In the late 1990s, three economists, David Autor at MIT, Frank Levy, and. Richard Murnane at Harvard, sat down with a question that turned out to be surprisingly hard to answer. They wanted to know which jobs would survive the coming wave of computerization and which would not.

Their conclusion, published in the Quarterly Journal of Economics in 2003, became one of the most cited frameworks in labor economics. The jobs that survive technological disruption, they argued, are not the ones that require the most education or the highest salaries. They are the ones that require the things computers cannot do.

At the time, that meant anything requiring physical dexterity, social intelligence, or creative judgment. The assembly line worker was vulnerable. The surgeon was not. The file clerk was vulnerable. The therapist was not.

What Autor, Levy, and Murnane could not have anticipated was how fast the line would move.

The question everyone is asking right now is the wrong question.

Will AI take my job is not a useful frame. It assumes jobs are static objects that either survive or disappear whole. They are not. Jobs are collections of tasks. Some of those tasks are vulnerable to AI replacement. Some are not. The question worth asking is not whether AI will take your job. It is which parts of your job AI is already doing better than you, which parts it will do better than you within three years, and what is left when you subtract those parts from the whole.

That subtraction tells you something important. If what remains after the subtraction is a thin slice of your current role, you are more exposed than you think. If what remains is the core of what makes your work valuable, you are more resilient than the headlines suggest.

There are four characteristics that make a task vulnerable to AI replacement. Not a job. A task. Keep that distinction in mind as you read.

The first is routine. If a task follows a predictable pattern, if it involves applying the same rules to similar inputs and producing consistent outputs, AI can do it. Not eventually. Now. Standard contracts. Routine data analysis. Expense reports. First draft emails. These tasks are already being automated at scale and the pace is accelerating.

The second is volume without judgment. If a task requires processing large amounts of information quickly but does not require deciding what to do with that information, AI can do it faster and more accurately than any human. Document review. Market research compilation. Financial modeling. The processing is the easy part. The judgment about what the processing means is harder.

The third is replication from examples. If a task produces an output that can be evaluated against past examples, AI can learn to replicate it.Writing that follows a formula. Code that solves a defined problem. Translation. Transcription. Anything where the quality of the output can be measured against a clear standard.

The fourth is low context dependency. If a task can be completed without understanding the full human context surrounding it, AI can handle it in isolation. Answering a standard question. Generating a standard report. The less a task depends on understanding the full picture of a specific human situation, the more replaceable it is.

Now here is what AI cannot do. At least not yet. And not well.

It cannot exercise genuine judgment in genuinely novel situations. It can pattern match against past situations with remarkable sophistication. It struggles when the situation has no real precedent. The lawyer handling a case with unusual facts. The doctor treating a patient with an atypical presentation. The manager navigating a team conflict with roots in a specific organizational history. These require judgment that cannot be reduced to pattern recognition.

It cannot build and maintain real human relationships. It can simulate them with increasing plausibility. The relationships that matter in professional life, the ones built on years of shared context, mutual trust, and genuine personal investment, are not simulable at the level that counts.

It cannot synthesize across genuinely different domains in ways that produce original insight. The cross-domain synthesis that produces a genuinely new way of seeing a problem remains stubbornly human.

And it cannot be accountable. Someone has to be responsible. Someone has to be able to say I made this decision and I stand behind it. AI can inform decisions. It cannot own them.

Look at your current role and ask four questions.

What percentage of my day involves tasks that are routine, high volume, replicable from examples, or low context? If that percentage is above fifty, your exposure is significant and growing.

What is left when those tasks are removed? Is what remains a thin slice of coordination and oversight? Or is it the core of what makes your work genuinely valuable?

Are the parts of my role that require judgment, relationships, cross- domain synthesis, or accountability growing or shrinking? If they are shrinking it is not because AI is making your role more efficient. It is because your role is being redefined around you.

And finally: if an AI system did the vulnerable parts of my job tomorrow, what would my employer actually need me for?

That last question is the one most people avoid. It is also the most important one to answer before someone else answers it for you.

