AI Entry-Level Jobs: The Seniority Trap Facing Firms

When a young corporate lawyer begins his career at a global firm, one of his first duties might be taking notes at client meetings. Nobody calls it training. It is billed as work. It looks like clerical labour. It is, in fact, the entire curriculum — an hour spent watching senior partners handle a difficult room teaches more than any course ever will.

That is the thing almost every organisation has now forgotten: the tedious work at the bottom of the ladder was never really about the work. It was a delivery mechanism. The memo, the model, the first-pass audit, the meeting notes — these were the physical process by which raw graduates were converted into people with judgment. The output was a by-product. The expertise was the product.

And we have just spent three years dismantling that mechanism because the output became cheaper to buy elsewhere.

The conversation about AI entry-level jobs has been framed almost entirely as a fairness problem — young people locked out, graduates struggling, a generation told to learn to code and then told not to bother. That framing is real but incomplete, and it lets executives file the issue under corporate social responsibility rather than under strategy. The sharper version of the story is this: companies have started decommissioning their own expertise supply chain, and the invoice does not arrive for about ten years.

The tedious work at the bottom of the ladder was never about the work. It was a delivery mechanism. The output was a by-product. The expertise was the product.

The data on AI entry-level jobs stopped being ambiguous

For a while it was possible to argue this was just a soft labour market. That argument is getting harder to make.

Research from the Stanford Digital Economy Lab, updated again in August 2026, finds that since 2022 employment among young workers in the 40% of occupations most exposed to AI has fallen by roughly 11%, while employment among young workers in the least-exposed 60% has risen about 10%. The mechanism matters as much as the number: the decline is driven primarily by reduced hiring rather than layoffs or voluntary departures. Nobody is being fired. The door is simply not opening.

The same pattern does not appear among experienced workers in those occupations, and it does not appear in entry-level roles with low AI exposure. As MIT Technology Review has documented, the effect is specific to early-career work that AI can reach — which suggests firms are substituting the technology for precisely the junior tasks through which people traditionally gained their first foothold.

The split inside the data is the most revealing part. Where AI automates a task — writing the code, handling the chat — entry-level hiring falls. Where AI augments a task, supporting problem-solving or accuracy-checking, employment stays flat or rises. The technology is not doing anything to the talent pipeline on its own. Deployment philosophy is doing it. Two firms with identical tools and opposite assumptions will end this decade with completely different institutional capability.

And here is the part that should genuinely unsettle anyone who runs a company: much of this was pre-emptive. Harvard economists analysing résumé and posting data across hundreds of thousands of US firms found the contraction began almost immediately after ChatGPT’s release — companies pulled back on entry-level hiring adjusting for automation they anticipated rather than automation that had already arrived. The ladder was pulled up in advance of a capability that, in many functions, still has not fully materialised.

Expertise Debt: the liability on no balance sheet

The retreat from AI entry-level jobs creates a liability that appears in no set of accounts: the gap between the expertise an organisation currently uses and the expertise it is currently manufacturing.

Call it Expertise Debt. It behaves unlike any other line item a CFO manages, and it has three properties that make it uniquely dangerous.

It is invisible while it accumulates. Your senior people are still senior. Quality holds. Client work ships. Nothing on the dashboard moves. The damage is happening inside a cohort that does not exist yet, which means it produces no signal at all — until the moment you need someone with eight years of pattern recognition and discover that the last person you hired on that trajectory joined in 2022.

It compounds silently. A missed hiring year is not a one-year problem. Expertise is sequential. The person who is not a junior this year is not a mid-level in four years and not a principal in nine. Skip three intake years and you have not created a three-year gap; you have created a hole in the org chart still visible in 2040, because the people who would have filled it went somewhere that let them learn.

It cannot be repaid with money. This is what separates Expertise Debt from technical debt or deferred capex. There is no Q3 budget that fixes it. Judgment is time-denominated. You can pay a premium for someone else’s ten years, but only if someone else invested those ten years — a rapidly diminishing pool if the whole industry defers simultaneously. And the whole industry is deferring simultaneously.

Read that as market structure and the next move becomes obvious: everyone bidding for a shrinking supply of experienced people while nobody produces new ones. That is not a talent strategy. That is a bubble in a labour market with a ten-year production lag.

The Verification Paradox

Here is where the argument about AI entry-level jobs turns from a human-resources concern into operational risk.

Almost every responsible AI deployment framework in existence rests on the same load-bearing assumption: a competent human reviews the output. Human-in-the-loop. Meaningful oversight. Expert validation. This assumption is what makes the whole arrangement defensible to a board, a regulator, and now an insurance underwriter.

