight and was unpredictable in prospect. The 2010 flash crash is a better model than the AI-apocalypse fiction. Something goes wrong, the system amplifies it, by the time humans intervene there’s substantial damage.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The ordinary version

This isn’t only about rare catastrophic events. The more common version is the steady, compounding degradation of quality in systems where AI deployment has outrun the verification layer.

Every piece of AI-generated marketing content that goes out without proper review is a minor instance of this. Every AI-drafted legal clause that slides through a busy deal cycle. Every AI-summarised medical note. Every AI-assisted underwriting decision. The individual errors are small. The aggregate effect is a lowering of the quality of everything the AI touches, slowly enough that it doesn’t trigger alarms.

Rowan’s 30 per cent ERW figure for construction is the analogue. In construction, rework has been running at around 20-30 per cent of project value for years. Nobody treats it as a crisis, because the cost is distributed and absorbed. The buildings still get built, eventually, at higher cost than advertised. The industry operates at a persistent productivity discount to what it could achieve.

AI is about to create the equivalent ERW overhang in every sector it touches. The costs will be absorbed in cost of sales, in customer complaints, in regulatory fines, in lost contracts, in the slow erosion of institutional credibility. No single event will trigger a reckoning. The cost will just sit there, spread across the economy, invisible on any individual balance sheet.

The NHS orthopaedic consultant Rowan spoke to put clinical rework at 60 per cent. That’s not a scandal that breaks. It’s a background rate that persists. AI deployment into clinical workflows, without proper verification, will push that number up before it pushes it down.

What “fast” should actually mean

Speed is useful when it’s matched to the system’s capacity to verify and correct. The virtue of a calculator isn’t that it’s fast. It’s that it’s fast and accurate and the user can check the output in the same cognitive register. The virtue of a well-designed automation isn’t speed either. It’s speed paired with observability, auditability, and a correction mechanism that runs at comparable speed to the production mechanism.

AI systems, as currently deployed, routinely fail this pairing test. The production speed is high. The verification speed is human-paced, because verification requires substantive understanding and that’s slow. The gap between production and verification is where the errors live.

A useful framing for any AI deployment: what’s the ratio of production speed to verification speed? If it’s 10:1 or higher, the deployment is running with a verification deficit. Either you accept the error rate and manage the consequences, or you invest in substantive human review and lose the speed advantage, or you build automated verification that operates at comparable speed to production.

Most deployments take option one, quietly, because options two and three are expensive. This is the setting in which faster means worse. Not always dramatically. Sometimes catastrophically, as in Knight Capital or 737 MAX. Usually just as persistent quality degradation, absorbed as the cost of doing business, showing up later in places the accounting doesn’t track.

Back to the Next Rung

Speed is the economic reason AI deployments displace workers. If the AI weren’t faster, the displacement wouldn’t happen. The productivity case depends entirely on the speed advantage.

The piece I want to keep making is that speed without verification is not productivity. It’s liability, booked to future periods. The displacement happens now. The rework cost comes later. The gap between them is where the optimists live, and where the workers lose their jobs.

If you want a clean framing for any AI deployment decision, it’s this: are we moving faster than we can verify? If the answer is yes, the deployment isn’t solving a problem. It’s storing one up. And in the meantime, it’s eliminating the middle-skill jobs that would have provided the verification if we’d kept them.

The Next Rung argument doesn’t require AI to fail catastrophically to land. It only requires AI to do what it’s clearly already doing: displace the middle layer of workers faster than the economy can absorb them, while producing output faster than the quality layer can verify it. Both effects compound. Both effects happen quietly. Both effects will eventually surface in ways nobody currently wants to talk about.

With thanks to Rowan Jackson, whose framing that “technology does not replace human error, it just makes it faster, in some cases more consequential” sits at the centre of this argument.

Primary sources: SEC report on the Knight Capital incident (2013); US House Committee on Transportation & Infrastructure, The Design, Development & Certification of the Boeing 737 MAX (2020); SEC-CFTC joint report on the 6 May 2010 flash crash; Bank of England reporting on the September 2022 gilt market events; Charles Perrow, Normal Accidents: Living with High-Risk Technologies (1984); Rowan Jackson, ERW paper, December 2024.

The most common response I get when I argue that AI is eliminating knowledge jobs faster than the economy can absorb the displaced workers is this: humans will just review the AI output. There’ll be a huge new category of AI auditors. Quality control becomes the new knowledge work. The next rung still exists, it’s just moved.

This is a comforting answer. It’s also wrong, and the arithmetic is not subtle.

A team of three human auditors can oversee the output of fifty AI agents. This is roughly the ratio you see in live deployments in customer service, content moderation, and document review. It’s the operational shape that makes the AI deployment economically attractive in the first place. If you needed one auditor per AI agent, you’d have saved nothing by deploying the AI. You’d just have shifted one human job to a different human job.

So the ratio is real, and the ratio is the point. If you displace fifty workers and need three auditors, you have forty-seven people without work. That’s not a transition to a new rung. That’s a net removal of forty-seven rungs.

This isn’t a hypothetical. It’s already visible in the companies that have rolled AI out at scale.

What the Klarna case actually shows

Klarna’s AI customer service deployment in 2023 was pitched as a case study in AI productivity. The company replaced around 700 customer service positions with an AI assistant. By 2024, Klarna was claiming the AI handled 75 per cent of all customer chats and did the equivalent work of those 700 people.

By spring 2025, Klarna was hiring customer service agents back. The CEO publicly admitted the quality was insufficient. The company moved to a hybrid model where AI handles routine queries and humans handle complex ones.

Here’s the part worth dwelling on. Klarna’s hybrid model is what the “QC rung” argument predicts. AI at the front, humans at the back, quality control distributed between them. It is also, measurably, not a return to the original headcount. The 700 displaced workers did not come back as 700 auditors. Some number came back, somewhere in the low hundreds based on reported hiring, to handle the residual work the AI couldn’t do. The rest are still displaced.

And the hiring Klarna has done is different in kind from what was lost. The new roles are targeted at flexible, part-time workers, remote, often students or parents. The original 700 were full-time salaried positions with career progression. The new structure is cheaper per head, less secure, and carries less institutional investment.

This is the QC rung in practice. It exists, but it’s narrower, lower-paid, and less stable than the work it replaced.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The content moderation precedent

The cleanest historical precedent for AI-augmented human quality control is content moderation at the big platforms. Facebook, YouTube, TikTok, and their peers deploy AI to triage posts, flag likely violations, and surface edge cases for human review. The humans make the final call on the most ambiguous content.

Meta, at its peak, employed around 15,000 content moderators globally, many through contractors like Accenture and Cognizant. The platform processes billions of pieces of content a day. The moderator-to-content ratio is something like one moderator per hundred thousand pieces of content. The vast majority of moderation is done by AI. Humans exist in the loop for the hardest cases, policy calibration, and training signal generation.

This is the shape AI quality control takes when it’s operating at scale. A small human cadre maintains the system, handles the exceptions, and generates the feedback that trains the next version. The headcount ratio is not one auditor per AI agent. It’s one human per thousands of AI interactions, or tens of thousands, depending on the domain.

And content moderation is a comparatively labour-intensive case, because the subjective calls are genuinely hard and the legal exposure is genuinely high. In domains where the AI is generating rather than judging, the ratio gets worse for the humans. A generative AI producing legal briefs can be reviewed by a human at roughly the rate a human could write them, because the review has to be substantive to catch hallucinations. But the AI produces them at ten times the speed. The math doesn’t give you ten auditors. It gives you one auditor checking every tenth brief, or reviewing all of them at reduced rigour, or rejecting the deployment as unsafe for the context.

