The pie gets bigger. The pieces don't redistribute on their own.

There's a comforting story going around about AI and jobs: automation always created more work than it destroyed, the economy grows, everyone reallocates, it'll be fine. There's a gloomier one too: a tidal wave of redundancy is weeks away. Both are too tidy.

I came across a good article on the Workonomics newsletter recently — "The Pie Gets Bigger. The Pieces Don't Redistribute Automatically." — and it names the thing both stories skip over. Growth doesn't reach displaced people by default. The pie can grow while the people who lost their slice are still standing there empty-handed. I want to build on that, because once you take it seriously, the interesting question stops being "how many jobs will AI destroy?" and becomes "what actually decides, job by job, which way this goes?"

First, the headline number

There isn't one. The credible forecasts disagree by an order of magnitude, and the famous figures don't measure what people think they measure. And I don't trust them — probably right in the short term, but I think they miss the rapidly growing capability of AI over the coming years; like many reports, they lag reality.

Goldman Sachs' much-quoted "25%" is 25% of work hours that are technically automatable — not a quarter of jobs gone. Their actual labour forecast is 6–7% of workers displaced over a ten-year transition, with unemployment up less than a percentage point if it's spread out. The Tony Blair Institute models 1–3 million UK jobs displaced over the long run — which sounds enormous until you notice the UK economy already loses and creates around 450,000 jobs a year in normal churn.

So anyone quoting a single dramatic number in the near future is either selling something or hasn't read the footnote. The honest position is a spread. The useful work is understanding the mechanism underneath it.

What actually decides it

Most times, automation does not eat a whole job. It eats tasks. Over six decades and 271 occupations, economists could find exactly one that was largely automated out of existence — elevator operators. Everything else got reshaped, not deleted. That will likely hold, for now.

Which means the outcome for any given role turns on a handful of forces, not on how clever the technology is:

Whether the demand can grow. This is the master switch. When demand for something is elastic — make it cheaper and people buy a lot more — automation can increase employment even as it automates tasks. That's the real lesson of the bank teller, the example everyone reaches for. ATMs quadrupled, yet teller numbers held for thirty years, because cheaper branches meant banks opened more of them. But — and this is the Workonomics point exactly — tellers shrank as a share of the workforce, and once branch demand stopped growing, headcount fell. Same machine, opposite result, depending only on whether the market could still expand.

So three rough modes, and most of the economy is a blend:

• Some work gets replaced outright — slowly, over varying timeframes. Routine voice support and back-office processing are furthest down this road.
• Some markets expand and soak up the freed effort — where cheaper output unlocks far more demand.
• And some markets can't expand. You don't need twice as much bookkeeping just because it got cheaper. Here, automation quietly means fewer people for the same output — even with no drama, no announcement.

There's a twist: the firms that use AI well will take more market share — but in the can't-expand sectors, that's a bigger slice of a shrinking human-served market. Winning and shedding people at the same time is not a contradiction. It's the most likely outcome in a lot of industries.

Whether your tasks are complemented or substituted. This is decided worker by worker, sometimes within the same job title. If the machine does the bit you were slow at and frees you for the bit you're good at, you get more valuable. If the machine does the bit you were the expert in, your wage falls. Two people with the same job description can land on opposite sides of that line.

And, counter-intuitively, how good the technology actually is. The economists Acemoglu and Restrepo have a lovely finding: it isn't the brilliant automation that hurts workers most — it's the mediocre kind. Technology good enough to replace people but not good enough to make the business much more productive. It displaces without growing the pie enough to rehire anyone. A lot of today's AI deployments are exactly this: "so-so" automation that trims a team without transforming the work.

The cases ahead — the same forces, four trades, four different endings. Jump to any of them:

• The taxi — the market explodes, and the drivers still lose
• Car insurance — the market contracts, squeezed from several sides at once
• Security — cheaper machines, yet far more security gets bought
• Care — the hopeful inversion: robots take the grind, the human part is what's left

Case study: the taxi

The framework above is abstract. Here's what it looks like when it arrives.

