For years, the AI-and-jobs debate has revolved around a deceptively simple question:
Which jobs will AI take?
It is a useful question, but it may also be the wrong one.
The reason is that AI does not need to eliminate an occupation in order to reduce the demand for people who perform it. It can eliminate tasks. Then it can reduce the number of hours required to complete those tasks. That can reduce billable hours, slow hiring, remove organizational layers and eventually reduce the number of people needed.
A job can survive on paper while becoming a much smaller economic opportunity.
That distinction matters because the first effects of AI may appear not as spectacular mass unemployment, but as slower wage growth, fewer entry-level opportunities, fewer hours of paid work and higher expectations of productivity from each remaining worker.
There is already evidence pointing in this direction. Anthropic’s March 2026 analysis found no systematic increase in unemployment among highly AI-exposed workers since late 2022, but it found suggestive evidence that hiring of younger workers had slowed in exposed occupations. Research summarized by Forbes from Apollo Global Management found that real wage growth in AI-exposed occupations lagged less-exposed occupations by 6.7 percentage points after 2023, despite no statistically significant decline in employment.
So perhaps the first stage of the AI employment revolution is not:
“Your job is gone.”
It is:
“Your job still exists, but society needs fewer hours of it, fewer people doing it, and perhaps eventually fewer people paid as much for doing it.”
That leads to a much bigger question.
If AI eventually reduces the amount of human labour required to produce what society needs, what happens to an economy built around human labour?
That is the paradigm shift this article explores.
1. The Question Everyone Is Asking
The conventional discussion usually produces lists.
Doctors will remain. Electricians will remain. Teachers will remain. Software developers will change. Accountants will be affected. Drivers may eventually disappear.
Such lists are useful for career planning, but they miss the deeper economic mechanism.
An occupation is not one indivisible activity. It is a bundle of tasks.
A lawyer reads documents, searches precedents, drafts arguments, communicates with clients, negotiates, checks facts and accepts responsibility for the final advice.
A programmer understands requirements, writes code, tests it, debugs it, documents it and maintains it.
An accountant collects information, checks records, reconciles accounts, prepares reports and explains financial consequences.
A customer-service employee reads a question, identifies the problem, searches information, follows a procedure and communicates an answer.
AI does not have to replace the entire person. It only has to make enough of those tasks unnecessary.
That distinction is already becoming visible in professional services. The American Bar Association, for example, has discussed how AI could substantially reduce billable hours as clients increasingly focus on outcomes rather than time spent. The broader shift is from paying for human effort toward paying for results.
The important economic variable, therefore, is not simply jobs.
It is human labour required per unit of output.
And that is a much more consequential variable for a labour-based economy.
2. AI Is Attacking Cognitive Friction
There is another way to understand what is happening.
AI is not simply attacking white-collar jobs.
It is attacking cognitive friction.
Whenever a task requires a person to know how to do something, AI can potentially reduce the amount of specialised knowledge required to perform it.
That distinction becomes important outside offices.
Imagine a future plumbing system in which a homeowner points a phone at a leaking pipe. The system identifies the component, explains the problem, identifies the replacement part and provides instructions. A sufficiently advanced system might even control a robotic tool that performs the repair.
At that point, the relevant question is no longer:
“Can AI become a plumber?”
A more interesting question is:
“Can technology make plumbing require less plumbing expertise?”
That is a much broader mechanism.
You don’t necessarily have to automate the plumber if you can make plumbing no longer require a plumber.
The same principle can operate in countless areas. AI can make software development easier. It can make tax preparation easier. It can make medical information easier to interpret. It can make electrical diagnosis easier. It can make vehicle maintenance easier. It can make legal documents easier to prepare.
The worker does not have to disappear immediately.
The expertise barrier can disappear first.
3. What Goes First?
The most exposed work tends to share several characteristics. It is digital, repetitive, rules-based, information-heavy and relatively easy to measure or check. It can often be performed remotely, and its outputs tend to be standardised enough to be evaluated systematically.
- routine customer support
- basic translation
- standard document preparation
- routine bookkeeping
- basic research
- administrative coordination
- repetitive legal work
- routine coding
- basic content production
- scheduling and back-office work
- certain forms of financial analysis
- routine educational assistance
Anthropic’s 2026 labour-market research illustrates the potential scale of this exposure. Its measure combines theoretical LLM capability with observed real-world usage and finds that actual AI coverage remains well below what current technical capabilities could theoretically cover. Computer programmers, customer-service representatives and financial analysts are among the more exposed occupations.
That gap is crucial.
Today’s AI usage is not necessarily the ceiling.
The distance between what AI can theoretically do and what organisations currently use it to do represents a potential future wave of change.
Evidence snapshot: Anthropic’s research specifically distinguishes theoretical AI capability from observed exposure, showing that current usage remains far below potential capability.
