The Missing First Rung

AI is making intelligence cheaper. But what happens to the people who are still learning how to use it?

Mohamed Alketbi | October 2026

A few weeks ago, I attended Claude Founder House in London.

I had the opportunity to meet people from Anthropic and get a closer look at how one of the world's leading AI companies is building and scaling its technology.

What interested me wasn't only the models themselves. It was the speed at which the economics of building things are changing.

We are approaching a point where a small group of people, equipped with sufficiently capable AI systems, can accomplish work that previously required an entire department.

As an AI student, this is fascinating. As someone building companies, it is an extraordinary opportunity.

But as someone about to graduate from university, I find myself thinking about something else.

What happens to the people who are supposed to be entering the workforce?

Not in twenty years. Now.

1. We are automating the beginning of careers

Consider how a software engineer traditionally becomes a senior engineer.

You graduate, join a company, fix bugs, write relatively simple functions, review code, break things, and slowly develop an understanding of how large systems work.

Eventually, you become the person who makes architectural decisions, handles difficult problems, and mentors the next generation.

There is a learning process here that takes years.

Now imagine that AI can perform a growing share of the tasks assigned to junior engineers.

A company might still need senior engineers. In fact, it may value them even more because they can supervise AI systems, evaluate their output, and make decisions that require experience.

But the company might need fewer junior engineers.

This creates an interesting problem.

If we automate the work through which people acquire experience, how do we produce experienced people?

The same question applies beyond software.

Junior lawyers spend years researching cases and drafting documents. Analysts build financial models. Consultants prepare presentations and research markets. Doctors learn through supervised clinical practice.

A significant part of early professional development involves doing relatively structured work, making mistakes, and receiving feedback.

These are also the tasks that AI systems are increasingly good at assisting with or automating.

We might be removing the first rung of the career ladder while keeping the upper rungs intact.

And a ladder without its first rung is not particularly useful.

The early evidence

A 2025 Stanford study, subsequently updated in 2026, found that employment among workers aged 22 to 25 in highly AI-exposed occupations was falling behind employment in less-exposed occupations.

The researchers estimated a substantial relative employment shortfall for young workers in the most exposed occupations.

This is not proof that AI caused every missing job. Economic conditions, interest rates, and changes in hiring all matter.

But it is consistent with a mechanism we should take seriously.

In the UK, graduate recruitment is already intensely competitive. The Institute of Student Employers reported an average of 141 applications per graduate vacancy in its 2026 survey.

An application is not a person, and these numbers alone cannot tell us how much AI has contributed to the pressure.

Still, imagine being a graduate today.

You have spent three or four years studying a subject, only to enter a market where the tasks you were trained to perform are becoming cheaper every few months.

We often tell young people to learn AI.

I agree. But that answer is incomplete.

If everyone learns AI, companies still need to decide whom to hire.

And if a senior employee using AI can perform the work of several junior employees, learning the tool does not automatically create more entry-level positions.

The problem is not simply a skills gap.

It may become an experience gap.

2. Jobs are not the same thing as work

I think we often confuse two different concepts when discussing AI.

A job is an economic arrangement. Someone pays another person to perform a collection of tasks.

Work is the activity itself.

AI can replace jobs without eliminating the underlying work.

Take customer support.

A company might previously employ twenty people to answer customer questions. An AI system could eventually handle much of that workload, leaving a smaller team to manage unusual cases, complaints, and escalations.

The company has reduced the number of jobs.

But customers are still asking questions. Problems are still being solved. The work continues, only through a different production process.

Software is another example.

If AI makes software development dramatically cheaper, we may see an explosion in the amount of software being created.

There are countless problems for which building software has never been economically worthwhile.

Lowering the cost changes that equation.

This is the optimistic case.

The pessimistic case is that demand does not expand quickly enough to absorb displaced workers, or that the new jobs require skills that existing workers cannot easily acquire.

Both mechanisms can operate simultaneously.

A technology can increase total economic output while reducing employment in particular occupations.