Artificial intelligence is going to create roles that do not exist yet. Every major technological shift in history eliminated some work and created other work that the people living through the shift could not have predicted. What is visible right now is the direction. The work that survives and grows is the work that is most deeply human. Judgment. Relationships. Synthesis. Accountability.

That is not a reason for comfort. It is a reason for urgency. Those capabilities do not develop by accident. They develop through deliberate choices about where you invest your professional energy.

The question is not whether your job is next. The question is what you are doing about it today.

Forwarded this? Subscribe free at replaced.co. If it made you think, send it to one person who needs to read it.

Further reading: The Autor, Levy, and Murnane paper that introduced the task- based framework for understanding which jobs survive technological change. Published in the Quarterly Journal of Economics in 2003, it remains one of the most cited papers in labor economics and is the intellectual foundation for how serious researchers think about AI replacement today. Read it here →

There is a particular kind of advice that older professionals give to young people trying to break into the workforce. It is well intentioned. It is almost always delivered with genuine conviction. And right now, in the summer of 2026, it is dangerously out of date.

Take the entry level job. Get your foot in the door. Do not worry about the salary. You are not there for the money. You are there for the experience. Do the work nobody else wants to do, learn the systems, prove yourself, and the rest will follow. Every career is a ladder. The bottom rung is where you start. The advice is not wrong. It describes exactly how careers worked for the last fifty years. It describes how the people giving it built their own careers. It is handed down in good faith from one generation to the next because it worked, reliably and repeatedly, for decades.What it does not account for is what happens when the bottom rung is removed.

In 2024, something started happening to the software industry that nobody talked about loudly enough to matter. Entry level developer jobs started disappearing. Not because there were fewer companies building software. Not because the industry was contracting. Because AI coding tools had reached the point where the tasks that used to require a junior developer, debugging, writing boilerplate code, handling routine tickets, could be handled faster and cheaper by a large language model with a senior engineer supervising it. The junior developer was not replaced by another developer. The junior developer was replaced by a tool that did not need health insurance, did not need onboarding, did not need a desk.

Senator Bernie Sanders recently cited in Fortune that software developer employment for workers aged 22 to 25 had fallen roughly 20 percent from its 2024 peak. Twenty percent. In two years. In the industry that was supposed to be the safe harbor. The place young people were told to go to build a career that AI could not touch.

The entry level job was never really about the work. That is the thing most people who give career advice forget to say out loud.

Yes, the work mattered. You learned things. You built skills. But the deeper function of the entry level job was something else entirely. It was the mechanism by which a person with potential but no track record could prove they had both. It was the on ramp to every job that came after it. You could not get the mid level job without the entry level job because the mid level job required experience and experience required the entry level job and that circular logic held together for fifty years because the entry level job existed.

Remove it and the circle breaks.

The twenty two year old who graduates in May 2026 with a computer science degree does not just face a difficult job market. She faces a job market that has restructured itself around the assumption that the work she was trained to do can be done more efficiently without her. She has the credential. She has the knowledge. She does not have the one thing every subsequent employer will require before they consider her.

She does not have the experience. And there is no longer a clear path to getting it.

The missing rung is not a software problem. It is the first visible sign of something that is going to move through nearly every industry that employs young people in entry level roles.

Marketing coordinators. Junior analysts. Research assistants. Editorial assistants. Customer support specialists. Administrative coordinators. Every role that exists to give someone with potential their first real professional experience is a role that AI replacement is evaluating right now for the same reason it evaluated the junior developer. Can this be done faster and cheaper without a human being at the bottom of the organizational chart?

In many cases the answer is increasingly yes. And when the answer becomes yes at scale, the ladder does not just get harder to climb. It loses its bottom rung in nearly every industry simultaneously.

Artificial intelligence is going to create extraordinary opportunities for the generation entering the workforce right now. The tools available to a twenty two year old today are more powerful than anything a thirty year old had access to five years ago. The ability to learn faster, build faster, and create things that previously required an entire team is genuinely remarkable. AI is not the enemy of the young professional. Used correctly it is the most powerful career accelerator in history.

But none of that changes the immediate problem. The path that previous generations used to build the experience that careers are built on is being removed faster than new paths are being created. The tools are extraordinary. The on ramp is gone.