The chart makes an important distinction: capability and adoption are not the same thing.
The future question is what happens as that gap narrows.
4. The Middle Gets Squeezed
This leads to one of the most important patterns in the emerging labour market.
The vulnerable area may not be the entire workforce.
It may be the middle.
At one end are occupations involving physical presence, unpredictable environments and difficult real-world manipulation.
At the other end are people whose value depends heavily on judgment, responsibility, trust, relationships, leadership or unusually deep expertise.
Between them lies a huge amount of standardised cognitive work.
That middle is where AI has a natural advantage.
Consider a simplified organisation. Senior people may be responsible for strategy, accountability, negotiation and relationships. Employees closer to the operational layer may handle physical work, unpredictable environments and direct interaction. Between them sits a substantial amount of analysis, documentation, coordination, information processing and routine decision-making.
AI can compress that middle.
One senior person with AI may perform analytical work previously distributed among several employees. One technician with AI assistance may perform diagnostic work previously requiring a specialist. One business owner may handle administration that once required several staff members.
The organisation may remain.
The organisational layer may not.
That creates a squeeze.
5. The Worker Doesn’t Have to Be Replaced
There is a second way in which AI can reduce labour demand that is easy to overlook: it can change the production process itself.
A striking example is emerging in the construction of AI data centers. A September 1, 2026 Wall Street Journal report, “See the New Building Techniques Turbocharging the Data-Center Boom,” describes how hyperscalers’ need for speed is driving time-saving construction techniques, including greater use of modular components and factory-based assembly. The broader objective is to compress construction timelines as demand for AI infrastructure accelerates.
The significance goes beyond data centers.
A drilling robot does not have to replace an entire construction worker. It only has to automate one repetitive part of the job. Prefabrication can move more work from the construction site into a factory. Better materials can reduce waiting time. Standardised components can simplify installation. Faster connection systems can eliminate repeated manual operations.
Each improvement may look small.
Together, they can substantially reduce the amount of human labour required to produce the same physical structure.
This is the important distinction.
The worker does not necessarily have to be replaced. The work itself can be reduced.
The same principle can eventually apply across construction, manufacturing, maintenance, logistics and other physical industries. AI can help redesign not only who performs a task, but whether the task needs to exist in its present form at all.
You don’t necessarily have to automate the construction worker if you can redesign construction so that far less construction work is required.
And that is precisely why the employment effects of AI could extend beyond the occupations that appear on conventional lists of “AI-exposed jobs.”
Sometimes the machine does not become the worker. It changes the process until there is simply less work left for the worker to do.
Source: The Wall Street Journal, “See the New Building Techniques Turbocharging the Data-Center Boom,” September 1, 2026.
6. The Automation of Maintenance
The same logic becomes even more interesting when we move from construction to the products themselves.
Suppose a repair technician currently spends four hours diagnosing and repairing a machine.
A future machine could have continuous sensors, automatic fault detection, remote diagnostics, modular components, simple access mechanisms, guided replacement and automated calibration.
The technician may still exist. But perhaps the machine now requires one hour of human intervention instead of four.
The worker has not been replaced. The work has been reduced.
Reversed Unplanned Obsolescence
The Automation of Maintenance
Planned obsolescence is familiar: products are designed, intentionally or otherwise, around relatively short replacement cycles.
The opposite possibility is more interesting.
Better components could reduce failures. Better design could improve access. Modular construction could simplify replacement. Sensors could diagnose problems. AI could guide repairs. Machines could increasingly monitor and maintain themselves.
The progression could look something like this:
Better components → fewer failures → easier access → modular replacement → automatic diagnosis → self-maintenance → self-repair → longer product life
This is not a prediction that every machine will become self-repairing. It is a direction worth noticing.
The best way to replace a repair job may not be to build a robot that repairs the machine. It may be to build a machine that can repair itself.
That is why the employment effects of AI cannot be analysed only by looking at occupations. We also have to examine the products and systems those occupations maintain.
7. AI Can Remove the Need to Know How
For most of human history, competence required knowledge.
If you wanted to drive a car, you had to learn to drive. If you wanted to repair a machine, you had to learn mechanics. If you wanted to write software, you had to learn programming. If you wanted to prepare complicated financial documents, you had to learn accounting.
AI introduces a different possibility.
Humans may increasingly specify what they want, while machines handle how to achieve it.
Consider a child sitting in an autonomous vehicle. The child doesn’t know how to drive. But the child can still understand something important: Slow down. Stop. Something is wrong. I feel danger.
If an automated driving system becomes capable enough, the technical skill of driving may become unnecessary for the passenger.
This is not a claim that autonomous driving has already reached that point. Driverless vehicles are operating commercially in multiple cities, but fully general Level 5 autonomy remains a much more distant proposition. Current safety evidence is encouraging, while broader conclusions still require more data and consistent evaluation.