Why the ratio can’t grow

The optimistic version of the QC rung argument assumes that as AI deployments scale, the auditor cadre scales with them. More AI means more auditors. New category of work, similar in shape to the old category, just shifted.

This doesn’t hold for three reasons.

First, the economic case for deploying AI depends on the ratio. If an AI deployment requires one auditor per AI, it’s simply a more complicated way of employing the person who used to do the job. Nobody invests in that. The deployments that happen are the ones where the ratio is heavily in AI’s favour. Which means the jobs that come back in the audit role are a small fraction of the jobs that went away in the production role.

Second, the audit role is itself being automated. Once enough human audit data accumulates, the system learns to flag its own low-confidence outputs, self-correct, and reduce the proportion requiring human review. The audit ratio degrades over time. Whatever QC rung exists in year one is narrower in year three.

Third, the audit work doesn’t scale geographically or demographically the way the production work did. A junior marketing manager in Leeds, a paralegal in Manchester, a business analyst in Cardiff, these were jobs distributed across the country, across demographics, across education levels, providing genuine middle-skill employment. The audit role concentrates. A small number of specialists in a small number of locations handle the quality control for work that used to employ tens of thousands. The distribution is gone.

This last point is the one that’s hardest to see from inside the productivity conversation. The Next Rung thesis isn’t just about net numbers. It’s about what happens to the geographic and social distribution of middle-skill work when the middle layer collapses. The audit rung, even if it existed at the scale of the lost rung, wouldn’t solve the distribution problem. And it doesn’t exist at the scale of the lost rung.

The error-rate problem

There’s a second reason the QC rung doesn’t solve the displacement question, and it comes directly from Rowan Jackson’s framework on errors, rework, and waste.

If AI output contains errors that look like competence, as I argued in the last piece, then the audit process has to be substantive. You can’t skim AI output and catch the fabricated citation or the contradictory clause. You have to read, check, verify, cross-reference. This is roughly the same work as producing the output from scratch.

Which means the audit role, where it exists, is not a cheaper version of the production role. It’s the same work, wearing different clothes. If the economics worked at one auditor per AI, it would also work at one producer per producer, which is where we started. The deployment only makes economic sense if the audit is done cheaply, which means superficially, which means the errors get through.

There are two outcomes from this. Either organisations accept the error rate and push the cost downstream to customers and counterparties. Or they invest in substantive audit and lose the productivity gain that justified the deployment. In practice, most organisations do the first, quietly. This is why the rework ledger matters: without it, the cost of bad AI output gets absorbed in places the accounts don’t show.

Rowan’s ERW work puts this cost at 20 to 45 per cent of operating expenses in industries where rework has been studied. In industries where rework is hidden in AI-augmented workflows, nobody knows what the number is. The plausible range is the same or higher. That’s a lot of invisible cost to be accumulating while reporting gross productivity gains.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The case of the legal paralegal

Consider what’s happening in law right now. Large firms are deploying AI-assisted contract review, legal research, and document drafting at scale. Allen & Overy rolled out Harvey in 2023. Clifford Chance followed. A number of US firms have built internal systems or licensed external ones. The productivity claims are substantial. Some firms are reporting that junior tasks which used to take two days now take two hours.

The QC rung prediction would be that paralegals and junior associates move into AI audit roles, reviewing the output, catching the errors, adding the legal judgment the AI can’t apply.

What’s actually happening is a reduction in paralegal headcount and a tightening of junior associate intake. The work that used to be done by two paralegals and a junior is now done by one senior associate with AI assistance. The audit is happening, but it’s being done by people senior enough that their time is too expensive to call an audit role. It’s just part of the senior job. The junior roles are gone, and with them the career pipeline that would have produced the next generation of senior associates.

This is not a new category of audit work. It’s a compression of the existing career ladder, with the bottom rungs removed. The AI didn’t create a new rung. It removed two.

What the audit case actually produces

When AI is deployed well, with substantive human oversight, what it actually produces is not a new audit profession. It produces a small number of highly skilled human roles at the top of the output chain, serving a much wider production base than was previously possible. The senior associate reviewing AI-generated briefs can now service three times as many matters. The senior editor reviewing AI-generated articles can now oversee ten times the output. The senior analyst reviewing AI-generated reports can cover a portfolio that used to require a team.

This is genuinely productivity-enhancing at the senior level. It is also genuinely displacement-causing at the middle and junior levels. The senior role doesn’t scale with the production increase. The junior and middle roles get eliminated in proportion to the production increase.

The net effect is fewer jobs, paid better, concentrated in senior specialists. This is good for the specialists who keep their jobs. It is catastrophic for the people who would previously have been junior specialists working their way up.

The book I’m writing, The Next Rung, is fundamentally about this collapse. The middle of the ladder is where the rungs are being removed. The top of the ladder continues to function. The people already on the top rung are fine. The people on the bottom rung have a hard time getting onto the second rung, because the second rung doesn’t exist any more.

The policy version

At the policy level, the QC rung argument is being made to justify doing nothing. If the market will naturally create new audit roles, there’s no need for industrial strategy, no need for transitional support, no need for retraining investment. The market will handle it.

This is the “it was fine last time” argument in a new coat. It assumes a mechanism that is demonstrably not operating at the scale required. It asks us to bet millions of livelihoods on a ratio that the economics explicitly forbid.

The honest version of the argument would accept that AI deployment produces fewer jobs than it displaces, and ask what we do about that. The dishonest version insists on a QC rung that the deployment data doesn’t support, because accepting the honest version means accepting that we need policy, and nobody wants that conversation.

What to actually do

If you’re an individual worker thinking about career positioning, the QC rung is a trap. Don’t plan your career around being the human auditor of AI output, unless you’re already senior enough to command the rates that make the role economically defensible. The junior and middle audit roles won’t sustain.

If you’re an employer thinking about deployment, be honest about the headcount arithmetic. If your AI deployment case assumes you’ll create quality control jobs to replace the production jobs you’re displacing, check your numbers. The ratio won’t support the story. And the invisible rework cost will catch up with you either way.

If you’re a policy maker, accept that the transition mechanism you’re relying on doesn’t exist at the scale you need. Previous transitions absorbed displaced workers because the new industries needed bodies. AI deployments specifically don’t need bodies. They need a small number of expert supervisors and otherwise they’re designed to operate without human labour. This is the whole point of them. Expecting them to generate employment is expecting the product to do the opposite of what it was built for.

The next rung isn’t being built. The QC rung doesn’t exist at the scale required. The question of what to do about that is the question this book is about.

With thanks to Rowan Jackson, whose ERW framework continues to inform the quality-and-rework dimension of this analysis, and whose observations about the limits of technology deployed onto broken processes (notably, “I have seen Salesforce destroy a company’s sales system”) echo the same structural point from a different direction.

Primary sources: Klarna public statements and press coverage, 2023-2025; Meta, YouTube, and TikTok content moderation staffing disclosures; industry coverage of Harvey (Allen & Overy) and legal AI deployments; ONS labour market data to early 2026; Rowan Jackson, ERW paper, December 2024; H.J. Harrington, Business Process Improvement, 1991.

In May 2025, the High Court of England and Wales ruled on a case called Ayinde v London Borough of Haringey. The headline wasn’t the dispute. The headline was that a witness statement in the proceedings contained 18 fabricated legal citations out of 45. The citations looked real. They had case names, court details, citation formats that matched proper legal conventions. They just didn’t exist. An AI had made them up, and the person submitting the statement hadn’t checked.

