How Much White-Collar Work Will AI Take? Forget the Number — the Direction Isn't in Question
How Much White-Collar Work Will AI Take? Forget the Number — the Direction Isn’t in Question
I’m not here to comfort you, and I’m not here to fearmonger either. Every institution has its own number for this — anywhere from 6% to 50% — and I have no interest in joining that fight. What I care about is the shape of the thing: how it happens, who it hits first, and why still having a job today won’t save you. The number is up for debate. The direction isn’t.
1. First, sort the layoffs into buckets
Lately your feed has probably been full of this: on YouTube, on Bilibili, on every platform, more and more people telling their layoff story — especially from Big Tech and IT roles. Watch enough of these and you start to feel like AI has already started eating jobs.
Let’s be fair first: most of the recent wave of Big Tech layoffs is not “AI did your job.” The evidence for large-scale, direct replacement is actually thin. One survey found roughly nine in ten executives admit AI has had “almost no impact” on employment at their own company; in most breakdowns of why companies are cutting staff, AI ranks fifth — behind the economy, restructuring, and shutdowns.
But this also isn’t the plain “we over-hired, now we’re correcting” story of 2022. The most accurate read is this: the companies that understand AI best are the ones front-running their own bet on its future. They’re moving huge chunks of headcount budget into AI infrastructure — Oracle cut roughly 30,000 jobs in one swoop to free up well over a hundred billion in cash; Meta laid off 8,000 people while posting record revenue at the same time — and, on the assumption that “AI will let us run leaner from here,” they’re restructuring and slimming down early, locking in margins and valuation ahead of the curve. Amazon’s CEO put it plainly: “In the next few years, as AI rolls out further, we’ll need fewer people.” Block’s Dorsey went further: “Most companies are still moving slow — within a year, everyone will be making the same structural call.”
In other words, the “reason” behind this wave of layoffs is AI, but the mechanism isn’t “AI took your job” — it’s “the people who understand AI best are betting on the future of jobs like yours, and moving first.” Mixed in, inevitably, is some plain old blame-shifting and “AI washing” — an ordinary cost-cutting exercise wrapped in an AI story to please the market.
And that’s exactly what worries me more. Because when the companies closest to AI, the ones who understand model capability best, are already reorganizing their people around the premise that “AI will make organizations smaller,” that is itself the strongest kind of leading indicator. What they’re betting on — AI actually, directly, at scale, doing the work — has only just begun to show itself. It hasn’t arrived yet. And when it arrives, it won’t look like a cyclical layoff:
- A cyclical layoff is a gust of wind — it passes, and the jobs come back.
- AI displacement is a change in climate — not a gust, but a drought that settles in for years: the water table drops a little further every year, and it never goes back to where it was.
So the layoff stories flooding your feed right now are more like a prologue: part correction from old over-hiring, part Big Tech getting into formation early for AI. The real structural replacement — the climate shift that doesn’t reverse — is still ahead. This piece is about that: how it happens, who it hits first, and, in the end, how much it takes.
2. Breaking the first illusion: white-collar workers think they’re out of reach
For decades, automation mostly ate blue-collar repetitive labor — assembly lines, warehouses, manufacturing. White-collar workers carried an implicit sense of safety:
My job lives in my head. Machines can’t reach it.
That sense of safety is failing. Not because AI has become “conscious,” but because three things are true at once.
First, most white-collar work is, at bottom, “language + rules + patterns.” Writing reports, answering emails, building spreadsheets, researching, writing code, reviewing contracts, making slide decks, scheduling, drafting copy, doing junior-level analysis — the inputs and outputs of this work are, by nature, text, structured data, and repeatable process. And that happens to be exactly the shape large language models and agents are best at.
Second, deployment is cheap and fast. Installing a robot arm on an auto line means retooling the production line and a heavy capital outlay. Buying an AI assistant, or standing up an agent, is a SaaS subscription, an API call, a few weeks of piloting. A CEO can say “we improved efficiency with AI” on an earnings call, shareholders understand it instantly, and the layoff rationale writes itself.
Third — and this is the crucial one — replacement happens at the level of the “task,” but companies keep their books at the level of the “headcount.” Almost no role disappears overnight, whole. The real path looks like this:
AI takes over 40% of a team’s tasks → work that used to need 5 people now takes 3 → two years later, headcount has quietly gone from 5 to 3.
