The Missing Middle: The Hidden Cost of Not Hiring Juniors in the Age of AI

It is March 2033. A mid-sized company has just finished rebuilding its operations around AI agents. Customer service, reporting, content, first-line analysis: all of it runs, mostly well. Then the board asks a simple question. Who owns these systems? Who checks them, catches their mistakes, and answers for their results? The company goes looking for experienced people to do that work, and finds very few. Every competitor is looking for the same people.

That scene is imagined. The chain of decisions that could produce it is not. It is being written right now, quietly, in hiring plans that look perfectly rational on a spreadsheet. This piece follows that chain from 2026 to 2036, using the signals already visible, and tries to keep a clear line between three things: what has been measured, what institutions forecast, and what is my own extrapolation.

2026: the door closes from the outside

The labour market of 2026 has a name. Economists call it low-hire, low-fire. In an August 2026 commentary, the Federal Reserve Bank of Cleveland described an unusual combination in an otherwise healthy economy: hiring, quits and layoffs all running low at the same time. People with jobs are rarely let go. People without them find it hard to get in.

The weight of that falls on one group more than any other. In its August 2026 update, Stanford's Digital Economy Lab found that employment of 22 to 25 year-olds in the occupations most exposed to AI stands about 19% below where it would be had it kept pace with less-exposed peers. A year earlier the gap was 15%. Experienced workers in the same occupations show no comparable shortfall.

The mechanism matters more than the number. The decline comes almost entirely from fewer young people being hired, not from more being let go. And it is concentrated in roles where AI substitutes for the work, while employment is flat or rising where AI complements it.

Graduate data points the same way. The Federal Reserve Bank of New York puts unemployment among recent college graduates at around 5.6% through the second quarter of 2026, with about 42% working in jobs that don't require their degree, the highest underemployment rate since 2020. Recent graduates used to do better than the market as a whole. That is no longer true.

Meanwhile, companies have started saying it out loud. According to Challenger, Gray & Christmas, AI was the leading stated reason for announced US job cuts for five consecutive months through July 2026, cited in more than 112,000 cuts this year, about a quarter of the total. Yet total announced job cuts are down 41% on the same period of 2025. Both facts are true at once, and together they describe the shape of this moment. Not a wave of firings, but a door closing quietly from the outside.

Two cautions before going further

• Layoff reasons are self-reported. Some companies have every incentive to present ordinary cost-cutting as an AI story, because it plays better with investors. The trend is meaningful. The exact numbers are partly narrative.
• AI is not the only cause. The Cleveland Fed analysis does not attribute the freeze to AI at all, and places it inside a decline in labour market fluidity that goes back decades. Interest rates and post-pandemic over-hiring explain part of what we see.

Neither caution changes the direction. They change how confidently we should read the pace.

2026 to 2031: from adding AI to rebuilding around it

The first phase of corporate AI, roughly 2023 to 2025, was experimentation, and it was mostly unprofitable. A widely cited MIT report found that around 95% of enterprise generative AI pilots produced no measurable impact on profit and loss, and blamed poor integration into real workflows rather than the models themselves.

What is striking is that hiring behaviour changed before the returns arrived. In April 2025, Shopify's chief executive told staff that teams should show AI could not do a job before asking for more headcount. That logic spreads easily, because it asks nothing of anyone already employed. It simply stops adding.

Some went further. In September 2025, Salesforce's chief executive said AI agents had allowed the company to reduce its customer support team from about 9,000 people to about 5,000.

The next phase is a redesign. Companies are learning that bolting agents onto old processes raises individual productivity but barely touches the balance sheet. The gains come from rebuilding the process around what agents do well, and only then deciding which people are needed, and where.

The potential scale is large. The McKinsey Global Institute estimates that technologies already demonstrated could, in principle, automate activities accounting for about 57% of US work hours, while stressing that this measures how much work could change, not how many jobs will be lost.

What the redesign produces is a different organisational shape. Fewer people at the base, doing first drafts, first analyses and first responses. More people in the middle, designing agent workflows, reviewing their output and answering for the results. Judgment stops being a quality we hope senior staff have and becomes the central function of the organisation.

Institutional forecasts expect the aggregate to come out positive. The World Economic Forum projects that by 2030 around 170 million jobs will be created and 92 million displaced worldwide, a net gain of 78 million, while 39% of today's skills become outdated.

Gartner expects AI to create more jobs than it eliminates from 2028. Others are far less optimistic.

In 2025, Anthropic's chief executive Dario Amodei warned that AI could eliminate half of entry-level white-collar jobs within one to five years.

The disagreement is real, but notice what the optimistic and pessimistic views share. Both expect the entry level to change first and fastest. The question is not whether the base of the pyramid shrinks. It is what happens to the people who would have climbed it.