That distinction matters.

The number of tasks society wants completed is not fixed. But neither is the number of paid jobs guaranteed to grow alongside those tasks.

3. Everyone can code. That doesn't mean everyone can build a company.

Something else is happening alongside the employment transition.

Software development is becoming accessible to a much larger population.

A person with limited programming experience can increasingly describe an application, generate its interface, connect a database, and deploy a working product.

This is remarkable.

But I have noticed an assumption emerging from this new capability.

If building software becomes easy, building startups must also become easy.

I don't think that's true.

There is a difference between producing a product and creating a business.

A business needs distribution. It needs customers, pricing, trust, and a problem important enough that someone is willing to pay to solve it.

AI can help you reproduce the functionality of an existing product. It cannot guarantee that anyone needs another version of that product.

In fact, as software becomes easier to create, differentiation may become harder.

Imagine a market where a thousand founders can build roughly the same application in a weekend.

The code is no longer particularly scarce.

What becomes valuable?

Understanding customers. Finding distribution. Building relationships. Knowing which problems matter. Developing judgment about what not to build.

This is something I think we underestimate when talking about the future of entrepreneurship.

We are making the act of creation cheaper, but not necessarily making the act of discovering valuable problems easier.

The next generation of successful founders may not be the people who can write the most code.

They may be the people who understand the world well enough to know what code is worth writing.

4. The economics of intelligence

Historically, human intelligence has been expensive.

If you wanted legal analysis, you hired a lawyer. If you wanted software, you hired engineers. If you wanted specialised research, you employed researchers or consultants.

There was a relatively direct relationship between the availability of skilled people and the cost of producing intellectual work.

AI changes this relationship.

Once a capable model has been trained, it can serve enormous numbers of users. The marginal cost of generating another response or performing another task can be much lower than employing a human professional for equivalent routine work.

Of course, AI is not free. Inference requires computing infrastructure, electricity, networking, and ongoing engineering. Complex tasks can also require substantial human verification.

But the direction is important.

The cost of accessing certain forms of cognitive capability is falling.

This creates a potentially enormous productivity opportunity.

It also raises a distributional question.

If one person using AI can produce substantially more economic output, who receives the additional value?

The employee?

The company?

The owners of the AI infrastructure?

The customer, through lower prices?

Economists Daron Acemoglu and Simon Johnson explore a related question in Power and Progress. Their historical argument is that technological progress does not automatically translate into broadly shared prosperity.

Institutions and economic power influence how productivity gains are distributed.

I think this is particularly relevant to AI.

We could build an economy that produces more goods and services with fewer workers, while concentrating a larger share of income among the owners of capital and technology.

Or we could build an economy in which lower production costs create new businesses, expand access to services, and improve living standards.

The technology alone does not determine which outcome occurs.

5. A world divided by access to intelligence

There is another dimension to this debate that receives less attention.

Most discussions about AI and employment focus on the United States, Europe, and China.

But what about the rest of the world?

Anthropic's Economic Index offers an interesting way to study this question. It examines how people use Claude across occupations, tasks, and countries.

One of its findings is that AI usage is not evenly distributed geographically.

This is not surprising.

Access depends on infrastructure, affordability, digital skills, language support, and the ability to integrate AI into productive work.

Consider two students.

One studies in London, with reliable internet, access to advanced models, university resources, and an ecosystem of companies developing AI products.

The other studies in a community where connectivity is expensive, electricity is unreliable, and access to computing resources is limited.

Their intellectual potential may be identical.

Their ability to participate in the emerging AI economy is not.

The World Bank's research on AI foundations shows that the infrastructure required to develop and adopt AI remains highly concentrated.

This creates a risk that existing economic inequalities will be reinforced.

Countries that already possess capital, infrastructure, and technical expertise may capture a disproportionate share of AI's economic benefits.

Countries without those advantages may become dependent on technologies they do not control.

But there is an alternative possibility.

AI could allow developing economies to bypass some traditional barriers.