The disruption is closer than people think

I'm a technology optimist, so let me say the optimistic part plainly: autonomous ride-hailing is going to be good, and it's coming fast. Tesla's Robotaxi is positioned to undercut Uber on price the moment it scales — because the driver is the single biggest cost in a ride — and the rollout will follow the cost gradient. The US first, where driver wages are high and the savings are largest. Europe next. The rest of the world more slowly, precisely because cheaper local drivers blunt the economic case for a while. Then Robovan arrives underneath even that, dropping the cost-per-seat again. China may run as its own market on its own stack. The scale being planned isn't a pilot: Tesla has applied to the Nevada Transport Authority for a permit covering up to 5,000 robotaxis in Clark County — greater Las Vegas — in the first twelve months alone. This is not a someday technology. It's a this-decade one.

And before anyone objects that autonomous trucks and vans can't work because a driver still has to load and unload — that's the task-versus-job point again. The loading stays; the driving goes; the role reshapes around the two ends. Fewer people, doing a different job.

The pie grows — and the drivers still lose

Here's the elasticity point from earlier, live. Make a ride cheap enough and people take far more rides — trips they'd have walked, skipped, or not been able to afford. The mobility market doesn't shrink; it expands, possibly hugely. By the logic of "automation grows employment when demand is elastic," this should be the happy case.

And this is where the optimism is earned, because the gain is far bigger than cheaper Ubers. Drive the cost of a ride low enough and you don't just expand an existing market — you hand reliable, on-demand transport to people who have never really had it. Older people who gave up the car keys and now ration their trips. Disabled people for whom every journey is a logistics problem. Young people, and the millions who can't afford to run a car at all. For most of them a robotaxi is cheaper than today's taxi, and quicker and far more useful than a bus that comes twice an hour — available at 6am or 11pm when public transport simply isn't. That isn't a convenience upgrade; it's access — to work, to healthcare, to family — for segments the transport system has quietly underserved for decades. The economic ripple from that is real, and it's broadly good. The pie doesn't just grow; it grows for people who were barely getting a slice before.

Except the expanded market is served by vehicles, not people. The pie gets much bigger and the original drivers get none of the new slice. A robotaxi fleet does create jobs — cleaners, chargers, depot technicians, remote operators watching for the situations the car can't handle — but it's a fraction of the headcount it replaces, in different skills, often in different places. This is the teller story with the comforting half removed: the teller's bank kept hiring tellers for thirty years because branches kept opening. A driver's "branch" is the car, and the car no longer needs them.

Who actually drove the taxis

The company-level story walks right past this.

Think about who drives. A lot of them are part-time. Many are running a second income to make the month work. Some fit it around caring, study, or a health condition that rules out a nine-to-five. And a good number landed there because it was the job you could start — no gatekeeper, no credential, no interview, just a licence and a car. For all its problems, gig driving has been one of the most accessible on-ramps in the labour market.

That's why this is worse than the teller case. The teller usually had a bank behind them — somewhere with adjacent roles to move into. The driver often took the wheel because the other doors were already shut. When automation closes the on-ramp itself, the redistribution doesn't just happen around these workers — it happens to the people with the least slack to absorb it.

So what do they do instead?

The honest answer is that we don't have one, and the glib answers are worse than no answer. "They'll retrain" is not a plan when the displaced group is part-time, cash-tight, and time-poor, and when the obvious next gig — delivery, warehousing, other driving — is on the same automation list a few years behind.

The question worth the work isn't "how many jobs will go." It's: which specific groups of people lose their on-ramp, and what is the realistic next rung for them? Match the displaced to the destination, honestly, group by group. Almost nobody is doing that. It's harder than a headline number, and it's the only thing that turns "the pie got bigger" into anyone actually eating.

Case study: car insurance

The taxi is the elastic case — the market explodes and the workers still lose. Car insurance is the opposite shape: a market that gets squeezed, from several directions at once, and may not stay independent at all.

I'll be honest that this one is more speculative than the taxi — I'm reasoning forward rather than pointing at a permit application. But the direction of travel is hard to argue with.

A market squeezed from several sides

Motor insurance is priced on risk — on things going wrong. Make fewer things go wrong and you shrink the product. And almost everything coming down the road reduces the number of claims or automates handling them:

• Assistance and autonomy mean fewer accidents. Even before full self-driving, the systems already shipping — automatic braking, lane-keeping, collision avoidance — cut exactly the prangs that fill the claims pile. Fewer crashes, fewer claims, lower premiums, smaller market.
• Cameras everywhere mean fewer thefts. A car that watches its surroundings, records continuously and phones home is far harder to steal and far easier to recover. Theft is a big slice of motor claims, and it starts to drain away.