The interesting question is therefore not:
“Will autonomous cars make mistakes?”
Of course they will. Humans make mistakes too.
The meaningful question is:
“Which makes fewer serious mistakes under comparable conditions?”
That is an empirical question.
If future evidence shows automated systems are substantially safer than human drivers, society will face an unusual situation: a human skill may become economically valuable not because humans perform it better, but because humans are accustomed to performing it.
The same pattern could eventually appear in many other domains.
AI does not have to give humans more expertise. It can make expertise less necessary.
8. The 90% Problem
Suppose AI initially removes 90% of the routine work in an occupation. The remaining 10% consists of judgment, exceptions, responsibility and difficult cases.
Humans remain. The occupation survives.
But then AI improves. It becomes better at exceptions, reasoning, diagnosis, planning and checking its own work. It begins attacking some of the remaining tasks.
The 90% figure is not a forecast. It is simply a way of illustrating the underlying mechanism.
The important point is that the work left behind by one generation of automation can become the target for the next generation.
Historically, automation often concentrated on repetitive tasks while leaving human supervision. AI is unusual because its improvement frontier includes increasingly sophisticated cognitive tasks.
The boundary between “routine” and “expert” work can therefore keep moving. What looks like the uniquely human part of a job today may become tomorrow’s automation target.
9. There May Be No Permanently Safe Job
This is why I am uncomfortable with lists titled “10 jobs AI will never replace.”
They confuse three different meanings of safety.
Technically safe
AI cannot currently perform the task adequately.
Occupationally safe
The occupation is likely to continue existing.
Economically safe
There will be enough demand for enough workers to earn a reasonable living from it.
These are not the same thing.
A plumber may remain technically difficult to replace. But if displaced workers from other occupations enter plumbing, the economic conditions of plumbing can change.
A profession can survive while becoming overcrowded.
That produces a very different kind of AI disruption.
10. The Second-Order Problem: Where Do the Displaced Workers Go?
Imagine that AI eventually eliminates a large number of middle-class knowledge jobs.
The displaced workers do not disappear. They still need income, housing, food, healthcare and transportation. They will also compete for the jobs that remain.
Suppose a driving job remains available. A person who left school after the 12th standard may previously have competed with other drivers. Now imagine that a displaced engineering graduate also wants that job.
The driving occupation has not been automated.
Its labour market has changed.
The same thing could happen with electricians, construction workers, hospitality employees, delivery workers, maintenance technicians, caregivers, mechanics and other occupations that initially appear relatively protected from AI.
So even a job that AI does not directly attack can become economically vulnerable because AI has changed the supply of workers competing for it.
This is the second-order effect that conventional job forecasts often miss.
AI does not have to automate a job to make that job less secure.
It only has to eliminate the alternatives.
11. The Wage Problem May Arrive Before the Job Problem
If AI makes a worker more productive, several things can happen. The worker can receive higher pay. The company can receive higher profits. Customers can receive lower prices. The company can hire fewer people. Or some combination of all four can occur.
There is no economic law saying that higher productivity must translate into proportionally higher wages.
The recent wage evidence is therefore important. Forbes, summarising Apollo Global Management research, reports substantially slower real wage growth in occupations with greater AI exposure, despite no corresponding broad collapse in employment. It also reports lower starting wages at AI-exposed companies after ChatGPT’s launch, particularly for junior and mid-level workers.
That suggests a possible sequence:
AI arrives → productivity rises → hiring slows → bargaining power changes → wage growth weakens → tasks disappear → positions consolidate → employment eventually falls in some areas.
The sequence is not inevitable, but it is economically plausible.
If that is how disruption unfolds, waiting for unemployment statistics to tell us that AI has changed the labour market may mean waiting too long.
The labour market can deteriorate without producing immediate mass unemployment.
12. The Productivity Paradox
Suppose a company needs 20 people to produce a certain amount of work.
AI allows the same output to be produced with 10 people.
The company has become twice as productive.
That is excellent for the company.
But what happens if every competing company obtains approximately the same productivity gain?
They do not necessarily create twice as much work. They compete. Prices may fall. Margins may change. Market share may shift. Total labour demand can decline.
This is why “AI will make everyone more productive” is not, by itself, an answer to the employment problem.
Productivity answers:
How much can we produce with fewer people?
It does not automatically answer:
What will those people do instead?
That second question is much harder.
New World and AI Taking Jobs Scenario: The Solution
Evidence & Further Reading
- Anthropic — “Labor market impacts of AI: A new measure and early evidence” (March 5, 2026). Research
- Forbes — “Why AI Will Cut Your Pay Before It Takes Your Job” (August 5, 2026). Article
- Federal Reserve — “AI and Coder Employment: Compiling the Evidence” (2026). Research
- The Wall Street Journal — “See the New Building Techniques Turbocharging the Data-Center Boom” (September 1, 2026). Article