By the end of November 2025, the UK had recorded 24 separate instances of AI-fabricated legal citations ending up in court proceedings. The global total was over 600. In July 2025 alone, more than 50 such cases were publicly reported across jurisdictions.

Judges have started imposing costs orders, referring solicitors to regulators, and in some cases warning of criminal prosecution for contempt. The Bar is producing guidance documents. Law societies are scrambling. None of this is solving the core problem, which is not that lawyers are using AI. It’s that AI is producing output that looks exactly right when it’s completely wrong, and humans cannot reliably detect the difference at speed.

This is the distinguishing feature of AI error. It’s not that errors happen faster. It’s that they look like competence.

Why this is different from previous technology errors

Every tool humans have ever invented has produced errors. Calculators get entered wrong. Spreadsheets get formula errors. Word processors have spell-check limits. Automation in manufacturing introduces systematic defects when the setup is wrong.

What’s different about AI error is the combination of three things that previously only occurred one at a time.

First, AI errors are fluent. A calculator error returns the wrong number. You can tell it’s a number. You can check it against what you expected. An AI error produces a fluent paragraph, a credible-looking citation, a recommendation that matches the surface pattern of a good recommendation. The error is wrapped in the linguistic signals of competence.

Second, AI errors are confident. There’s no uncertainty signal. The system produces the wrong answer with the same tone it produces the right answer. In earlier automation, a lot of the signalling came from hesitation, partial completion, or visible edge cases. The machine would stop, throw an error, flag the anomaly. AI systems don’t stop. They produce, and they produce fluently.

Third, AI errors are at scale. One person can generate the volume of output that used to require a team. The rate of production outruns the rate at which any human can plausibly check the work. Even if the error rate per output is low, the absolute volume of errors is high, and the proportion that escapes detection rises as volume rises.

Rowan Jackson’s framework for errors, rework, and waste identifies the same pattern in industrial settings. When processes produce outputs faster than the quality system can verify them, errors accumulate and propagate. In a factory, the errors eventually become visible because the product is physical. In knowledge work, the errors stay invisible until someone downstream acts on them and something breaks.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The bricklayer analogy

Rowan used this example in correspondence with me, and it’s worth sitting with. If a bricklayer places a wall five centimetres out of alignment, the error is immediate and physical. Someone walks past, notices, measures. The wall has to come down and be rebuilt. The cost is painful but bounded. You pay for the bricks twice, the labour twice, the time twice.

If an AI generates a legal clause that contradicts another clause in the same contract, the error is invisible. The document looks well-formed. The clauses read clearly on their own. The contradiction only surfaces when something triggers the two clauses at the same time, which might be years after signing, and might involve a dispute where the cost is orders of magnitude larger than the original drafting fee.

Rowan mentioned directly that his team has corrected clients’ contracts that had clauses directly contradicting each other. These weren’t AI-generated. They were the result of humans copy-pasting from templates without reading carefully, which is the same failure mode AI is now industrialising. The human version was rare. The AI version will be common, because the production speed is higher and the verification pressure is lower.

The cost of the bricklayer’s five-centimetre error is visible in the rework budget. The cost of the contract error is hidden until it activates. By the time it activates, the person who drafted it has moved on, the firm that employed them has billed the fee, and the client is the one holding the liability. This is not a technical problem. It’s a structural one.

The Air Canada precedent

In February 2024, the British Columbia Civil Resolution Tribunal ruled in Moffatt v Air Canada. The facts are simple. A customer asked Air Canada’s website chatbot about bereavement fares for a grandmother’s funeral. The chatbot gave incorrect information. The customer relied on it and booked a full-fare ticket. Air Canada then refused to honour the bereavement rate the chatbot had promised.

Air Canada’s defence was remarkable. It argued that the chatbot was a separate entity whose misrepresentations were not the airline’s responsibility. The tribunal rejected this, ruled the airline liable, and awarded around $650 plus fees.

The damages are trivial. The precedent is not. The case established that a company is liable for the misrepresentations made by its AI, and that “the AI said it” is not a defence. Every deployment of customer-facing AI since Moffatt sits in this legal context, whether the deploying company realises it or not.

More importantly, the case demonstrated that a fluent-sounding AI error can produce a legally binding commitment. The customer relied on it. The tribunal found reasonable reliance was warranted because the information appeared competent. The fluency was the problem. If the chatbot had stammered or said “I don’t know”, the customer would have checked. Because the chatbot sounded like it knew, the customer acted on it.

Multiply this across financial services, insurance, healthcare, legal services, government. Every AI interface that produces output in natural language is now a potential source of binding commitments by the deploying organisation, whether the output is accurate or not.

The detectability gap

Here’s where the problem gets structural. Consider how humans check each other’s work.

A junior associate’s contract draft is reviewed by a senior associate. The senior associate has years of pattern recognition. She’s seen enough bad drafts to know what a weak clause looks like, what a common miswording is, what pattern of structure suggests the junior didn’t understand the underlying transaction. The review is possible because bad human work tends to carry signals of its badness, at least to a trained reader.

A junior associate’s AI-generated contract draft, handed to the same senior associate, carries different signals. The clauses are grammatically clean. The structure matches template conventions. The terminology is consistent. The language signals of bad work are absent. What’s left is substantive error, which requires substantive review. That takes approximately the same time as writing the contract from scratch.

This is the core problem with quality control at scale for AI output. The surface signals that used to help reviewers triage bad work have been erased by the AI. Every piece of output now demands substantive checking, because the superficial checks no longer discriminate. Either reviewers slow down to do substantive review on everything, in which case the productivity gain evaporates, or they speed up and let substantive errors through, in which case the rework cost shows up later somewhere else.

There’s no middle path. The old quality control system was calibrated to a world where poor work looked poor. That world is gone.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The hallucination metric problem

Vendors of AI systems report hallucination rates. OpenAI, Anthropic, and Google all publish benchmarks. The numbers have been coming down. State-of-the-art systems in 2025 are reporting single-digit hallucination rates on standard evaluation sets.

This sounds reassuring. It isn’t.

A single-digit hallucination rate means roughly one in every 15 to 25 outputs contains a fabrication. If you’re producing legal documents, medical summaries, or financial analyses at a rate of hundreds per week, you’re producing dozens of hallucinated outputs per week. Even if each one has only a small probability of causing material harm, the portfolio-level risk is substantial.

More importantly, benchmark hallucination rates are measured on evaluation tasks where the ground truth is known. The hallucination rate on novel, open-ended, commercially relevant tasks is generally higher and is not measured in any systematic way. The vendor reports the benchmark number. The user acts on the novel task. The gap between the two is invisible.

And the benchmark itself is a lagging indicator. Every time a category of hallucination becomes publicly embarrassing, the next model version is tuned to address it. This creates the impression that hallucinations are being solved, when what’s actually happening is that the hallucinations are being redistributed into categories that haven’t yet been measured.

What this does to the next rung argument

The original Next Rung thesis argued that AI is automating cognitive work across the board, leaving displaced workers without a clear next job to move into. One of the standard objections to this argument is that humans will still be needed for quality control. Humans will review the AI output, catch the errors, ensure quality. The displaced workers can become AI auditors.

This objection looks much weaker once you understand the detectability gap. AI errors don’t announce themselves. They require substantive expert review to catch, which takes roughly the same time as doing the original work. You don’t get a productivity gain if you have to substantively review every output. You just shift where the labour happens.