Put those together and the pool the whole industry feeds on gets shallower year on year. This isn't even the can't-expand market from the framework — it's a step worse. It's a market that actively contracts.

And there's a further squeeze, on the cost of settling each claim rather than the number of them. Today a disputed claim leans on human witnesses, who are unreliable. Put enough camera-equipped, autonomous cars on the road and that changes: a car with no stake in the crash can file what it saw. Liability stops being argued and starts being calculated automatically. That hollows out the most labour-intensive, judgement-heavy part of the business — the loss adjusters, the investigators, the fraud teams, the whole back-and-forth of who-hit-whom — and it quietly kills "crash-for-cash" fraud, which only survives in the absence of a reliable witness.

People will balk at the idea of every passing car logging what it sees, and they're right to; but the objection is answerable — incident-triggered capture, anonymisation, regulated access — and the economics are relentless. Cheaper, faster, fraud-proof settlement is exactly the kind of saving that gets forced through in the end, objections and all.

The slow tail, and the fast cliff

There are two clocks running here, and they tell different times.

The slow one is reassuring: the car fleet turns over slowly. The average car on the road is well over a decade old, so even once every new vehicle is safer and harder to steal, the existing fleet — and the premiums it pays — takes the better part of fifteen years to age out. On that clock this is a gentle, manageable decline, with plenty of time to build a bridge.

The fast clock isn't reassuring at all. The bigger threat to the people who work in insurance isn't the gradual fall in claims — it's a structural collapse that could land in a single product cycle. I speculate that at some point the manufacturers will underwrite the cars themselves, and the independent insurer will be cut out.

Why them, and not the incumbents? Because the carmaker already has what an insurer spends billions trying to approximate — it knows how you drive. While the car is still manual, the manufacturer sees every input in real time; and increasingly it can intervene — trim the speed, prime the brakes, flag the risk. Once you both price the risk and partly control it, underwriting stops being a separate industry and becomes a line item on the car. And for the autonomous fleet, manufacturers may not just be able to self-insure but have to — because liability shifts from the driver to the machine, and the maker is the party left holding it.

So the same trade faces a slow erosion from below and a fast decapitation from above. Either way, the independent motor-insurance market — brokers, underwriters, claims operations — is on the wrong side of it.

Who actually does this job

Insurance is full of exactly the work the labour market can least afford to lose.

Motor insurance runs on people in clerical and customer-facing roles — claims handlers, call-centre staff, underwriting assistants, brokers, loss adjusters — often concentrated in regional offices in towns where they're a steady white-collar employer. It has long been one of the more accessible white-collar on-ramps: no degree required, trainable on the job, friendly to people returning after caring, and disproportionately staffed by women doing exactly that. It's the office equivalent of the gig-driving on-ramp — a door that opens without a gatekeeper.

When the market shrinks and consolidates into a handful of manufacturers' captive operations, those jobs don't move — they disappear, in the specific places that depended on them.

So what do they do instead?

The same honest non-answer as the taxi, with the same trap. The obvious next step — other claims work, other admin, other call-centre roles — sits on the same automation list, a few years behind. "Learn to code" is no kinder here than "they'll retrain" was for the driver.

And the regional concentration makes it worse. When a town's insurance office is its anchor white-collar employer, its quiet contraction isn't a statistic spread thinly across the economy — it's the same few hundred people, in the same place, losing the same on-ramp at once.

A word on the robots

Everything so far has been software and sensors — autonomy, cameras, models. The next two cases are physical, and they wait on a machine that isn't here quite yet: the general-purpose humanoid robot.

Humanoid robots are not in the labour market today. It will likely be a couple of years before we see them in any noticeable number, and years more — plausibly a decade or two — before they scale enough to shift whole sectors. On this article's clock, this is the slowest tail of all.

But two things make them worth thinking about now. The slowness is exactly what makes the displacement manageable — if anyone is looking. And, more important, a humanoid reaches work that fixed automation never could. Factory robots conquered the structured, repetitive, bring-the-work-to-the-machine world decades ago. What they could never touch was physical labour that's distributed, unpredictable, and happens in human spaces — a building to be patrolled, a person to be lifted. A machine shaped like us, that can go where we go, is the thing that finally reaches those jobs. That's why the two hardest cases — security and care — are robot cases.