So what does the person standing at the bottom of a ladder with a

missing rung actually do?

That is the question this newsletter is going to spend a lot of time on. Not the thirty year view. The Monday morning view. What do you do right now, today, when the entry level job that was supposed to be your first step does not exist anymore.

We are going to figure that out together.

Forwarded this? Subscribe free at replaced.co. If it made you think, send it to one person who needs to read it.

Further reading: The Fortune article mentioned above on Bernie Sanders, AI ownership, and the proposed American AI Sovereign Wealth Fund: Read it here →

In 1914, Henry Ford did something that confused almost everyone who worked for him.

He doubled their wages.

His competitors thought he had lost his mind. His shareholders were furious. The newspapers called it an experiment in social engineering.

Ford called it common sense. He was paying his workers $5 a day, roughly double the going rate for factory labor. Ford understood that mass production required mass consumers, and his famous wage increase became the clearest symbol of that idea. You cannot build things faster than people can afford to buy them. The factory only works if someone can afford what it makes.

That is not a story about generosity. It is a story about a man who understood something fundamental about how an economy actually works. Something so obvious that it takes a certain kind of genius to see it clearly while everyone around you is missing it entirely.

The economy is not a machine that produces things. It is a loop that circulates value. Money flows from companies to workers in the form of wages. Workers flow that money back to companies in the form of purchases. The loop runs. Everyone benefits. The moment you interrupt the loop the whole system starts to break down in ways that are slow at first and then suddenly catastrophic.

Henry Ford understood the loop. He understood that his factory was only valuable if someone could afford to buy what it made. He understood that the workers standing on his assembly line were not just a cost to be minimized. They were the customers who justified the existence of the assembly line in the first place.

That insight is now one hundred and twelve years old.

And we have forgotten it entirely.

Here is a scenario that is happening in boardrooms across America right now. A CEO sits across from her chief technology officer. He shows her a presentation. Slide three is the one that matters. It shows the current cost of processing a customer service inquiry with a human employee versus the projected cost of processing that same inquiry with an artificial intelligence system. The human costs $18 per inquiry when you factor in salary, benefits, training, and management overhead. The AI costs $0.04.

She approves the transition. The stock goes up. The quarterly earnings call goes well. Fourteen hundred people working in customer service were replaced. This is not a story about one CEO or one company. It is a story about a system that was designed to optimize for one number. Shareholder value. That system is working exactly as intended. The problem is not the people making the decisions. The problem is that nobody designed the system to ask the question that Henry Ford asked in 1913. Who is going to buy the things?

The American economy is, at its core, a consumer economy. Seventy cents of every dollar of economic activity in this country comes from people spending money on things. Food. Housing. Cars. Phones. Clothes. Healthcare. Entertainment. The number is so large it is almost abstract.

Seventy percent. That means manufacturing, government spending, exports, and every other economic activity you can think of account for the remaining thirty cents. The consumer is not a participant in the American economy. The consumer is the American economy. And the consumer has a job. For now. Which means the most important business insight of the last century still applies today.

The $0.04 per inquiry calculation does not account for that. The quarterly earnings model does not account for that. The stock price does not account for that. But Henry Ford accounted for it. In 1914. Standing next to an assembly line in Detroit.

The fourteen hundred people who were replaced were not just a cost center on a balance sheet. They were seventy cents on the dollar. They were the people buying the groceries and paying the rent and going out to dinner on Friday nights and purchasing the phones and the cars and the streaming subscriptions that constitute the economic activity of their communities.

When they lose their jobs they stop doing those things. Not immediately. There is a savings account. There is severance. There is the hope of another position. But eventually, if the AI replacement is large enough and sustained enough, the spending stops. And when the spending stops the businesses that depended on that spending start to struggle. And when those businesses struggle they cut their own costs. And the first cost they cut is labor.

The loop breaks. Slowly at first. Then faster. None of this is an argument against artificial intelligence. The technology is extraordinary. What it has the potential to do is genuinely remarkable and it is going to keep getting more remarkable. AI is already identifying cancers that human radiologists miss. It is accelerating drug discovery by decades. It is solving engineering problems that stumped the best minds in the world for generations. The promise is real and it is enormous and some of it is going to be beautiful.