And substantive review scales worse than production. A senior lawyer can draft one contract at a time. She can’t review ten AI-generated contracts at a time with the same rigour. The production side scales with compute. The review side scales with human expertise, which is rare, expensive, and slow to develop.

The practical outcome is not a new category of AI auditor jobs. It’s a quality degradation spread across the economy, where the output volume goes up and the average quality goes down, with the losses absorbed by customers, patients, clients, and downstream processes that can’t detect the errors until they’ve already acted on them.

What to do

For any organisation deploying AI into a workflow where errors matter, three actions are non-negotiable.

First, assume the AI is producing some proportion of fluent-but-wrong outputs, and design the review process around substantive verification rather than surface checking. The days of skim-reading junior output are over. Every piece of AI-generated work that goes external needs line-by-line review by someone qualified to spot substantive error.

Second, track error rates downstream, not just at the point of production. If your customer service AI is giving wrong answers, the metric that matters is customer complaint volume and churn, not the chatbot’s internal confidence score. The error rate at the point of production is often low. The error rate at the point of consequence is higher.

Third, accept that the productivity gain from AI in high-stakes work is smaller than the vendor claims, because most of the saved time on production will be spent on verification. If you’re going to deploy AI in legal, medical, financial, or safety-critical contexts, budget for the verification overhead from day one. If the business case doesn’t work once you include it, the deployment probably shouldn’t happen.

The speed is real. The scale is real. The quality signalling, for the first time in the history of technology, has been decoupled from the quality itself. That’s what’s new. That’s what matters.

With thanks to Rowan Jackson, whose observation that contract drafting errors are a particularly hidden category of rework, and whose example of clauses that directly contradict each other within the same document, sharpened the legal section of this piece.

Primary sources: Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin); Moffatt v Air Canada, 2024 BCCRT 149; Damien Charlotin’s AI Hallucination Cases Database; the Bar of England and Wales guidance on AI use, 2025; Vectara hallucination leaderboard; Matthew Syed, Black Box Thinking.

In July 2025, a research lab called METR published the result of a randomised controlled trial on AI coding tools. The setup was rigorous. Sixteen experienced open-source developers, average five years of experience on the mature codebases they were working on, completing 246 real tasks, screens recorded, timings verified. They could use any AI tools they wanted on half the tasks. On the other half, no AI. Same people, same projects, same work.

Before the trial, the developers predicted AI would make them 24 per cent faster. After completing the tasks, they reported feeling 20 per cent faster. The actual measurement showed they were 19 per cent slower.

Forty-four points of gap between what they believed was happening and what was happening. Not because they were lying. Because they couldn’t feel it. The rework was distributed across the day, the reading of AI output, the correction of AI output, the debugging of AI output, the re-prompting when the first attempt was off. Each of those micro-moments felt productive. Stacked up, they were net-negative.

That’s the productivity mirage. You can’t see it from the inside.

The numbers you’ve been quoted

McKinsey: AI could automate 30 per cent of US work hours by 2030. GitHub’s own research: Copilot makes developers 55 per cent faster. Goldman Sachs: generative AI could raise global GDP by seven per cent. Klarna: its AI assistant handled 2.3 million customer conversations, doing the work of 700 full-time agents.

Every one of those numbers is gross. None of them is net of rework, correction, quality loss, or downstream repair.

The GitHub 55 per cent figure came from a controlled experiment where developers wrote a specific HTTP server function. The task was narrow, well-defined, and closely aligned to what Copilot had been trained on. The METR number is from real work in mature codebases where the developer has to hold the existing architecture in their head. Both numbers are real. They just measure different things.

Guess which one gets quoted in board papers.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

What gross means in a P&L

Imagine you run a marketing team of ten people producing 100 pieces of content a month. You roll out an AI tool. Output doubles to 200 pieces a month. The productivity dashboard shows a 100 per cent gain. Leadership reallocates the team, lets three people go, congratulates itself.

Now count what actually happened. Of the 200 pieces, 60 needed substantial rewriting because the AI output was bland, factually off, or missed the brand voice. That’s 60 pieces at a couple of hours each, call it 120 hours of rework that didn’t exist before. Another 20 pieces got published with minor issues that nobody caught until a customer pointed them out. That’s customer trust costs, hard to measure, paid in tiny increments over months. A further 30 pieces got deployed in campaigns where the performance data came back 15 per cent below prior benchmarks, because the work was technically complete but commercially weak.

The productivity dashboard still shows a 100 per cent gain. The rework sits in a different budget line. The customer complaints sit in CX. The campaign underperformance sits in media. The three people who got let go sit on LinkedIn.

The CFO sees the headline number. The CMO’s bonus is paid against the headline number. The AI vendor’s case study uses the headline number. Nobody is looking at the rework ledger, because nobody set one up.

This isn’t a thought experiment. It’s the Klarna story.

The Klarna case, in detail

Klarna replaced around 700 customer service staff with an AI assistant in 2023. By February 2024, the AI was handling 75 per cent of customer chats, around 2.3 million conversations, in 35 languages. The company publicised cost savings of around $40 million annually. It was used as a poster child for AI-driven cost cutting. The CEO, Sebastian Siemiatkowski, spoke at conferences. Stock analysts cited it.

By spring 2025, Klarna was hiring customer service agents back. The reason, in Siemiatkowski’s own words: “We focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.”

Translated out of CEO-speak: the AI deployment saved money on the headcount line and cost money everywhere else. Customer complaints went up. Satisfaction scores went down. Complex issues didn’t get resolved. Customers who needed a human got a loop of generic AI responses. Some of them churned. The lifetime value loss from those churned customers almost certainly exceeded the saved salary budget, though nobody has published that number because nobody is forced to.

Klarna’s current model is hybrid: AI for routine, humans for escalation and complexity. This is the right model. It’s also, notably, more expensive than the original AI-only model and less profitable than just having humans who were trained properly in the first place. The rework cost of the original deployment was the entire cost of running the AI deployment, plus the customer damage, plus the cost of rebuilding the customer service function afterwards.

That cost never appeared on the productivity dashboard. The productivity dashboard still shows “75 per cent of chats handled by AI”.

The difficulty of measuring net

Rowan Jackson’s work on errors, rework, and waste (ERW) identifies the core problem. Rework doesn’t appear in any standard financial reporting structure. Companies account for cost of goods sold, cost of labour, and cost of overhead. They don’t account for cost of rework. They don’t account for cost of error. They don’t account for cost of work that was done but didn’t need to be done, or work that was done but had to be redone because it was wrong.

Rowan’s ERW estimate across industries is 20 to 45 per cent of operating expenses. In construction, the Get It Right Initiative puts avoidable error at around 21 per cent of project turnover, and total costs as high as a quarter of project spend when you count latent defects and indirect effects. These are huge numbers that don’t show up in financial statements because there’s no line for them.

AI deployments are now being layered on top of this existing invisibility. You can’t find the rework cost of a broken AI workflow in the accounts, because you couldn’t find the rework cost of the manual workflow that preceded it either. The AI vendor can claim productivity gains without ever having to reconcile them against the rework line, because the rework line doesn’t exist.

This isn’t a minor methodological quibble. It’s the main reason optimistic productivity estimates and pessimistic real-world outcomes keep diverging. You can’t manage what you don’t measure. Nobody is measuring.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

What a net view looks like

A net productivity view for an AI deployment needs four things that most deployments don’t include.