Case study: security

The disruption

Security is one of the first places a humanoid — and its cheaper cousin, the roaming autonomous patrol vehicle — lands, because so much of the job is presence: being there, watching, deterring, reporting. A camera on wheels or legs that never sleeps, never gets bored, and costs less than a wage does a great deal of that.

Two pulls at once

This is the elastic case again, and more two-sided than most. One pull replaces guards directly with cheaper machines. The other is the taxi lesson in reverse-gear: make security cheap enough and people buy far more of it. Roaming autonomous patrols for streets, car parks and campuses — the same kit a police force will use — and always-on premise security for the retail units, warehouses and offices that could never justify a human guard before. The security market doesn't shrink; it expands, possibly a lot.

There's a twist specific to this trade. Human guards have, for years, been told not to physically intervene — the liability of an injured employee outweighs the stolen goods, so the modern guard is an "observe and report" role, not a "tackle the shoplifter" one. That restriction doesn't apply to a robot. A machine can be sent into the confrontation a human is now forbidden from: it can be damaged, but it can't be hurt, traumatised, or sued over. So robots don't merely match the guard — they re-open a part of the job the humans were pulled out of.

And the layer above grows. Someone has to run the fleets, watch the feeds, handle the exceptions, hold the contracts — security management expands even as the patrolling headcount falls.

Who actually does this job

Guarding is one of the great accessible on-ramps: shift work, low barrier to entry, no degree, frequently the job taken by older workers, migrants, students, and people topping up a second income. It asks for reliability more than credentials — which is exactly why so many people for whom other doors are shut end up there.

So what do they do instead?

Here the news is a little better than the taxi, and worth saying plainly. The growing management-and-monitoring layer is a next rung, and some guards will climb it. But it's a narrower ladder than the frontline it replaces — far fewer people run a fleet than walk a beat — and the rung sits higher, asking for different skills. For everyone who doesn't make that step, the honest problem returns: the expanded market is patrolled by machines, and the new jobs sit above the displaced worker, not in front of them.

Case study: care

This is the only case that's a net positive for jobs — a social revolution rather than a grim trade-off. It runs opposite to all the others.

The status quo is the scandal

Start with where we are, because it's bleak. Care — of the old, the disabled, the chronically ill — is understaffed, underpaid and physically brutal, and vast numbers of people live in conditions that are, frankly, torturous: rationed visits, undignified waits, families coping alone. Everyone knows it is broken. It is a problem every government promises to fix and none can afford to — the bill for properly funding social care has defeated administration after administration, the current one no more than the last. The money isn't there, and pretending otherwise hasn't worked for thirty years.

What the robots actually change

This is where a machine that can lift, fetch, clean, monitor and be present through the night changes the equation — not just by being cheaper than a carer, but by doing the part of care no human should have to do alone: the lifting that wrecks backs, the 3am checks, the relentless physical grind that burns carers out and that we have never been able to staff. Over the slow robot timescale, that isn't job destruction. It is filling a gap we have catastrophically failed to fill any other way.

The new human role

Take the crushing physical load off the carer and what remains is the human half — company, conversation, reassurance, knowing the person. A new role comes into focus: the care companion, whose work is presence and relationship rather than lifting and cleaning. Less physically destroying, more human. For once the technology doesn't hollow the job out — it removes the part that was breaking people and hands back the part that was the point.

That is the optimistic shape this whole article has been circling: let the machines take the work that was never really about being human, and they hand back the work that is. Care is where that trade is at its most stark — and its most worth getting right.

Why "it'll reallocate" isn't a plan

Put the cases together and you get the uncomfortable middle position. The evidence doesn't support a sudden imminent jobs apocalypse — automating firms tend to grow, and across history the net employment effect has been close to a wash. But the distributional effect is real and brutal in its specifics. Most of the rise in US wage inequality over the last forty years traces to one thing: workers whose routine tasks got automated, in industries that automated fastest. The economy adjusted. Those particular people, mostly, didn't.

New roles appear — but they appear around the displaced worker, not for them. The teller didn't become a data scientist. The pie grew; her slice didn't.

None of this is an argument against AI. The growth is real and worth having. It's an argument that the redistribution is a choice and will need focus, not an automatic consequence — and right now it's a choice almost nobody is making on purpose. The window to build the bridge between displacement and recovery is open while adoption is still early and uneven. It won't stay open if the wave arrives fast instead of slow.

The pie is going to get bigger. Whether the pieces reach the people who need them is still up to us.