There’s a story going around about artificial intelligence and work. You’ve probably heard it; AI will hollow out the middle tier of knowledge jobs (squeezing analysts, paralegals, junior consultants) while leaving plumbers and CEOs more or less intact. The bottom does physical work that machines can’t replicate. The top makes judgment calls machines can’t match. The middle generates and processes information, which is precisely what large language models are built to do. Clean logic, but I believe it is almost certainly wrong.

Wrong in ways that will lead companies to make bad decisions, governments to write bad policy, and individuals to spend the next decade preparing for the wrong future.

The problem is that AI doesn’t hit tiers, it hits tasks. Every level of work, from the shop floor to the C-suite, includes tasks that machines can now perform reasonably well, alongside tasks that become more valuable the moment AI gets involved. There’s a fault line running through every job in every industry, and it is going to matter a lot which side you sit on.

That sounds ominous, but it isn’t. The capabilities AI can’t touch are the ones we’ve always known matter most: the ability to walk into a room and read it, to sit with a mess until it starts making sense, to hold two contradictory truths in your head long enough to find a third option. We’ve spent the last few decades treating those skills as soft, secondary, the stuff you mentor into people after they’ve proven they can crunch numbers. AI is about to remind us that they were never secondary at all.

The Comfortable Narrative and Why It Sells

The middle-squeeze argument isn’t stupid. GPT-4 passed the Uniform Bar Exam in the top 10% of test-takers [1]. It can churn out working code, legal summaries, and financial models on demand. Research suggests that AI could replace more than 50% of the tasks performed by market research analysts and 67% of those performed by sales representatives, compared with just 9–21% of the tasks handled by their managerial counterparts [2]. A significant portion of early-career knowledge work involves pattern execution, and pattern execution is now inexpensive. Those who noticed this and raised the alarm are correct about the mechanism, but they’re wrong about the target.

The three-tier model treats “the bottom” and “the middle” as a single category. It draws clean lines between categories when the actual economy is a sprawl of overlapping roles, in which a nurse exercises more real-time judgment in a single shift than a business analyst does in a month, and where a “senior” manager might do nothing all day that a well-crafted prompt couldn’t handle. The model is a diagram, not a map, and people who navigate by diagrams end up lost.

Forget the tiers for a minute. Look at what’s happening on the ground—data entry clerks, gone. Call centre operators are going. Basic bookkeepers, filing staff, first-line customer service agents, transcription typists, and scheduling coordinators. These aren’t middle-tier knowledge workers. By any honest measure, these are bottom-tier roles. And they’re vanishing not in some speculative future but this year, last year, the year before. The World Economic Forum’s latest Future of Jobs report found that clerical and secretarial roles are expected to experience the largest absolute decline among job categories over the next five years [3]. Independent analyses indicate that data-entry keyers face a 95% risk of full automation and project a 25% decline in those jobs by 2031 [4]. The bottom of the office economy is already washing away.

A company paying 200 offshore data processors $15,000 each faces a $3 million annual wage bill for work that is entirely pattern-based. The return on replacing that with AI is not a close call. There’s no agonising boardroom debate. No one writes think pieces about data entry clerks. The contracts are not renewed. The seats quietly empty. Studies show companies keep AI’s role in layoffs under the radar to avoid bad press [5]. The bottom falls out, and hardly anyone in power notices, because the people affected have never been visible enough to matter in the narrative.

Compare that with replacing a junior analyst earning $70,000 at a big consulting firm. The work might be partially automatable, sure. But the optics are different. Those junior analysts are the children of the decision-makers. They went to the same universities. Social friction slows the replacement. The bottom tier has never had that protection.

This is the part the popular narrative glosses over because it’s less interesting to write about. How AI threatens paralegals makes a provocative headline. How it threatens people already earning close to nothing is just uncomfortable. However, the scale of bottom-tier displacement is enormous; it’s occurring faster and affecting far more people. In the U.S. alone, call centres employ roughly 3 million workers [6]. Far more than the combined total of paralegals and junior consultants. One bipartisan bill introduced in Congress this year, the Keep Call Centres in America Act, explicitly cited AI automation as a threat to those millions of jobs [6]. If you want to know where the pain is sharpest right now, look down, not across.