If you’re inside it, it feels like “I’m still employed.” But from a macro view, the total number of white-collar jobs is shrinking. That’s exactly why personal experience badly underestimates the true speed of replacement.
3. Why it’s inevitable: this is arithmetic, not emotion
A lot of people treat AI replacing humans as something distant and science-fictional. It’s actually just a cold piece of arithmetic.
Employing a white-collar worker costs a company far more than salary: social insurance, office space, management friction, recruiting, training, mistakes, office politics, turnover risk. An AI system’s cost is a subscription fee, plus a handful of engineers to maintain it, plus reasonably predictable output quality.
Here’s the sentence that matters:
Replacement doesn’t require AI to score 100. It only requires that the cost of AI doing 70-point work be lower than the cost of a human doing 80-point work. Once that’s true, the knife falls.
A lot of white-collar workers overestimate how much the market is willing to pay for their extra margin of quality. In most scenarios, good enough plus cheap wins.
And white-collar work is more of an assembly line than most people want to admit. Look at where the time in a few “respectable” jobs actually goes:
| The title says | What the day actually is |
|---|---|
| Lawyer | Research, summarizing, drafting boilerplate clauses, organizing due-diligence materials |
| Accountant | Reconciling books, categorizing entries, producing reports, filling out tax forms |
| Software engineer | CRUD, bug fixes, writing tests, wiring up APIs, implementing spec |
| Marketer | Ad-creative variants, SEO copy, competitor trackers, weekly reports |
| Consultant | Applying frameworks, writing up interview notes, building decks, running benchmarks |
| HR | Screening resumes, writing job descriptions, running onboarding, answering policy questions |
| Designer | Revisions, resizing assets, templating, assembling materials |
None of this is “low value” — but it’s highly templatable. AI always eats the templatable chunk first, leaving behind signature authority, accountability, relationships, and a handful of real judgment calls. The problem: headcount is counted in whole people. When 80% of someone’s work gets taken over, the company doesn’t keep paying them full salary for the remaining 20% — it has one person absorb the leftover 20% from three people’s jobs.
4. Clearing up the muddy numbers
I’m not going to compute my own precise percentage — that’s what institutions do for a living, and they haven’t agreed among themselves either. But you should at least understand how muddy this water is, and why.
First, remember this: these numbers look like they’re fighting, but they’re mostly not measuring the same thing. Some measure “exposure / impact” (the widest net), some measure “share of work that can be automated” (medium), some measure “jobs that actually disappear” (the narrowest), and some only measure the entry-level slice. Mix them together and you get absurd conclusions — like reading the IMF’s “about 40% of global employment is affected by AI” as “40% will lose their jobs.” This is the single most common, most easily exploited confusion in this whole debate: a large share of “affected” means augmented — made more productive, paid more — not eliminated.
Once the units are sorted, three reference points are enough to build a ruler that spans the whole spectrum:
- The floor that’s already happened (measured, not predicted): Stanford’s Brynjolfsson team found that entry-level white-collar roles with high AI exposure have already shrunk by about 13%. This isn’t “what might happen” — it’s already on the books.
- The official midpoint (the one worth anchoring to): Goldman Sachs analyzed over 800 occupations and landed on a baseline — over the next 10 years, roughly 6%–7% of US employment will be truly displaced by AI, with a range of 3%–14% depending on assumptions.
- The most aggressive ceiling: Anthropic CEO Dario Amodei has warned that AI could eliminate half of entry-level white-collar jobs within 1–5 years, pushing US unemployment to 10%–20%. Take this one with a grain of salt — he’s selling AI, and has an incentive to talk it up — but he’s also one of the people on the planet who understands what these models can actually do.
From 13% to 50%, from “already happened” to “alarming,” notice what’s constant: they’re all arguing about speed and magnitude — nobody is arguing that it won’t happen. That’s exactly why I don’t want to get dragged into a fight over the precise decimal point in this piece: when even people on opposite sides only disagree on degree, chasing the decimal is missing the point.
5. Instead of fighting over the number, learn to take it apart
Institutions hand you a conclusion without showing their work. What’s actually useful isn’t memorizing some percentage — it’s having a ruler in your hand that can take any number apart. So instead of giving you “the number I calculated,” I’ll give you the method for dismantling one.