The question is not whether we build it. We are building it. The question is whether we build the economic infrastructure alongside it that ensures the people it replaces have somewhere to go and something to eat while it does the extraordinary things it is capable of doing. So what happens to the economy when the people who were supposed to buy the things cannot buy the things?

Henry Ford knew the answer.

We just seem to have forgotten.

Forwarded this? Subscribe free at replaced.co. If it made you think, send it to one person who needs to read it.

Further reading: The Henry Ford museum's own account of the January 1914 announcement, including the original board documents and the competing theories for why Ford actually did it. Read it here →

On May 11, 1997, the greatest chess player who ever lived sat down in a Manhattan office tower and resigned after nineteen moves.

Garry Kasparov had held the world championship for twelve consecutive years. Expert commentators predicted an easy win when IBM challenged him to a rematch against their chess computer, Deep Blue. One said he would win every game.

The final game lasted less than one hour. Deep Blue sacrificed a knight on the eighth move. Kasparov never recovered. At move nineteen he pushed back his chair, stood up, and walked off stage. It was the first time in his career he had resigned without exhausting every possible defense.

He said afterward: “For the first time in the history of mankind, I saw something similar to an artificial intellect.”

Something similar. Not the thing itself. Because what Kasparov had just lost to was not intelligence. It was something that looked, from the outside, exactly like intelligence, but was built from something else entirely.

Deep Blue did not actually understand chess.

It evaluated 200 million board positions per second using 32 processors running in parallel. For every possible move, it calculated the likely responses, then the responses to those, branching dozens of moves ahead, assigning a numerical score to each possible future and selecting the move with the highest score.

It did not know what a knight was. It had no sense of what it meant to win or lose. It processed positions and selected moves, faster than any human brain could follow, across more possible futures than any human could hold in mind simultaneously.

That was enough. Against the greatest chess player who ever lived, it was more than enough.

This is what AI actually is. A pattern engine, operating at a scale no human can match. When you ask a modern AI system to write a paragraph, analyze a contract, read a medical scan, or generate a financial model, it is doing the same thing Deep Blue did. It has processed vast amounts of human output, identified the patterns in that output, and learned to produce outputs that match those patterns. It does not understand what it is doing. It recognizes patterns with a precision and speed that makes understanding unnecessary.

Every major technology company, every major government, and most of the largest pools of capital in the world are currently racing toward one specific destination.

Artificial General Intelligence.

AGI is not a faster chess computer. It is the point at which a single system can learn and perform any cognitive task a human can perform, across any domain, without being specifically trained for each one. A system that can diagnose the patient, argue the case, design the building, compose the music, negotiate the contract, run the experiment, manage the project, and teach the class. Every job that exists because it requires a human being to think.

When Deep Blue beat Kasparov it proved that a machine could master one cognitive domain completely. AGI is the point at which a machine can master all of them.

And here is the part that makes AGI different from every previous wave of replacement in history.

Once it exists, it can be copied. Instantly. At essentially zero marginal cost.

The steam engine took decades to spread from mill to mill. The assembly line required factories to be built city by city. Every previous wave of automation was constrained by the cost and time required to deploy it at scale. The replacement was real but it arrived gradually enough that some workers and some communities could adapt before the next wave hit.

AGI does not work that way.

The moment a general intelligence capability exists, every company on earth can deploy it simultaneously. The moment the system exists and an API goes live, a lawyer in Des Moines, a radiologist in Houston, a financial analyst in Chicago, and an accountant in Phoenix are all competing with the same system at the same time. The replacement that took the Mahoning Valley steel industry thirty years does not take thirty years anymore.

Kasparov walked off that stage in 1997 having seen something similar to an artificial intellect. He was right to be unsettled. He was looking at the early version of something the entire world is now racing to build into the general version.

Deep Blue could only play chess.

What comes next can play every game at once.

Forwarded this? Subscribe free at replaced.co. If it made you think, send it to one person who needs to read it.

Further reading: IBM’s own account of how Deep Blue was built, how it worked, and what the 1997 match actually proved about the relationship between computing power and human intelligence. Read it here →