The Fracture Line

Here is what I think is actually going on. Every tier of work, bottom to top, is splitting along a single fault line. On one side: work that is fundamentally about processing information or executing well-defined patterns. On the other hand, work that requires physical presence, human connection, contextual judgment, or creative synthesis. AI wipes out the first kind. It makes the second kind better.

At the bottom of the org chart, the split is stark. A data entry clerk is on the pattern side. Gone. A disability support worker is on the human side. Not going anywhere. Their work improves because AI can handle the crushing paperwork that currently consumes a third of their shift, freeing them to do the work for which they were hired: being present with a human being who needs them. Attentive, responsive, adaptive in ways no algorithm touches.

A call centre operator reading from a script? Pattern side. Gone. An aged-care nurse who reads a patient’s face and realises something is wrong before the monitors catch it? Human side. Better equipped than ever. A filing clerk? Gone. A tradesperson standing in a 1940s bathroom trying to figure out how to run new plumbing through walls built before standards existed? Irreplaceable, and now faster at quoting and scheduling because AI handles the back-office burden.

The bottom tier is not destroyed uniformly. It fractures. The pattern work disappears. Human labour becomes scarcer and more visibly essential as commoditised routine vanishes around it. Whether that translates into better pay is a political question, not a technological one. However, the value is becoming increasingly difficult to ignore.

Ford CEO Jim Farley has been making this case publicly. He calls it the “essential economy”: all the jobs that involve moving, building, and fixing real things in the real world. The U.S. has a massive shortage of workers in these skilled trades, with hundreds of thousands of unfilled roles in manufacturing and construction even as AI advances [7][8]. Farley’s point is plain: we’ve overemphasised the white-collar career track while demand surges for people who actually make and maintain the physical world. AI isn’t about to weld a pipe or rewire a house or hold an older adult’s hand. The AI boom itself needs human workers to build and maintain the data centres that power it [8]. The human side of the economy isn’t shrinking. In many areas, there’s a desperate need for more of it.

The Middle Is Not One Thing

The same divide runs through the middle tier, and here the popular narrative comes closest to being right while still missing the point.

Yes, a lot of middle-tier knowledge work is pattern processing dressed up in business casual. The analyst who builds the same spreadsheet model every quarter with different numbers plugged in. The marketing coordinator follows a templated content calendar. The project manager is responsible for maintaining the Gantt chart and for sending status emails. These roles are badly exposed. Gartner predicts that by the end of 2026, 20% of large companies will use AI to flatten their management structures, potentially eliminating more than half of current middle-management positions [9]. When an AI can shuffle data and send updates instantly, the person whose job was to shuffle data and send updates has a problem.

But sitting right next to those people, sometimes sharing the same title, are workers doing something completely different. The manager who walks into a room where two teams are at war and walks out with everyone aligned. The strategist who looks at the same data as everyone else and spots the one insight no one else saw. The team lead, who knows Sarah is about to quit before Sarah even realises it, because they pay attention to humans rather than dashboards.

AI does not displace these people. They get rocket fuel. The strategic thinker who used to spend 60% of their time gathering and organising information can now do it in minutes. The hours they get back go to the high-level thinking that has always been their real contribution, the contribution that admin has crowded out for years. The good manager who wasted half their week on reporting and scheduling now has that time back to develop people, build culture, and make calls that require reading the room rather than a spreadsheet.

We’ve had an entire generation of talented people buried under busywork. Strategic minds drowning in admin. Natural leaders spend their days on process management rather than people management. AI doesn’t just threaten the weak middle. It strips away the noise and lets the signal through. Companies that understand this will restructure around it. They’ll retain their best managers and strategists and provide them with better tools. They’ll stop replacing the template-followers and process-runners. The org gets thinner in the middle, yes, but not uniformly. It thins on the pattern side. It depends on the judgment side. That’s not a hollowing out. That’s an upgrade.

The Top Stays, But Not for the Reasons People Think

The standard argument says the top tier is safe because leadership can’t be automated. True enough, as far as it goes, but it’s also flattering. Much of what passes for senior leadership isn’t visionary. It’s sitting in meetings, thumbing through decks someone else built, signing off on recommendations someone else made, and doing expensive hand-waving that a decent prompt could simulate.