For any scary big number to actually become “someone lost their job,” it has to pass through three gates — and each one shrinks it substantially:
A scary big number
↓ Gate 1: Task exposure — how much work is technically within AI's reach
↓ Gate 2: Economic viability — of that work, how much is actually cheaper and practical to hand to AI
↓ Gate 3: Hours to headcount — of the hours saved, how much actually becomes an eliminated position
= The real net job loss- Gate 1 (exposure) filters out a first batch: white-collar work really is mostly language, rules, and structured data — AI’s home turf, no argument there — but “within reach” doesn’t mean “will be touched.” The IMF’s “40% affected” figure is measuring exactly this gate.
- Gate 2 (viability) filters out another batch: technically possible doesn’t mean economically sensible. Integration costs, compliance and liability risk, customers who won’t accept it, organizational inertia, legacy systems dragging things down — a large chunk of “things AI could do” gets stuck at “not worth doing yet.”
- Gate 3 (headcount conversion) is the most important cooling valve, and the strongest argument against doomsaying: hours saved don’t convert 1:1 into layoffs. Some people get reassigned to new demand, some become “the same person doing more work,” and organizational friction eats into the rest. Historically, roughly half of the productivity gains automation freed up got absorbed by “so we just did more.”
So keep this ruler handy. Next time someone throws “half of white-collar jobs will disappear” at you, run it through the three gates: that claim requires all three gates to be wide open simultaneously, at maximum — which isn’t realistic. In the real world, a scary number that’s passed through all three gates usually has only a fraction left standing — which is also why I’d rather hand you a direction that “sounds unscary but holds up” than a headline built to scare you.
But here’s exactly what I want to say next: even if the net reduction really is just a fraction, the damage is nowhere close to just that fraction.
6. The real damage isn’t in the word “unemployment”
If hearing “the net reduction is just a small fraction” made you exhale — “okay, so most people are fine” — you’ve just fallen into the second illusion.
“Net headcount reduction” is a mild-sounding total. But what actually reshapes people’s lives is the part that never shows up in that net number, even as their circumstances are turned upside down. Break the jobs apart and, over the coming years, roughly three things happen at once:
- Positions disappear outright or get merged: headcount is genuinely gone, and these people genuinely need to find a new path.
- The role still exists, but the headcount is cut in half and hiring is frozen: a team of 8 becomes a team of 3; you’re still there, but promotion, raises, and any leverage you had are gone.
- The entry door closes: junior roles used to be the pipeline that trained people up into mid- and senior-level talent, and AI replaces exactly the entry-level, standardized, easy-to-supervise work first and most completely. Once the pipe is cut, the path a lot of people assumed was just there — “put in a few years, work your way up” — is simply gone.
Put the first two together and it’s clear: the number of positions genuinely, substantively hit is far larger than that comforting net-reduction figure — it’s a sizable slice of the whole. Which is why looking only at “net headcount” badly underestimates what’s actually happening.
And there’s something with a wider reach, something more people will feel firsthand, than “will I lose my job”: the entire white-collar ladder is being pulled out rung by rung.
Unemployment is acute pain for a minority. But the ceiling coming down is a slow, chronic squeeze almost nobody escapes. Think about the path that used to just exist by default — join a company, put in a few years, get promoted, lead a team, get a raise. That path only worked because there were open slots above you and rungs below you:
- The slots disappear: when a team shrinks from 8 people to 3, the management and senior roles it needed shrink with it — the “one level up” position over your head might simply be eliminated.
- The rungs get pulled up: junior roles get replaced by AI, meaning you no longer have direct reports, no one to hand your grunt work to — you’re likely doing senior-level judgment work and your own junior-level execution at the same time, more work, without your title or pay necessarily moving.
- Your leverage weakens: once a company discovers “fewer people plus AI” also gets the job done, your “give me a raise or I walk” card stops working — because if you walk, that seat may not get refilled at all.
The result: even if you’re never laid off, you’ll probably still find yourself stuck in place — unable to move up, unable to get a raise, and possibly losing ground in real terms against inflation. This isn’t the visible catastrophe of “I lost my job” — it’s the harder-to-name sinking feeling of “I’m working just as hard, but I feel less and less traction, and I can’t see where this goes.” And that second feeling is exactly what more people are experiencing, or about to.