2033: the missing middle

This is the part of the story the aggregate numbers hide.

Entry-level work was never only about output. The first version of the deck, the analysis nobody else wanted to run, the support tickets, the tedious reconciliation: that was how people learned. Each task was a repetition. Enough repetitions, and a beginner developed the thing no job description can list but every senior role depends on. Judgment.

AI absorbs exactly those tasks, and does them well. So the company keeps the output and quietly loses the training. A junior working with AI in 2026 can produce more than any junior in history, while getting fewer of the repetitions that used to turn juniors into experts.

Now run the arithmetic forward. The people not hired as juniors in 2026 would have been mid-level by around 2029, and senior by around 2033. They will not be there. This is a projection, not a measurement, but it rests on a fact that is hard to argue with: senior people are made from junior people, and there is no other supply.

The paradox is what makes it serious. The agent-based organisation needs more experienced judgment, not less: people who can tell a sound output from a confident mistake, who understand the business well enough to design the workflow, and who can be accountable when it fails. Anthropic's research on how people use its models points the same way: long-standing users get noticeably better results than newcomers, and more skilled users benefit most. AI rewards expertise it does not itself create.

That is the missing middle. Not a shortage of jobs, but a shortage of the experienced people needed to run the new ones, created by decisions that looked like efficiency at the time.

The Gulf runs a different clock

Writing from Dubai, it is hard not to notice that this story is largely a Western one. The Gulf is running on a different clock, driven less by corporate cost-cutting than by state-led investment in the infrastructure of AI itself.

Stargate UAE, a planned 1-gigawatt compute cluster built by G42 with partners including OpenAI, Oracle and NVIDIA, has its first 200-megawatt phase scheduled for 2026. Around that build-out, employer hiring intentions have been among the strongest in the world.

ManpowerGroup's Employment Outlook Survey recorded a net employment outlook of +60% for the UAE in the second quarter of 2026, one of the strongest readings anywhere.

It has not been a smooth line. The outlook fell to +17% for the third quarter, after a first half of the year disrupted by regional conflict, before recovering to +41% for the fourth. Hiring in the region follows investment cycles and geopolitics far more than the quiet substitution visible in the United States.

That leaves an open question I don't think anyone can answer yet. Is the Gulf skipping the freeze because it is building the infrastructure layer itself, or is it simply a couple of years behind the same curve? If it is the latter, the region has an unusual advantage. It can see the missing middle coming, and design around it before it arrives.

2036: two possible endings

By the mid-2030s, the story could resolve in one of two broad ways.

• Apprenticeship, reinvented. Companies deliberately fund learning roles whose purpose is to build judgment rather than to produce output. Juniors review agent work, own small workflows end to end, and learn the business by being accountable for real results early. Training becomes a strategic cost, treated the way R&D is treated today.
• The experience premium. Nobody coordinates. Each company makes the individually rational choice not to train, experienced judgment becomes scarce, and its price rises sharply. The gap between those who got in before the door closed and those who didn't becomes one of the defining divides of the decade.

The first ending is not the default. It requires companies to spend on something whose payoff arrives years later and may end up benefiting a competitor. That is exactly why it is worth deciding now.

What leaders can do now

If you are responsible for hiring, the decision that matters is not whether to use AI. It is what you do with the entry level once AI can cover its tasks.

• Keep a junior intake, on purpose. Treat it as an investment in future judgment, sized deliberately, not as the first line to cut when budgets tighten.
• Redesign junior roles around judgment. Give juniors the work of reviewing, correcting and owning AI output, not the work AI now does alone.
• Measure learning, not only output. Track how quickly people reach independent judgment, not just how much they produce with AI's help.
• Pair juniors with the owners of AI workflows. The people accountable for agent workflows are the natural teachers of the next generation.
• Name an owner for every AI workflow. Accountability is where judgment gets built. Without an owner, there is nobody to learn from when something fails.

What could make this wrong

A forecast is only credible if it names its own failure modes. These are the ones I would watch.

1. The returns never arrive. If most AI deployments keep failing to pay back, companies will reverse course and hire juniors again, and today's freeze will look like a hype cycle.
2. AI is mostly a label. If "AI" in layoff announcements is largely a respectable name for ordinary cost-cutting, this is more economic cycle than structural shift.
3. Capability plateaus. A slowdown in what models can do would stretch every phase out and give organisations time to adapt.
4. Policy responds. A visible rise in youth unemployment is exactly the kind of problem that produces fast political answers, from hiring incentives to apprenticeship requirements.
5. Macroeconomics dominates. Interest rates and the business cycle may explain more of the youth hiring slump than technology does.

Each of these would change the speed. None of them changes the underlying logic. Expertise is built through practice, and practice is what the entry level used to provide. Whatever the pace, that is the thing worth protecting.