AI Isn't Taking the Job. It's Taking the Apprenticeship.

The number that changed how I think about this is not about job losses at all.

Stanford's 2026 AI Index reported that employment for software developers aged 22 to 25 fell by roughly a fifth since 2024. Over the same period, employment for developers aged 30 and above continued to grow.

Separate work from Stanford's Digital Economy Lab, led by Erik Brynjolfsson, found a similar pattern more broadly — around a 13% relative decline in early-career employment in the occupations most exposed to AI.

Read those two facts together and the familiar headline falls apart. This is not a profession being replaced. It is a profession that has stopped hiring at the bottom while continuing to hire in the middle.

Which is a different problem, and I think a more serious one.

We Have Run This Experiment Before

Automotive manufacturing has been through several waves of this, and the pattern is consistent enough to be useful.

When robots arrived in body shops, the confident prediction was that welders would disappear. They did not disappear. The trade moved up a layer.

What emerged instead were robot programmers, cell maintenance technicians, integration engineers, and people who could diagnose why a weld gun was drifting out of specification on the second shift. The work became more skilled and more valuable.

But I want to be honest about the two parts of that story which usually get left out.

There were fewer of them. A line that needed forty people needed fifteen with different skills. Net employment on that line went down, even as the individual jobs got better.

They were often different people. Not everybody who was excellent at manual welding became excellent at robot programming. Some did, brilliantly. Others were retrained onto something they never really took to, and a number left the industry. That transition was managed well in some plants and badly in others, and the difference was almost entirely about how much effort leadership put into it.

So when I hear that AI will create more jobs than it destroys, I believe it. I have watched it happen. I simply do not think that statement means what people use it to mean.

The Trades Being Created Right Now

They are appearing faster than most engineering organisations are noticing.

The clearest is the AI evaluation engineer — someone who builds and maintains the frameworks that determine whether an AI system is actually working. Test suites, regression pipelines, criteria developed alongside domain experts. Two years ago this was one bullet inside a machine learning job description. It is now a standalone title with its own market, and salaries in the United States running in the region of fifty-odd dollars an hour and upward.

Alongside it: agentic AI engineers, posted in early 2026 by organisations including Anthropic, Salesforce, EY, Deloitte and Accenture — roles that did not exist in their hiring pipelines two years earlier. Then LLM operations, context engineering, AI infrastructure. In manufacturing specifically, AI oversight and audit specialists are being cited as a genuinely net-new category.

Notice what nearly all of them have in common. They are jobs verifying, maintaining and governing the automation. That is exactly what happened in the body shop.

And one cautionary note worth carrying, because it shows how fast this moves. Prompt engineer was the celebrated new profession of 2023. In a Microsoft-commissioned survey of 31,000 workers across 31 countries, it ranked second from last among roles companies expected to add. Prompting turned out to be a skill, not a job — absorbed into other roles within about two years.

Some of what is being called a new trade today will go the same way. The ones that persist will be the ones with a verification or accountability function attached, because those cannot be absorbed into everybody's job without becoming nobody's.

The Rungs, Not the Ladder

Here is what I think the Stanford numbers are really telling us, and it is the part that worries me.

AI is not removing the job. It is removing the apprenticeship.

Think about what junior engineering work actually consisted of. Checking drawings. Running tolerance stack-ups. Producing standard simulations somebody senior had specified. Writing up test reports. Chasing a supplier for a document.

None of that was economically valuable. That was never the point.

The point was that after two years of checking other people's drawings, you had developed something you could not have been taught directly — a feel for when a design is going to be difficult, before you can say why. After a hundred tolerance stacks, you stop calculating and start noticing.

That work is now the most automatable work in the building. And it is being automated first, precisely because it is routine, high-volume and low-risk.

So the organisation keeps its senior engineers, becomes more efficient, and quietly stops manufacturing new ones.

The bill for that does not arrive this year. It arrives in about a decade, when the people who learned the old way start retiring and there is a thinner layer than expected behind them.

The Honest Numbers, With Their Caveats

The optimistic case is real and I do not want to dismiss it.

The World Economic Forum's work points to substantial net job creation — on the order of 170 million roles created against 92 million displaced by 2030. Goldman Sachs has made the longer historical argument: more than 85% of US employment growth since 1940 came from technology-driven job creation, and around 60% of American workers today are in occupations that did not exist in 1940.

That history is genuinely reassuring about the destination. It says very little about the journey.

And the caveats deserve equal billing. Roughly 77% of emerging AI roles require a master's degree or equivalent experience. McKinsey data indicates that 56% of displaced workers in automation-heavy sectors report difficulty moving into new roles even when those roles exist and are nearby.

I would also treat the projections themselves carefully. The WEF figures are employer intention surveys rather than econometric forecasts, they have historically overstated both creation and displacement, and different editions of the same report carry noticeably different numbers. They describe what executives expect to do, which is not the same as what will happen.

Net positive is not the same as evenly distributed, and it is certainly not the same as painless.

What a Leader Can Actually Do

Not much about the macro picture. Quite a lot about the specific twenty people you are responsible for.

Notice what your juniors are no longer doing. Make an actual list of the tasks that have quietly moved to a tool in the last eighteen months. Then ask what those tasks were teaching, and whether anything has replaced that.

Rebuild the apprenticeship deliberately. If nobody learns by checking drawings any more, then the judgment that used to come from it has to be transferred another way — reviews where a junior engineer defends a decision, exposure to the meeting where the trade-off is argued, being made to predict an outcome before the tool produces one.

That last one matters more than it sounds. If a young engineer predicts the answer before running the analysis, and is sometimes wrong, they are still building the model in their head. If they only ever read the output, they are not.

Be honest with people about the shape of it. The trades that are appearing are real, they pay well, and they are open to engineers who move early. They are also fewer, and they demand different aptitudes. Telling people the transition will be smooth because the aggregate numbers look fine is not kindness. It is how the badly-managed plants handled the robots.

The industries that got this right did not do it by predicting the future accurately.

They did it by taking the retraining seriously about two years before they had to.