The work that AI now finishes in seconds is the same work that used to turn beginners into experts. The employment data has started to show the cost, and the governance frameworks that regulated finance depends on quietly assume it is not happening.
Everything I know after over two decades in technology, I learned by doing work that AI now does in seconds. Overnight production support, chasing a batch job that failed at 3 am. Reading other people's code until I could feel where the defects lived. Reconciling numbers that refused to match until I understood how the whole system actually moved money and data. None of it was glamorous. All of it was formative.
That work is disappearing, and the disappearance is being recorded as a productivity gain. That accounting is correct as far as it goes. What it leaves out is that every knowledge profession has used exactly this kind of work, for at least a century, as its apprenticeship. This article looks at what the evidence now shows, what it does not yet show, why regulated financial services is the place where the problem stops being philosophical and becomes priced, and what a deliberate response would look like.
What the evidence shows
Until recently this was a hunch traded in hiring meetings. It is now measurable, and the most careful measurement comes from the Stanford Digital Economy Lab. In its August 2026 update to the study it calls Canaries in the Coal Mine, the researchers report that employment among workers aged 22 to 25 in highly AI exposed occupations now stands about 19 percent below where it would be had it kept pace with similarly aged workers in less exposed occupations. The shortfall was 15 percent in July 2025. The divergence has widened steadily since the team first documented it in August 2025.
Three details in that research matter more than the headline number. First, the adjustment operates mainly through reduced hiring of young workers rather than through layoffs. Nobody is being pushed off the ladder; the bottom rungs are simply not being installed. Second, experienced workers in the same occupations show no comparable gap. The effect is concentrated exactly where formative work used to live. Third, the authors are explicit that they do not see widespread, economy wide job displacement associated with AI. This is not a story about AI destroying work. It is a story about AI removing the entry point to it.
The pattern repeats wherever the data is granular enough to see it. SignalFire's 2025 State of Tech Talent report found that new graduates accounted for just 7 percent of hires at large technology companies, with new graduate hiring down 25 percent from 2023 and more than 50 percent from 2019. At startups the figure was under 6 percent. The firm's 2026 report continues the series.
The demand side of the shift is visible too. Microsoft's 2026 Work Trend Index reports fifteenfold year over year growth in active agents inside Microsoft 365, rising to eighteenfold in large enterprises, and frames the change with a sentence worth reading twice: as AI and agents take on execution, our own agency expands. That is true for the people who already have judgment. The question this article asks is how the next cohort acquires it.
The trade press has begun asking the same question in plain language. American Banker put it in a headline: when AI does entry level work, how do rookies learn the ropes?
What the evidence does not show
A research grade argument has to be honest about its limits, so here are the ones I see.
The Stanford findings are observational. They compare young workers in more and less AI exposed occupations over the same period, which controls for economy wide conditions such as interest rates or a post pandemic hiring correction, but it cannot prove that AI adoption is the cause of the gap rather than something correlated with it. The steady widening of the gap as adoption deepened is consistent with the causal story. It does not settle it.
The SignalFire data covers the technology sector, which adopted these tools earliest and most aggressively. It is a leading indicator, not a picture of every industry.
Microsoft's report is written by a company that sells agents, and its framing that agency expands is optimistic by design. It may also be right for experienced workers. Nothing in it contradicts the youth specific finding; it simply does not address it.
And the mechanism I describe below, that removing formative work removes the formation of judgment, is an argument from professional experience and from how expertise is known to develop, not a measured outcome. The measurement will arrive in about a decade, when today's 24 year olds are asked to supervise systems they never learned to do the work of. The point of this article is that waiting for that measurement is itself a decision.
Judgment is compressed experience
Here is what the productivity accounting misses. In every knowledge profession, the grunt work was never only output. It was the curriculum.
A junior lawyer reviewing thousands of discovery documents is not just finding the relevant ones; she is building an instinct for what relevant looks like. A junior credit analyst writing memos few people read is learning what risk feels like before it becomes a loss. A junior engineer on production support at 3 am is learning how systems fail, which is a different and more valuable education than learning how they work. An anti money laundering analyst clearing hundreds of false positive alerts is learning, slowly and expensively, what a true positive looks like.
Expertise is compressed experience. The compression happens through volume, repetition and consequence: doing the thing many times, being wrong, and finding out. Every profession built its career ladder so that the bottom rungs generated this compression while producing cheap, useful output as a byproduct.
AI has taken the output. The unresolved question is whether it has also taken the compression, and the honest answer is that it has, unless we rebuild it deliberately. Watching an AI do the work compresses almost nothing. Reviewing AI output without ever having produced the work yourself is like grading translations in a language you never learned. You can check the grammar. You cannot catch the lie.
Why regulated finance is the stress test
I spend my working life in financial services and I want to use it as the evidence here, not because this is only a banking problem, but because regulated industries are where the argument stops being philosophical. Here, judgment is not a desirable quality in senior staff. It is the product the regulator is buying when it licenses an institution.