A rural hospital might access diagnostic assistance that would otherwise require scarce specialists. A student could receive personalised education. A small business could communicate with international customers without employing a large administrative team.

These are genuine opportunities.

The important distinction is between having access to an AI product and possessing the capability to create economic value with it.

A country does not necessarily need to train a frontier model.

It does need reliable infrastructure, education, local expertise, and institutions capable of using and governing AI effectively.

Otherwise, we risk replacing one form of technological dependence with another.

6. Intelligence is becoming a dual-use technology

The same capabilities that make AI economically transformative also create difficult safety questions.

Consider biology.

Models capable of reasoning about complex scientific information could accelerate drug discovery, support medical research, and help scientists understand diseases.

That could be extraordinarily valuable.

But increasingly capable systems may also lower barriers to certain forms of biological misuse.

This is why AI safety research increasingly examines not only what models can accomplish, but what harmful capabilities they might enable.

The point is not that every powerful model will be dangerous.

It is that capability improvements can have benefits and risks at the same time.

We cannot evaluate progress solely by measuring what a system can do.

We also need to understand the circumstances under which those capabilities are deployed, who has access to them, and which safeguards exist.

The same principle applies to employment.

A model becoming better at coding is a technical achievement.

Whether that achievement improves the lives of software engineers, displaces junior workers, or enables new companies is an economic and institutional question.

Benchmark scores do not answer it.

7. Who is responsible for the next generation?

This brings me back to the question I started with.

If companies increasingly automate junior-level work, who will train the next generation of senior professionals?

It is tempting to say that the market will adapt.

Historically, technological change has created new occupations and industries. That may happen again.

But adaptation takes time, and the costs are not distributed equally.

A company has an incentive to reduce costs today. It may have much less incentive to train workers who could leave for another employer tomorrow.

Yet the economy as a whole depends on the continuous development of skilled people.

This is a coordination problem.

Universities cannot solve it alone. They can teach theory, technical foundations, and analytical methods, but professional judgment often develops through practical experience.

Businesses cannot ignore it either. If every organisation hires only experienced professionals, the supply of experienced professionals eventually becomes constrained.

Governments have a role because the consequences extend beyond individual companies.

We may need to rethink how professional training works.

Perhaps AI should become part of supervised apprenticeships rather than simply replacing junior employees. Perhaps universities and employers should collaborate on real projects where students learn to evaluate, correct, and manage AI-generated work.

Perhaps companies that benefit significantly from automation should invest more deliberately in developing future talent.

None of these approaches guarantees success.

But doing nothing is also a choice.

And it may be an expensive one.

8. The question is not whether AI will become more capable

I am optimistic about artificial intelligence.

I study it, work with it, and build products around it. I have seen how much it can expand the capabilities of individuals and small teams.

I believe it could improve healthcare, education, scientific research, and the productivity of entire economies.

But I am increasingly convinced that the most important questions about AI are no longer exclusively technical.

We know how to measure whether a model can solve a mathematical problem or generate working code.

We are much less prepared to measure whether the economic value created by these systems is being distributed fairly.

We can track the cost of inference, model performance, and adoption rates.

But are we tracking the number of graduates who cannot enter their chosen professions?

Are we measuring whether workers' wages rise alongside AI-driven productivity?

Are we examining whether developing economies are building their own capabilities or becoming more dependent on foreign infrastructure?

These questions matter because the success of AI should not be measured only by how much work machines can perform.

It should also be measured by what humans are able to achieve because those machines exist.

From Dubai to London, and across countries with very different economic circumstances, the challenge is ultimately the same.

How do we make intelligence more abundant without making opportunity more scarce?

I don't think the answer is to slow down every technological advance. Nor do I think we should assume that the market will automatically distribute its benefits.

We need to develop human capability alongside machine capability.

Because if we build systems that can perform the work of experts, but gradually eliminate the opportunities through which people become experts, we may discover that we have optimised for the wrong thing.

We are making machines better at work. We should be equally concerned with making sure humans still have a future in it.