The fault line doesn’t stop at the C-suite door. The CEO making novel strategic decisions under genuine uncertainty, sensing a market shift before the data confirms it, inspiring thousands of people to change direction? Untouchable. The senior VP whose main function is relaying information between floors and who hasn’t had an original thought since 2019? Exposed. The timeline is longer. The severance is better. The fracture is the same.

Amazon CEO Andy Jassy recently told employees straight: as the company uses more AI, “we will need fewer people doing some of the jobs that are being done today,” and he expects the corporate workforce to shrink [10]. That’s not because AI has become CEO. This’s because much high-level office work is more routine than is acknowledged.

The gap between genuine leaders and title-holders is about to become impossible to hide. A great CEO with AI tools can process more information, test more scenarios, move faster, and spend less time on the operational grind that used to eat up their calendar. AI doesn’t replace them. It reveals, with brutal clarity, which senior people were actually thinking and which ones were just occupying the chair.

The Apprenticeship Problem Is Real

One thing the middle-squeeze crowd gets genuinely right is the apprenticeship crisis.

For decades, grunt work has served a double purpose. It completed the work and trained personnel. The junior lawyer doing document review wasn’t just reviewing documents. She was learning how contracts actually work, developing an instinct for what looks wrong, building the foundation that would make her a good senior lawyer in ten years. The cub reporter writing community briefs was learning to find sources and work under deadline pressure. The junior coder fixing bugs was learning the codebase from the bottom up.

When AI takes those tasks, the training pipeline breaks. You can’t skip straight to judgment. Judgment grows from experience, and experience comes from doing the work. A senior risk manager who has never spent years in the weeds with actual data may lack the intuition to distinguish a real problem from statistical noise. The AI will flag anomalies throughout the day. Knowing which ones are nothing and which are catastrophes requires the kind of intuition you only build by doing the work we just automated.

We’re already seeing early signals. Stanford researchers found a 13% decline in new entry-level job postings in AI-affected fields since late 2022, characterising it as the fastest and broadest shift in early-career labour dynamics in years [11]. LinkedIn’s data confirms it: generative AI tools are handling many of the simple coding, debugging, and administrative tasks that used to fall to new graduates [12]. Anthropic CEO Dario Amodei warned that AI could eliminate half of all entry-level white-collar jobs, potentially pushing unemployment to 10–20% [14]. If those graduates cannot secure senior roles, who will fill them in ten or twenty years?

The solution is not to pretend AI doesn’t exist and keep armies of juniors busy with busywork for the sake of tradition. That’s nostalgia dressed up as training policy. The solution is to redesign early-career roles so that AI handles the mechanical aspects while the junior person addresses judgment calls under supervision, earlier than before. Let the new analyst use AI to gather and process data, but make their real job interpreting it. Have them present to the senior team in month three rather than month eighteen. Let the junior lawyer rely on AI for document sifting and spend their time analysing tricky cases alongside a partner.

Think about what this actually means. The traditional model required people to spend years on mechanical tasks and labelled it training. Much of it was just cheap labour disguised as development. The new model, when implemented correctly, places people in the deep end of judgment work from the start, with AI as their research assistant and a senior professional as their guide. That’s not a loss. That’s an acceleration of human capability. Some law firms are already experimenting, moving away from the billable-hour model for junior associates and toward hands-on apprenticeships in which AI handles document preparation [13]. It’s early. Most companies are just cutting headcount and hoping the pipeline sorts itself out. It won’t. But the blueprint for something better is right there.

The Historical Pattern That Actually Matters

People love bringing up the Luddites, the printing press, and the steam engine. The parallels are real enough. Technology destroys specific jobs, creates new ones, the species adapts, and nobody mourns the lamplighter. But the parallel that matters most concerns which jobs were destroyed. It concerns the fate of the survivors.

Every single time, they got more human.