First, a baseline quality score for the work before the AI rolls out. Customer-facing metrics where relevant. Error rates where available. Completion time and first-pass acceptance rate. Without this baseline, there’s no way to tell whether the AI-augmented output is better, same, or worse.

Second, a rework tracker. How many outputs needed correction after initial production? How much time did the correction take? Who did the correction? If the rework happens outside the team that deployed the AI, it needs to be traced back, not hidden in a different budget.

Third, a downstream quality check. Does the work actually do what it was meant to do, once it’s out in the world? Customer satisfaction, campaign performance, error rates in dependent processes, support ticket volume. AI that generates plausible output that doesn’t perform is worse than slower human work that performs.

Fourth, honest headcount accounting. If the AI deployment justified reducing headcount by three, but the remaining team plus the AI takes the same total hours to produce the same total net value, then no productivity gain actually occurred. The savings came from paying fewer salaries for the same amount of net work, which is a different phenomenon. It’s a distributional change, not a productivity change.

Most organisations running AI deployments right now are not doing any of this. They’re reporting gross output volume, dividing it by people on the org chart, and declaring victory.

The macro version

The same pattern plays out at the level of the whole economy. The UK Office for National Statistics, as of early 2026, has not yet detected a meaningful AI-driven productivity uplift in the national accounts. This is despite widespread adoption in the professional services sector, the spread of Copilot licences across enterprise IT estates, and the deployment of AI in financial services, healthcare, and retail.

If the gross productivity claims were real, they’d be visible at the macro level by now. The fact that they aren’t suggests one of three things is happening. Either the productivity gains are real but offset by rework and error at a scale that neutralises them; or the gains are captured by capital owners and not passed through to consumers in measurable ways; or the gains are real for a minority of users and non-existent or negative for the majority.

All three of these are plausible. All three of them mean the “AI is a productivity revolution” narrative is, at best, incomplete.

Why this matters for the Next Rung thesis

In the original Next Rung post, I argued that the three historical mechanisms for absorbing displaced workers were all cracking at once. New task creation, demand expansion, and skill transferability.

Demand expansion is the one most likely to stay intact in a standard economist’s model. Even if new tasks don’t materialise and skills don’t transfer, real productivity gains should expand the economy and create new demand somewhere. That’s the mechanism that made the Luddites wrong for two hundred years.

But demand expansion only works if the productivity gains are real and spread broadly. If gross productivity claims are hiding invisible rework costs, the gains are smaller than they look. If capital captures what remains, the gains don’t spread. If the deployment pattern is “replace the worker, absorb the rework in a different line item, keep the savings”, then the displaced worker gets neither a new job nor cheaper goods.

This is why Rowan’s ERW framework matters for the book. It’s not a side argument about quality. It’s a direct attack on the consolation prize that displacement discourse usually falls back on.

Action for anyone reading this

If you’re running an AI deployment, set up a rework ledger before you start. Measure baseline quality. Track correction time. Check downstream performance. Report net, not gross. If you can’t do this, don’t make productivity claims.

If you’re evaluating a vendor’s productivity claim, ask for the net number. If the vendor can’t produce it, assume it doesn’t exist.

If you’re writing about AI productivity, at minimum cite the METR study alongside the McKinsey forecast. Readers deserve to see both sides of the mirage.

If you’re setting corporate strategy on the assumption that AI will deliver headline productivity gains in the next three years, hedge. The Klarna reversal is not an isolated case. It’s the leading indicator.

The next piece in this series looks at why AI errors are particularly hard to catch, and what that means for quality in systems where the output volume outruns the human capacity to verify. If the productivity mirage has a specific mechanism, that’s where it lives.

With thanks to Rowan Jackson, whose ERW framework informs the financial analysis in this piece, and whose Klarna-related observations (shared separately) sharpened the case study.

Primary sources: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (METR, July 2025); Klarna public statements and press coverage, 2023-2025; Get It Right Initiative UK construction research; ONS productivity data to early 2026; McKinsey Global Institute, The Economic Potential of Generative AI, 2023; GitHub Next research on Copilot productivity, 2022.

In 1990, the managing director of an insurance company had a suspicion that his operation was inefficient. He didn’t know the scale. He brought in a cost accountant, dug into the numbers, and found that errors, rework, and waste were costing the business £170 million a year. Not in a rounding-error sense. In a “how is this company still operating” sense.

The group CFO launched a global review across all sixty countries where the firm ran. The findings were worse than the UK number. The CEO started an improvement programme that ran for years.

None of that £170 million had ever appeared on a financial report. It was sitting inside salary lines, inside overhead, inside the cost of everyone who had a job because the work wasn’t getting done right the first time.

The story comes from a paper by Rowan Jackson on what he calls ERW: errors, rework, and waste. He’s been working on this problem for decades. His working figure for how much of operating expenditure it consumes across industries is 20 to 45 per cent. In construction, the Get It Right Initiative (GIRI), which has done the most serious UK research on this, puts avoidable error at around 21 per cent of project turnover, and total costs including latent defects and indirect effects at up to a quarter of project spend. Rowan uses 30 per cent as a working figure once you include the waste that never gets counted. That’s roughly £10 to £25 billion a year in the UK construction sector alone.

Here’s the part that matters. None of those costs appear in the P&L in a way that lets anyone see them clearly. They’re absorbed. They’re the three people doing the work of two because the system needs redundancy. They’re the fourth draft of a document because the first three were wrong. They’re the brick wall pulled down and rebuilt because it was set five centimetres out of place. They’re the customer who doesn’t come back.

If this sounds familiar, it’s because it’s the same mechanism that makes AI productivity claims so misleading. And that’s the part worth sitting with.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Why this isn’t a construction problem

The reason I’m writing about ERW in a series about AI displacement is that I kept running into a gap when I was arguing the Next Rung thesis. The gap was this: if AI delivers genuine productivity gains, displaced workers might at least benefit from cheaper goods and services as consumers, even if they can’t find work as producers. That’s the standard economist’s consolation. Output goes up, prices come down, everyone’s real income rises. The Luddites were wrong because this happened, repeatedly, for two hundred years.

The problem with applying that logic to AI is that nobody is measuring the productivity gains net of ERW. They’re measuring them gross.

A developer using an AI coding tool ships a feature in half the time. Productivity doubled. But if that feature has a bug that takes two developers a day to find and fix in production, the gross number is a lie. The fix is in a different budget line, probably in a different quarter, possibly attributed to someone else. The productivity number stays up. The ERW cost gets absorbed somewhere the accountants don’t look.

METR ran a randomised controlled trial in early 2025 on experienced open-source developers using frontier AI tools. The developers predicted the tools would make them 24 per cent faster. After using them, they reported feeling 20 per cent faster. The actual measurement, from screen recordings of the work, showed they were 19 per cent slower.

Let that sit. A 40-point gap between perceived and actual productivity, in a field where the gains are supposed to be obvious and everyone has strong priors in favour of the technology. The participants weren’t lying and they weren’t stupid. They were doing what humans do when the rework happens invisibly, gradually, at a lower cognitive register than the work itself. It disappears into the background.

And that’s in code, where the feedback loop is tight. The code either runs or it doesn’t, eventually. In knowledge work where the outputs are documents, reports, summaries, and decisions, the feedback loop is much longer. Sometimes it never closes.

Rowan’s NHS moment

Rowan told an NHS orthopaedic consultant about the 30 per cent construction figure. The consultant laughed and said the rework rate in the NHS is closer to 60 per cent.