But verification capacity is not a separate skill you can hire for. It is the residue of production. You can spot the flaw in a financial model because you have built enough of them to know where models lie. You can tell a legal memo has missed something because you have written a hundred memos and been corrected on ninety. Recognition is built by production. There is no shortcut, and there never has been.

So the structure we are building looks like this: AI performs the junior work, juniors are no longer hired to perform the junior work, the tasks that produced verification capacity disappear — and the entire governance model still depends on verification capacity being abundant. Academic work has begun calling this pipeline compression: the experiential path that builds senior judgment narrows even while final outputs still look screened, so present stability in high-stakes domains can mask long-run erosion of the very capacity that keeps the system safe.

Your controls look fine. Your controls are running on borrowed judgment with no replenishment schedule.

Your controls look fine. Your controls are running on borrowed judgment with no replenishment schedule.

The honest counter-argument

There is a real case on the other side of the AI entry-level jobs debate, and it deserves to be made properly rather than strawmanned.

Junior professionals frequently love these tools. When large professional services firms gave legal staff AI assistants, it was often the junior staff who were most enthusiastic — because the software removed drudgery that never taught them much anyway. Nobody’s judgment was meaningfully improved by their four-hundredth precedent search. The optimistic version says a first-year at an AI-enabled firm can now reach, in months, work that used to take several years to earn access to.

That case is plausible, and if it holds it is genuinely good news. But it depends on a condition almost nobody is meeting: the time freed must be deliberately reinvested in judgment formation. If drudgery is removed and the recovered hours are converted into higher throughput — more matters, more clients, more output per head — then the firm has harvested the efficiency and skipped the compensating investment entirely. Acceleration requires somebody to consciously design what the accelerated path looks like. Absent that design, you have not compressed the apprenticeship. You have deleted it and told a flattering story about the deletion.

The distinction is not philosophical. It is visible in a budget line. If nobody has one, the reinvestment is not happening.

Rebuilding the ladder deliberately

The organisations that navigate the AI entry-level jobs shift well will not be the ones that hire the most juniors. They will be the ones that stop treating expertise formation as a by-product of billable work and start treating it as a manufacturing process with its own inputs, cost, and yield.

Separate learning work from economic work. For a century, apprenticeship was free because someone paid for the output. That subsidy is gone. Training has become an explicit cost rather than a hidden one, which means it must appear in a budget and have an owner — or it will not happen. As practitioners in the legal sector have already begun arguing, as on-the-job learning becomes scarcer at ground level, structured education has to expand to compensate.
Teach by reconstruction, not review. A junior who only ever checks AI output learns to recognise surface plausibility. A junior who builds it manually first, then compares against the machine, learns where the machine is wrong and why. The sequence is not interchangeable. Build, then compare — never compare alone.
Protect an error budget. Expertise is built primarily from being wrong under supervision. If AI absorbs all the low-stakes work, the first mistake a young professional makes will be on something consequential. Somebody has to deliberately manufacture safe places to fail, because the workflow no longer produces them for free.
Measure formation, not throughput. If your only early-career metric is output volume, AI wins that comparison permanently and you will conclude — correctly on that metric, and disastrously — that juniors are unnecessary. Track capability acquisition instead: what can this person now do unsupervised that they could not do six months ago? That is the only number describing whether your expertise factory is still running.

The institution question

Strip away the operational framing and something more fundamental sits underneath.

An organisation is, in its oldest and most durable sense, a machine for converting inexperienced people into experienced ones. That is what a guild was. That is what a firm was. The output — the ships, the audits, the buildings — was how the machine paid for itself, but its real function was continuity: taking someone who knew nothing and, over years, making them into someone who knew what the institution knew.

A company that stops doing this is no longer a company in that sense. It is a set of contracts with people who acquired their judgment somewhere else, running on a stock of capability it consumes and does not replace. It works beautifully right up until the last generation that learned properly begins to retire.

The Federal Reserve Bank of New York put recent-graduate unemployment at 5.6% in the fourth quarter of 2025, with underemployment elevated alongside it. That number is normally read as a story about young people’s prospects. Read it the other way — as a running tally of what the disappearance of AI entry-level jobs is quietly costing. It is a measurement of how much expertise the economy chose not to manufacture that quarter.

The firms that keep the ladder intact through this decade will not be doing it out of generosity. They will do it because in ten years, when every organisation on earth has access to identical models and identical agents, the only genuinely scarce input left will be people who know which outputs to trust — and the only way to have them will have been to have made them.

Judgment is the last constraint. It was always going to be. The strange thing is how quickly we dismantled the one process that produced it.

An organisation is a machine for converting inexperienced people into experienced ones. A company that stops doing this is running on a stock of capability it is consuming and not replacing.