When ATMs rolled out, teller jobs didn’t vanish. ATMs made it cheaper for banks to open branches, so they opened more of them, and teller employment actually rose [17][21]. The job changed. Tellers stopped counting cash and began conducting relationship work: advising customers, solving problems, and building trust. When spreadsheets killed manual calculation, accountants didn’t disappear either. The ones who thrived used the freed-up time for deeper financial insight. When the camera arrived, painting didn’t die. It was liberated. Freed from the obligation to reproduce reality, painters went on to create Impressionism, Expressionism, and Abstraction—an explosion of creativity, released by the very technology that was supposed to kill the art.

This is the pattern. Not “his tier lives, that tier dies.” Within every field, the mechanical fraction shrinks, and the human fraction expands. The question for each person, each role, each company is always the same: how much of the work was mechanical, and how much was human? What’ss different this time is the scale of the amplification. The camera freed painters from realism, but it didn’t hand them better brushes. The spreadsheet freed accountants from arithmetic, but it didn’t help them think more clearly about strategy. AI does both. It removes the mechanical burden and upgrades the tools available for the remaining human work. A great doctor with AI-assisted diagnostics doesn’t just do less paperwork. They catch things they would have missed. A great architect with AI modelling doesn’t just draw faster. They explore possibilities they never would have had time to imagine. The mechanical capabilities get wiped out, yes. But for people on the human side of the line, the tools for doing the work that matters have never been this good.

The Distribution Problem We Have to Solve

When manufacturing jobs moved to cheaper countries, the work still went to people. A laid-off factory worker in Ohio meant a newly hired factory worker in Vietnam or Mexico. Money continued to flow into a community’s payroll. Globalisation was painful and uneven, but economic energy stayed in the labour market.

AI doesn’t work like that. When it replaces a team of data processors or a platoon of call centre agents, the value doesn’t flow to another group of workers. It flows upward. Into margins. To shareholders. Into the senior people overseeing a learner operation. The work vanishes from the labour market entirely. MIT economist Daron Acemoglu has warned that the digital revolution since 1980 has already contributed to the hollowing out of middle-class jobs and the widening of inequality, and that the unchecked adoption of AIrisks repeating that pattern on a more extreme scale [18].

Companies are finding it easiest to use AI to replace outsourced and offshore labour first, then pocket the savings. One firm saved $8 million annually by spending $8,000 on an AI tool, largely by eliminating outsourced roles [20]. That’s spectacular efficiency if you own the company. It’s a disaster if you were doing the work. The solo founder running a business that once required a team of twelve is real and growing. Exciting if you’re the founder. Less exciting if you zoom out and notice the economy just lost eleven jobs, with the revenue going to one person and some cloud-compute fees.

“Upskill” is the word everyone reaches for. And people should learn new things. But not everyone can climb the value chain simultaneously. There isn’t room at the top for every displaced middle manager, just as there wasn’t room in the middle for every displaced clerk.

This problem is solvable. It isn’t a law of physics. It’s a distribution question, and distribution questions have policy answers. Tax structures that don’t penalise human labour while subsidising automation. Education funding that aligns with where jobs are going. Proper valuation of care work, which has been scandalously underpaid for decades, is about to become one of the most important sectors in the economy. Social safety nets that treat transition as an investment rather than a handout. Acemoglu and others have proposed concrete measures: reducing taxes on human labour, increasing taxes on AI-driven capital gains, and building stronger support systems for workers in transition [21]. These aaren’tanti-technology positions. They’re pro-human positions. The productivity gains from AI are substantial. The question is whether we build systems that spread those gains or let them pool at the top. That’s a choice. We’ve made it badly before. We have a chance to make it better this time, if we move fast enough.

So What Is Actually True

AI will hit the bottom tier first and hardest in raw numbers. Data entry, call centres, routine processing. This is happening now, accelerating. In banking alone, analysts project that up to 200,000 jobs could be eliminated over the next few years as AI automates back-office and customer-service tasks [22]. The middle-squeeze crowd isn’t wrong that middle-tier knowledge work is in the firing line, but they’re wrong that it’s the primary target. The bottom is where cost pressure is simplest and organisational friction is lowest.

Within each tier, the division lies between pattern work and human work. Pattern work is automated regardless of whether it sits at a higher tier. The nurse gets freed from paperwork. The tradesperson gets faster at admin. The good manager actually gets time to listen and manage. The great strategist gets tools that match the quality of their thinking.