Consider the two governance frameworks most institutions are building their AI programmes on.
The US Federal Reserve's SR 11-7 guidance on model risk management, in force since April 2011, rests on a principle it calls effective challenge: critical analysis by objective, informed parties who can identify model limitations and produce appropriate changes. The guidance says plainly that effective challenge depends on a combination of incentives, competence and influence. Institutions have spent fifteen years engineering the incentives and the influence, through independent validation functions and reporting lines. Competence was assumed to arrive on its own, because it always had. It arrived through the work.
The European Union's AI Act goes further. Article 14 requires that the natural persons overseeing a high risk AI system be enabled to properly understand its capacities and limitations, to remain aware of the tendency to automatically rely or over rely on its output, to correctly interpret that output, and to decide in any particular situation not to use the system or to override or reverse its output. Read that list as a job description. Every item on it is a description of judgment, and every item presumes the person holding the job acquired judgment somewhere.
Now run the clock forward. The people qualified to challenge an AI credit model in 2035 would, in the old world, be doing manual underwriting today. The people qualified to catch a subtly wrong automated claims decision would be handling claims by hand right now. The analyst who can tell a real sanctions hit from a false positive would be clearing alerts. They are not, because the AI is. We are writing governance frameworks whose supervisory layer we have stopped manufacturing, and the frameworks themselves have no clause that notices.
That is why I keep describing this as a systemic question rather than a human resources inconvenience. A bank that cannot produce senior judgment in ten years has a safety and soundness problem. An industry that cannot has a financial stability question. The same logic runs through medicine, law and software, everywhere society relies on a human being able to say the machine is wrong and be right.
Aviation already had this argument
Finance does not have to reason from first principles here, because another safety critical industry went through it a decade ago. In January 2013 the US Federal Aviation Administration issued Safety Alert for Operators 13002, warning that continuous use of autoflight systems could lead to degradation of the pilot's ability to quickly recover the aircraft from an undesired state, and encouraging operators to build manual flight operations into both line operations and training.
Notice the structure of that response. The regulator did not argue against automation, which had made flying dramatically safer. It recognised that the skill the automation relied on in an emergency was being eroded by the automation itself, and it asked operators to deliberately preserve a share of manual work as training load. Aviation reached that conclusion when the automation removed the middle of the pilot's task. AI removes the whole of the beginner's task, so the case for a deliberate response is stronger, not weaker.
Four design principles, each with a measure
The response I hear most often is that this will sort itself out, as it did with calculators, spreadsheets and search. Those tools removed drudgery from the middle of tasks. They did not do the analyst's entire job and hand it back for approval. AI does. So the first five years of a knowledge career have to be redesigned on purpose, by employers, because no university can simulate consequence. From what I have seen work inside technology organisations, the redesign has four parts, and each can be measured.
Produce before you review. No one supervises work they have never done. Before a new analyst approves AI generated reconciliations, code or credit memos, they produce a meaningful volume of that work by hand, with their errors surfaced and discussed. Measure: the share of reviewers in each function who have completed the production requirement before gaining approval rights.
Preserve a deliberate share of manual work. This is the aviation lesson applied to knowledge work. Route a fixed share of live volume, even five or ten percent, through humans without AI assistance, not as punishment but as training load, and rotate everyone through it. Measure: manual share by function, reported alongside the automation rate rather than buried beneath it.
Teach AI supervision as a craft. If the junior job is now to check the machine, teach checking as a discipline. Give juniors AI output seeded with known errors and score what they catch, with seniors grading the grading. A cohort that has learned to distrust fluent output is the most valuable asset an AI era institution can build, because AI failures are fluent by nature. Measure: seeded error detection rates over time, by cohort.
Make explanation the deliverable. The fastest way to force compression is to require the junior to defend the answer rather than deliver it. If the AI drafted the credit memo, the analyst presents it without notes to someone senior whose job is to ask why. What cannot be explained is not approved. Measure: explanation pass rates, and the proportion of AI outputs rejected at that step.
None of this is exotic. All of it costs money in precisely the quarter in which AI was supposed to save money, which is why it will not happen by drift. It happens when a board asks a direct question: show me the ten year plan for producing the people who will supervise the machines.
The question for your own institution
The Stanford researchers were careful to say that their data shows displacement at the entry level, not a collapse of work. I believe them. This article is not an argument that AI destroys careers. It is an argument that AI has quietly privatised the cost of training and socialised the risk of not training, and that in regulated finance the risk lands on the public.
Every leader I know is proud of what AI now does for their organisation. Few can answer a simpler question. In your institution today, what does a 24 year old do that will make them worth listening to at 40? If the answer is that they watch the AI, you do not have a talent strategy. You have a decade long bet that judgment will appear from nowhere.
I learned my profession on the bottom rung. The rung is gone, and I understand why; I helped automate some of it. But the people my generation will eventually hand these systems to are 24 years old right now. Somebody has to decide, on purpose, how they earn their scar tissue. The machines will not do it for them. That, at least, is still our job.