There’s one more layer, harsher still, that acts on every job that’s still alive — basic supply and demand. Positions are shrinking, but the people who’ve been pushed out, plus the people already competing for these same roles, haven’t shrunk in step. So the same category of white-collar job quickly develops an obvious talent surplus: too many candidates chasing too few seats, and price falls naturally.
Which means: even if your job survives, its premium is evaporating. When there’s a long line outside willing to do your job just as well, or even for less, your leverage in any negotiation with your employer disappears — raises get harder to win, and getting lowballed becomes the norm. And the moment you actually leave that job and go back into the market, you hit this wall even more directly: resumes vanish into silence, fewer interviews come back, and the offer you finally land, if you land one, pays less than you expected. Finding the equivalent job, and getting the equivalent pay — two things that used to be taken for granted — both become genuinely, seriously hard.
So for an individual, the real risk exposure was never that comforting total, “how much did net headcount shrink.” Add “positions gone” to “roles severely thinned out,” and for any specific person, the likely outcome is: the seat might still be there, but it’s getting more crowded and worth less every year. This doesn’t contradict “net headcount only fell a little” — one is talking about a total, the other about seats and headroom. The most common ending won’t be the dramatic “AI took my job” — it’ll be the slow-boil version: “I’m still working. I’m just worth less every year, and I have fewer places to go.”
7. This is not a small group’s problem
Percentages are abstract; this needs some sense of scale. I won’t stack up my own pile of “X hundred million people” estimates here — that’s just estimation on top of estimation dressed up as precision. One authoritative reference point is enough: Goldman Sachs estimates roughly 300 million full-time jobs worldwide are exposed to AI automation.
“Exposed” is the widest measure, and far from “unemployed.” But even after shrinking through the three gates above, what’s left is nowhere near a small group — it covers the majority of knowledge-work jobs in developed economies and major Chinese cities. In other words: this isn’t localized pain in some niche occupation. The entire white-collar class is standing on the same downward curve — the only differences are timing and depth.
8. Not every white-collar job suffers equally: by industry
Time to knock down the last, and most comfortable, illusion: the average.
The biggest problem with “the white-collar workforce nets out a small reduction” is this: it averages “barely a scratch” together with “half an industry wiped out,” then hands everyone the same reassurance. But you don’t live in the average — you live in your specific slot. Roughly speaking (these numbers are for scale, not precision): the overall figure might sit around 10%, but if you’re in a high-risk industry, your slot’s number might be 40% or 50%. The gentler the average looks, the more it means someone else is absorbing the half that got averaged away on your behalf.
So the real question isn’t “how much will white-collar work overall get replaced” — it’s “which tier is my slot in.” Replacement doesn’t move in lockstep; sorted by risk, the gap is enormous:
| Tier | Typical jobs | Why |
|---|---|---|
| Hardest hit | Customer service, content/copywriting, data entry and labeling, translation, back-office finance, junior developers, junior legal/audit/investment-banking analysts | Output is standardized, documentable, easy to supervise, easy to replace |
| Moderate | General software engineering, marketing, product/project management, consulting, HR, training, operations/admin | Large chunks of task work get absorbed, roles get “seniorized,” total headcount shrinks |
| Slow, but not immune | Healthcare back office, heavily regulated finance front office, government/state-enterprise clerical work | Regulation and inertia can stall it 3–5 years, but leaner headcount eventually arrives |
| Relatively resilient | Heavy in-person physical presence, heavy sign-off liability, heavy relationship/negotiation work (trial lawyers, operating surgeons, enterprise sales, 0-to-1 founders) | Accountability, relationships, and genuine uncertainty can’t be outsourced to a model |
Looking at it another way — by person, not job title — who’s most at risk:
- New graduates with no moat: no relationship capital, no accountability track record — “AI plus one or two interns” is the cheapest possible replacement for you.
- Mid-career people, roughly 35–45: high salary, work that’s easy to decompose into tasks, and distance from front-line output — cutting them has the best “return on investment.”
- Specialists who only know one tool: once that tool gets absorbed into an AI platform, “I only know Photoshop” or “I only know Excel macros” turns into a commodity anyone can get for free.
- Fully remote white-collar workers: in an employer’s eyes, the most substitutable by either global labor or AI.
- Vendors, contractors, freelancers: the first line item cut when a client trims budget.