The displaced individuals were, in most cases, already employed in jobs vulnerable to cost pressures, including offshoring, outsourcing, and routine efficiency gains. AI accelerates and sharpens that displacement. That’s not a reason to be indifferent to their situation. It’s a reason to build better systems to support them. And “things will work out eventually” is not a strategy.

The historical pattern holds. The internet didn’t kill books. Streaming didn’t kill television. AI will change industries, not erase them.

This brings us to the conversation about what we teach and why we keep getting it wrong. AI literacy? Sure. Teaching kids to code? Fine. But the tools are the easy part. Any smart kid can learn to prompt ChatGPT or write some Python. What we desperately need to teach, and what most education systems are catastrophically bad at, is the stuff on the human side of the fault line. How to think through ambiguity. How to communicate under pressure. How to lead a group of people who disagree. How to care about the work and the people affected by it. How to ask the question nobody else thought to ask.

When global employers were surveyed about the most important skills for the coming years, the top responses were analytical thinking, resilience, flexibility, and leadership [15]. Not narrow technical abilities. Harvard Business Review found that foundational skills like collaboration, adaptability, and problem-solving are likely to matter more for long-term career success than specialised technical knowledge [16]. Learning how to work with humans and navigate complexity beats learning the latest framework that an AI will probably handle by the time you graduate.

We’ve spent years developing these “soft skills.” The label was “always ” wrong. They’re the hardest to acquire, the hardest to fake, and, in the age of AI, the hardest to replace.

What’s Left

If you’re trying to work out where you stand, stop thinking about your tier and start thinking about your tasks. Pull apart your working week. How much is pattern execution: following checklists, applying formulas, turning the crank on a routine process? How much is genuine judgment, human interaction, creative problem-solving, or physical skill? That ratio tells you more about your future than your job title ever will.

When you strip away the pattern work, when you remove the busywork and the checkbox-ticking and the information-relaying that fill so many working days, what’s left is what people actually care about—solving real problems. and connectingwith other humans. Building something that matters. Making a call that takes courage. Teaching someone younger how to think. That’s the other side of the fracture. Not a diminished, residual existence in which humans do whatever the machines can’t be bothered to do. The opposite. The work that remains is the work we’ve always known we have. We couldn’t get to it; we were buried under the other stuff.

The machines are becoming more sophisticated, and that won’t stop. The question was never whether AI would change work. The question was whether we’d be ready for the work that actually matters once the machines took over the work that didn’t.

Sources

Did n’tenAI, “GPT-4 Technical “report, 2023. GPT-4 scored in the 90th percentile on the Uniform Bar Exam.

[2] Brookings Institution analysis of AI task exposure across occupational categories.

[3] World Economic Forum, “Future of Jobs R”port 2025.”

[4] Will Robots”Take My Job? Bureau of Labour Statistics projections for data entry keyers.

[5] Studies on corporate AI adoption and workforce communication strategies.

[6] U.S. call centre employment data; Keep Call Centres in America Act, 2025.

[7] Jim Farley remarks on the “essential emeconomy and skilled trade shortages.

[8] National Association of Manufacturers: construction and data centre labour-shortage data.

[9] Gartner forecast on AI-driven management restructuring, 2024.

[10] Andy Jassy’s internal memo to Amazon employees on AI and workforce planning.

[11] Stanford Digital Economy Lab, entry-level job posting analysis, 2024.

[12] LinkedIn Workforce Report on generative AI and early-career task displacement.

[13] Legal industry apprenticeship model shifts reported in ABA Journal and The American Lawyer.

[14] Dario Amodei’s interview remarks on AI and entry-level employment.

[15] World Economic Forum employer survey on workforce skills, Future of Jobs Report 2025.

[16] Harvard Business Review, “The Skills That “ill Matter Most in the Age of AI.”

[17] U.S. Burea” of Labour Statistics, bank teller employment data 2000–2018.

[18] Daron Acemoglu, various publications on automation, inequality, and labour market hollowing.

[20] Case study on AI replacing outsourced operations; widely reported in business media.

[21] Acemoglu & Johnson, “Power and Progr”ss,” 2023; ATM/telle” employment research.

[22] Bloomberg Intelligence: banking-sector AI workforce projections.