I don’t know if 60 per cent is right. Neither does the consultant, really, because nobody is measuring it at that granularity. But think about what a number in that range would mean. NHS England’s resource spending is running at around £187 billion a year (2024-25). 30 per cent of that is £56 billion. 60 per cent is £112 billion. The UK defence budget for 2025-26 is £62.2 billion.

Put it another way. If NHS rework is anywhere near that range, the hidden cost of getting things done twice is large enough to run the country’s entire defence capability. It sits inside the budget lines nobody argues about in Parliament because nobody can point to it.

This isn’t an NHS-specific problem. Matthew Syed’s Black Box Thinking puts preventable medical accident deaths in the US at around 1,000 a day, which would be the equivalent of two fully loaded Boeing 747s crashing every morning. Aviation responds to that kind of failure rate by grounding fleets and rewriting procedures. Healthcare treats it as background noise, because the individual deaths are absorbed into individual case files, and the system-level number never makes it onto any report.

The through-line from insurance to construction to healthcare is the same. When errors and rework are invisible to the accounting system, they get baked into the cost of doing business, and the only way anyone finds them is by going looking.

AI is about to make this much worse, because AI gives you a tool that generates output at a speed that outruns your ability to verify it.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The productivity lie in plain sight

Here’s the trap. The way AI is being sold into the economy right now is through gross productivity claims. GitHub Copilot makes developers 55 per cent faster. McKinsey estimates 30 per cent of US work hours automatable by 2030. Klarna’s AI customer service handled 2.3 million conversations in 35 languages, the equivalent of 700 full-time agents.

Every one of those numbers is a gross figure. None of them is net of rework, error, or the work required to correct what the AI produced. And when you go looking for the net figure, what you find is sobering.

Klarna’s story makes this cleaner than most. The company replaced 700 customer service staff with AI in 2023. It reported aggressive cost savings. By spring 2025, customer satisfaction had dropped enough that the CEO, Sebastian Siemiatkowski, admitted publicly that “we focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.” Klarna started hiring human agents back. The company now runs a hybrid model, with AI handling routine queries and humans handling escalations and complex cases.

Notice what’s missing from Klarna’s original productivity claims and every similar announcement. The rework cost. The complaint volume. The customers who walked away. The brand damage. The eventual rehiring budget. These are real costs, but they appear in different lines of the P&L from the “AI savings” line, and often in later periods. The initial productivity claim gets made, quoted, and celebrated. The correction happens quietly, if it happens at all.

This is the ERW pattern, just applied to cognitive work instead of factory floors. And it’s doing the same thing it did in manufacturing: letting leadership celebrate gains that don’t net out, while the real costs accumulate somewhere the quarterly reporting doesn’t cover.

What the next rung argument needs

My original post on The Next Rung made the case that AI is breaking the three mechanisms that have absorbed displaced workers in every previous technological transition: new task creation, demand expansion, and skill transferability. Rowan’s ERW framework sharpens the demand-expansion argument in a way I hadn’t fully worked through.

The standard story goes: productivity gains lower prices, lower prices expand markets, expanded markets create new work. It assumes the productivity gains are real and durable, not gross numbers waiting to be eroded by rework.

If ERW runs at 20 to 45 per cent of operating expenses across the economy, and AI deployments are being sold on gross productivity claims that don’t account for it, then the demand-expansion mechanism is weaker than the optimists claim, because the actual cost reduction is smaller than reported. Some of what looks like productivity is just cost-shifting: work that used to be done by one person with a salary line is now done by an AI plus a different person doing corrections, often across different budget centres.

That matters for displacement. If the AI deployment delivers real net productivity, displaced workers at least benefit as consumers. If it delivers gross productivity that gets eaten by rework, the savings don’t materialise, the consumer doesn’t benefit, and the displaced worker has neither a job nor cheaper goods to buy.

This is the economic equivalent of a sealed box. Capital captures the announced gains. Labour loses the jobs. Consumers don’t get the price cut because the rework ate it. The only people who come out ahead are the ones selling the AI.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

What to do about it

The honest version of the AI productivity conversation has to start by measuring net, not gross. Every organisation deploying AI into a workflow should be running a rework ledger alongside the productivity dashboard. How many outputs had to be corrected? How many downstream errors traced back to AI-generated content? How many hours of human review does each AI-assisted task actually consume? What was the quality score before and after, on the customer-facing side?

Rowan’s ERW work, Harrington’s process improvement methodology from the early nineties, the Toyota Production System, Danaher’s operating model, these aren’t legacy industrial frameworks. They’re the only serious methods the business world has ever developed for measuring what’s actually happening in a workflow, rather than what leadership wants to be happening. They apply to AI deployments the same way they apply to car plants.

If you’re running an AI rollout and you’re not tracking rework, you’re not running a productivity programme. You’re running a perception programme. They’re different things, and only one of them shows up on the balance sheet in the end.

The next piece in this series goes further on what that looks like in practice. For now, the point is simpler. The productivity gains you can’t measure are the ones you shouldn’t claim. And the costs that don’t appear on your financial report are the ones that will eventually eat it.

With thanks to Rowan Jackson, whose paper on Errors, Rework, and Waste (ERW) provided the framework and several of the examples in this piece, including the £170 million insurance case, the NHS consultant’s 60 per cent figure, and the broader observation that errors, rework, and waste are the silent killers of organisational performance.

Sources: Get It Right Initiative (GIRI) construction error research; METR 2025 study on developer productivity with AI tools; Moffatt v Air Canada 2024 BCCRT 149; Klarna AI customer service reversal coverage, 2025; NHS England financial directions 2025-26; UK defence budget data from House of Commons Library; Matthew Syed, Black Box Thinking; H.J. Harrington, Business Process Improvement (1991).

I’ve seen five of them in the last 24 hours. Different people, different parts of the country, a couple in the US. None of them know each other. None of them would describe what’s happened to them as part of anything.

They’re all posting the open-to-work card. The green ring on the profile photo, the “this isn’t a post I expected to write,” the bit at the end asking for likes and shares to help it reach the right people. Which is fair enough. As an individual move it’s one of the boxes you check. Everyone else does it, so you do it too, and you’ve got as much chance as the next person. I’m not knocking anyone for posting one. If I were in their position I’d probably do the same.

But here’s what struck me, and it took five of them in a day for it to land.

Every single one is looking for a job in the exact thing they’ve just been made redundant from.

The senior marketing manager wants another senior marketing role. The operations lead wants another operations lead role. And in the normal logic of a career, that’s the sensible thing. You’ve got ten years of doing the thing, so you go and find someone who needs the thing done. That logic has held for every previous generation of worker who lost a job. The role still existed somewhere. You just had to go and find the version of it with your name on it.

The logic only works if you’re the one being displaced.

If the role itself is being made redundant, going to find the same role somewhere else is going back to the same edge of the same cliff, just at a different company. You’re not unlucky. You’re early. And the thing you’re applying for is the thing that’s quietly being removed from org charts in the same month you’re applying for it.

That’s the part that nobody’s saying.

Go and read the comments under any of these posts. I did. They’re full of decent people being kind. “That’s a shame.” “Good luck, all the best.” “Can’t believe it’s happened to you.” Genuine, warm, well meant. And not one of them says the actual thing.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Nobody says: this happened because you’re in a senior marketing role, and a senior marketing role is one of the easiest things in the building to displace, because an AI can do most of the day-to-day now. Nobody says: the reason it’s hard to find another one is that the people hiring have worked that out too. Nobody says: the gap between the role you want and the roles that exist is going to keep widening, and the next one might not last either.