The relatively safe people are always the same three types: people who bear real legal or financial liability for outcomes (someone has to pay damages, go to jail, or lose a license when things go wrong), people who control scarce resources (capital, licenses, key clients, political access), and people who can use AI as leverage to produce what used to take a whole team — though this last group, in the end, is tiny.
9. Why I call 2026–2030 the “concentrated outbreak”
Someone will ask: if this is so severe, why hasn’t it happened at scale yet? Because the curve is slow, then fast:
- 2023–2025: The capability exists, companies pilot it, but most layoffs get blamed on “the economic cycle” or “cost-cutting,” and AI’s actual contribution stays hidden.
- 2026–2028: The first wave of “we cut 20% of headcount with AI and the business runs fine” showcase companies appears. This is the trigger — once one company proves it works, competitors copy the playbook fast.
- After 2028–2030: Organizations start rebuilding around smaller teams by default, and the entry-level door narrows as a permanent feature, not a temporary squeeze.
The reason replacement lags isn’t that AI isn’t capable — it’s that organizational change is slow: legacy systems, regulation, inertia, human relationships. But inertia can only delay this, not make anyone immune to it. This slow-then-fast shape is also exactly why I don’t buy the “everything collapses tomorrow” version of doomsaying: it’s happening in stages, one layer at a time.
10. What I’m not saying (don’t skip this part)
To avoid being read as fear-mongering, I need to lay out the honest other half — and I’d argue these points make the earlier argument more credible, not weaker:
First, total employment probably won’t collapse. New roles will appear: AI operations, data, compliance, human-AI collaboration design. But two “buts” apply: the new roles are usually fewer than the ones that disappear, and they carry a higher bar to entry. For an average white-collar worker, the net effect still leans negative.
Second, I might be overestimating the speed. Legacy systems, regulation, unions, organizational inertia could all make the rollout slower than I expect. If you want to push back on me, this is the strongest angle — but note that it only changes the timeline, not the direction.
Third, and crueler than the “total” — it’s the “distribution.” This is the single point I most want you to remember. GDP might keep climbing, corporate profits might keep hitting records, but the median white-collar income could stagnate or fall, while wealth accelerates toward capital and a tiny number of superstar individuals. For most people, the crisis they’ll actually feel isn’t the clean, sharp pain of “I lost my job” — it’s the harder-to-name, chronic sinking feeling of “I’m still working, but I and my work are worth a little less every year.”
A closing note: the number can be argued, the direction can’t
If there’s exactly one thing you take from this piece, it shouldn’t be any specific percentage — every institution has its own, and I’m not here to referee that fight. What you should actually remember is this:
Whatever the final number turns out to be, the direction is no longer in doubt: over the next 5, 7, even 10 years, AI will drive a real, substantial decline in both the total number of white-collar jobs and the income levels of the people who hold them. This is close to predictable at this point — and it’s becoming a consensus across very different camps.
That’s the part that matters. When the IMF, Goldman Sachs, the World Economic Forum, AI company CEOs, and independent researchers — using completely different methods, standing in completely different positions — all end up pointing at the same direction, and the only disagreement left is “how fast” and “how much,” with nobody arguing “it won’t happen,” this stops being one person’s alarmism and becomes a fact already sitting on the table. When even people on opposite sides of an argument only disagree on speed and magnitude, and nobody disputes the direction itself, the most dangerous posture is pretending it isn’t there.
So I know this piece is uncomfortable to read. But I’d rather you feel uncomfortable now than keep walking forward holding onto a few illusions that are already failing:
- The illusion of the “iron rice bowl” — slow doesn’t mean immune.
- The illusion of “just put in a few more years and I’ll be safe” — the pipeline you’re counting on to carry you up is being cut off at the entrance.
- The illusion of “I’ll go get a certification, learn a standardized skill” — standardized, certifiable ability is exactly the kind AI flattens first.
As for what to do once you’re clear-eyed about this — where to stand, what to build, what’s genuinely scarce in the AI era — that’s a different piece. This one, I only owe you the wall coming down, so you can see what’s actually behind it.
The institutional numbers cited here (Stanford, Goldman Sachs, the IMF, Amodei, and others) are all public estimates — anchors for calibrating a range, not precise prophecy. What I’m betting on in this piece is the direction, not any particular decimal point. I’m glad to argue with you about any one of the three gates — just don’t wave it away with “this is just fear-mongering.”