They don’t say it for one of three reasons. Either they don’t know it. Or they know it and don’t want to be the person who says it under a post where someone’s already having the worst week of their year. Or, and this is the one that sits with me, they half-know it and saying it out loud would make it real. As long as nobody names it, it stays a run of individual bad luck. A few people who didn’t quite land the next thing. The moment somebody names it, it becomes a pattern, and the pattern has a direction, and the direction includes the person reading the comment.

So everyone keeps it as bad luck. It’s easier. It costs nothing to type “good luck, all the best.” It costs a lot to type what’s actually happening.

This is what the slow version of a crisis looks like. Not a crash. Not a headline. Not a number on the six o’clock news. Just a steady drip of individually-explainable exits, each one with a perfectly reasonable story attached, each one absorbed and moved past, each one followed by a sympathetic comment and a scroll to the next thing in the feed.

There’s a phrase for the way societies live inside a situation everyone can feel and nobody will name. Everything stays the same until it suddenly doesn’t. You keep finding ways to read the evidence as normal, right up until you can’t.

The redundancies are real. They’re happening now, this week, to people you’re connected to. But they don’t arrive as displacement. They arrive as a green ring on a profile photo and a brave caption and 188 likes. The wave is already here. It just doesn’t look like a wave from inside it. It looks like a coffee request.

The displacement will not be televised.

It’ll be posted, one open-to-work card at a time, and we’ll all wish them luck, and we’ll scroll on.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The timeline is phased. The first phase is quiet and already underway. The second phase breaks. The third phase reshapes. You need to understand the shape of all three because what you do now determines your position in all three.

Phase One, 2026-2028, is the quiet squeeze. Hiring freezes in knowledge work sectors. Contractor positions disappearing. Quiet restructuring that doesn’t show up in headlines because HR departments know the labour market is watching. Real unemployment rises while official figures stay flat because people leave the job market entirely rather than accept precarious work. The people in exposed roles start to notice that the stability they expected isn’t coming. The political class hasn’t developed a coherent response yet because the crisis doesn’t feel urgent if you’re not the one being squeezed. This phase feels like a normal downturn. It isn’t.

Phase Two, 2028-2032, is the breaking point. Enterprise AI adoption reaches critical mass. Competitive pressure forces the holdout companies to adopt. The pace of displacement accelerates past the pace at which retraining or natural job transition can absorb it. Housing markets face pressure as people who expected to service 25-year mortgages lose income. Political systems strain. The gap between “this is a technical transition” and “this is a systemic crisis” becomes impossible to deny. Unemployment peaks. Political systems are forced to choose between managed transition or something worse. This is the phase where institutional responses are determined. The choices made here, whether to tax AI gains or let them concentrate, whether to fund retraining or assume the market will sort it, whether to expand the social safety net or leave it where it is, those choices determine which scenario unfolds.

Phase Three, 2032-2040, is the reshaping. Whatever institutional response emerged in Phase Two plays out at scale. Managed Transition looks like the post-war settlement being rebuilt at higher velocity. Neo-Feudalism looks like the slow consolidation of the extractive model. Fragmentation looks like separate regions pursuing incompatible strategies producing unexpected interactions. The black swan, if it hits, makes everything unpredictable. But by 2032, the main direction is set.

The window for influencing this is open now and closes around 2032. Not absolutely. Politics can surprise. Crises can force rapid reversals. But the window for voluntary transition, for policy made in calm rather than panic, for solutions chosen rather than imposed, that window is roughly six years wide. After that, the structural decisions have been made and the space for modification narrows sharply.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Here is what I think we do with that window.

First, demonstrate that managed transition works at scale. The evidence from pilots is clear: UBI doesn’t cause work participation to collapse, retraining works when funded properly and taken seriously, people respond to genuine opportunity. But pilots don’t move political systems because pilots are small and expensive and can be dismissed as unrepresentative. What moves political systems is scale demonstration: a city, a region, implementing genuine managed transition policies and showing outcomes. If even one major economy implements something close to managed transition and the results are decent, the probability of other economies adopting increases. If none do, the path toward neo-feudalism becomes harder to contest.

Second, rebuild the connection between productivity and wages. For 60 years after the Second World War, productivity growth and wage growth tracked together. Since the 1980s, they’ve diverged. Productivity up, wages flat. That disconnect is not a law of nature. It’s a policy choice, or more accurately, a policy non-choice: the absence of mechanisms that would force productivity gains to translate into wage gains. The mechanism exists. It’s called taxation and redistribution. Implementing it is not economically impossible. It’s politically contested. Demonstrating that it’s possible, that you can have high productivity and shared gains simultaneously, is the necessary precondition for the policies that enabled it post-war to seem non-radical now.

Third, protect the people in the blast radius through the transition period. The displacement isn’t happening to everyone simultaneously. It’s happening to orchestration-heavy roles first: coordinators, analysts, junior consultants, middle managers, marketers focused on process rather than strategy. These are the people with mortgages, kids in school, the lowest tolerance for income shock and the highest political voice. The people who can articulate what’s happening and demand response. If these people are left to fend for themselves, the political response will be chaos. If they’re supported through the transition, the response can be managed. The difference is not charity. It’s structural stability. You don’t let the politically articulate middle class fall into precarity without consequences. Those consequences are bad for everyone.

Fourth, invest in the capabilities that survive. AI automates pattern-based, documentation-heavy, coordination-heavy work. It doesn’t automate trust, genuine problem-solving, human relationships at scale, creative work where originality is the product, skilled trades with physical presence, care work. The education and training infrastructure should reflect this. Teaching people to code is not the solution because most of those coders are not going to be hired; the competitive supply of code-writers will exceed demand. Teaching people genuine problem-solving, communication, the ability to navigate ambiguity, relationships and trust-building, those investments do translate into resilience. The system that measures this is Enterprise Skills at K12 level, but the principle applies across the lifespan. Invest in human capabilities that an algorithm can’t replace. Don’t invest in the illusion that tech skills will save you.

Fifth, maintain optionality. The scenarios are not precise. Nobody knows exactly how this plays out. The people and institutions that survive are the ones with enough flexibility to respond when the actual future arrives. Don’t optimise for one scenario. Build resilience. Diversify income, skills, relationships, assets. Have multiple plans, not one plan with contingencies. The worst position you can be in is locked into one future with no ability to adapt. The best position is flexible across scenarios.

Those five aren’t complete. They’re the places where the window is open and closing fastest. They’re also the places where individual action and institutional action reinforce each other. If you build a diversified, resilient working life, you create demand for education systems that teach resilience. If you invest in trust networks and community, you create the social infrastructure that managed transition needs. If institutions demonstrate that managed transition works, they create permission for individuals to position themselves differently. The system and the individual aren’t separate. They feedback.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

Here’s what I know and what I don’t.

I know the displacement is real and its scale is comparable to the largest labour market disruptions in modern history. I know the current trajectory, without policy intervention, trends toward concentration. I know the window for voluntary, policy-driven response is open and closing. I know the precedent: when managed transition has been implemented, it works.

I don’t know if the political will exists to implement it. I don’t know if the pace of AI development will exceed the pace of institutional response. I don’t know if the black swan is 5% or 50% in reality. I don’t know which scenario we end up in. That’s why I’m not betting on one outcome. That’s why I’m suggesting moves that make sense across all four.

The book is the full argument: what’s happening, why it matters, what the futures look like, what you do now. This Substack is where the argument lives and where it develops as the evidence changes. I’ve said enough in these posts that you know roughly where I’m standing and roughly what I think matters. The rest is the details and the lived experience of what this transition actually feels like, which is what the book explores.

The window is open. Not for much longer. What you build in the next six years shapes what you’re positioned in when 2032 arrives.

There’s a concept in nuclear physics called critical mass. It’s the minimum amount of fissile material needed to sustain a chain reaction. Below that threshold, individual atoms split but the energy dissipates. Nothing much happens. Above it, each reaction triggers others, and the thing becomes self-sustaining.

Enterprise AI adoption crosses that line somewhere around 2028.

This is Chapter 7 from Part Two. It’s the chapter where the gradual pressure of Part One becomes unmistakable. Where hiring freezes and contractor squeeze-outs give way to explicit redundancy programmes that cite AI as the reason. Where the denial stops being possible.

The mechanism matters because it explains why individual skill or excellence becomes insufficient protection. Your job isn’t disappearing because you’re bad at it. It’s disappearing because your employer is losing the competitive race against companies that have already rebuilt their operations around AI.

Here’s the frame: by 2028, 88 percent of organisations are using AI in at least one business function. That sounds like the revolution has already happened. The number that actually mattered was different. Thirty-one percent of use cases had reached full production in 2025, and that number was doubling annually. Full production means the AI isn’t being tested. It’s doing the work. Decisions are flowing through it. People who used to do those tasks are now doing something else, or nothing at all.

Run the maths. If 31 percent of use cases hit production in 2025 and the doubling rate held, 60 percent were in production by 2027. By 2028, the question wasn’t whether your company used AI. It was whether your company had been redesigned around it.

The productivity gap became impossible to ignore. Top AI-native startups were generating average revenue per employee of $2.5 million. Traditional companies averaged $200,000. That’s not a competitive advantage. That’s a structural difference so large it operates on different economics entirely.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The decision that killed companies was choosing between two bad strategies. Adding AI to a legacy structure is like putting a jet engine on a horse cart. You get noise and expense and the cart falls apart. The companies that pulled ahead were the ones that built new carts. But rebuilding from scratch required capital, will, and organisational design changes that most mid-market firms couldn’t execute while maintaining business-as-usual.

So they tried to manage decline instead. Cutting costs. Consolidating departments. Reducing headcount by attrition. These measures saved money in the quarterly numbers. They didn’t change the underlying economics. Every quarter the gap widened. Every quarter the catch-up cost increased. After a certain point, catching up became impossible.

This pattern, technology historians call the incumbent’s dilemma. The incumbents weren’t stupid. They were rational actors making decisions within a framework that no longer applied. Improve incrementally. Protect margins. Retain talent. All sensible advice in stable competitive environments. All catastrophically insufficient in one undergoing phase transition.

The adoption curve was the fastest in human history. The telegraph took 80 years to peak. Television took 15. Generative AI achieved 78 percent organisational adoption within a single year. The reason is structural. Previous technologies required hardware. AI required an internet connection and will.

This compression of the adoption curve had a specific consequence. Previous technology transitions gave industries decades to adjust. The cloud transition gave them fifteen years. AI gave them roughly five, from mainstream availability in 2023 to restructuring necessity by 2028. That’s not enough time for labour markets to absorb displaced workers. It’s not enough time for retraining programmes to scale. It’s not enough time for most people to fully understand what’s happening.

The tipping point hit different sectors on different timescales, but the sequence was consistent. Professional services went first. Legal services felt it when major firms deployed AI contract review systems that processed in hours what paralegals had processed in weeks. The entry-level pipeline, the three years of document review that had always been the route into partnership-track careers, was the exact pipeline AI automated first. Junior associate hiring dropped 20-30 percent at the major firms between 2026 and 2028. Revenue held steady. Headcount didn’t. Revenue per partner increased, which looked like success on a balance sheet and looked like a closed door to law graduates who’d spent three years and six figures earning a qualification for roles that no longer existed.

Consulting followed with a six-month lag. Accounting had been on notice longest. For the professional class, critical mass meant something very specific. It meant the economic viability of their roles was no longer a matter of individual performance. It was a function of their employer’s competitive position. And their employer’s competitive position was deteriorating.

This is where the abstraction meets the mortgage payment. You can be excellent at your job. Skilled, diligent, well-regarded. And if your employer is on the wrong side of the AI transformation curve, your excellence is irrelevant. The company isn’t failing because you’re bad at what you do. It’s failing because a competitor is doing what you do with a quarter of the people.

The Next Rung is a book on how AI is quietly dismantling the middle of knowledge work, and what you can do about it before the market decides for you: pre-order it before it publishes in January.

Pre-Order The Next Rung

The professional class had never experienced this particular form of economic threat. Manufacturing workers knew it. They’d lived through plant closures where individual performance didn’t matter because the whole facility was uneconomic. But professionals, the people with degrees and certifications and career paths that climbed through predictable stages, had always operated on the assumption that competence provided security.

The system stopped providing it. Not because professionals had done anything wrong. Because the system had changed state.

The quiet squeeze became loud in the most ordinary possible way. Not riots. Not protests. Emails. The email from HR. The restructuring announcement. The meeting request from your manager with no agenda attached. The job listing for your role, rewritten around AI capabilities, posted while you still held the title.

Between 2025 and 2028, displacement felt like bad weather: unpleasant, temporary, someone else’s problem. By 2028, the weather had become climate. When your company announces its third restructuring in eighteen months and the new job descriptions all include AI proficiency as a core requirement, and the roles being eliminated are the ones that look exactly like yours, the denial architecture starts to crumble.

The cascade didn’t stop at companies. It moved through supply chains, client relationships, entire industries. A company going AI-native didn’t just reduce its own headcount. It changed the economics for everyone connected to it. The outsourcing firm that provided back-office support. The staffing agency that placed contractors. The training company that ran professional development. The landlord who leased the office space. The sandwich shop on the ground floor.

Commercial real estate felt it early. Office vacancy rates in professional services districts climbed steadily from 2027 onward. The recruitment industry was hit with particular force. Staffing agencies found their market contracting from both directions. Fewer roles to fill because companies needed fewer people. Fewer candidates to place because displaced professionals were competing for a shrinking pool.

The professional training industry collapsed. The companies that ran leadership courses, project management certifications, and professional development workshops found that their customer base was cutting spend at unprecedented rates. Only 13 percent of firms intended to invest more in training in 2025, down from 32 percent the year before. The professional development industry had existed in the space between what workers could do and what employers needed them to do. When AI closed that gap, the industry lost its reason for being.

By 2029, the question had changed. In 2025, people asked: will AI take my job. By 2029, they asked: how long before it takes mine. The shift from if to when was the psychological critical mass, the moment denial became impossible.

What changed wasn’t the information. The data had been available for years. What changed was the inability to maintain the story that it wouldn’t happen to people like you. When the person sitting next to you at a dinner party, someone with the same degree, the same career trajectory, the same kind of mortgage, tells you they’ve been made redundant because an AI system now does what their department used to do, the abstract becomes concrete.

Critical mass wasn’t a theory any more. It was happening in real time.

The full chapter walks through the mechanism in detail. It shows you how the competitive cascade works, what board-level AI governance looks like, and why the traditional strategies for managing technological disruption failed. It explains the incumbent’s dilemma and why some companies became uncompetitive not because they were poorly managed, but because they made rational decisions within a framework that no longer applied.

It’s not comforting. It’s not designed to be. It’s designed to explain what you’re seeing, and to help you position before the wave hits your particular door.