The Work That Made Them
A beginner with today’s tools can produce competent work on his first day. In the same period, the jobs where beginners used to start have been opening less often.
Those two developments may not share a cause. Together they raise a question: where does a beginner get the experience the tool cannot supply on its own?
The gap the tool closes
The effect of these tools is not evenly distributed.
A study of customer support agents, published as NBER working paper 31161 and later in the Quarterly Journal of Economics, found AI assistance raised productivity about 14 percent on average, roughly 34 percent for novices, and almost nothing for experienced agents.
A field experiment with 758 Boston Consulting Group consultants found the same shape. On tasks the tool handled well, consultants completed 12 percent more work, 25 percent faster, at more than 40 percent higher quality. Below-average performers improved 43 percent against their own baseline. Above-average performers improved 17.
The support study also found something less convenient for the simple story. When the AI went down, workers kept part of their gains relative to their pre-AI baseline, and the effect was stronger among those who had used it longest. The researchers describe the mechanism as disseminating the practices of the more able workers, moving newer ones down the experience curve faster.
That matters. The tool is not only handing out documentation. It was built from what capable people actually do, and some of that transfers.
And the gap it opens
The same consulting experiment included a task placed outside what its authors called the jagged frontier, the uneven boundary of what the tool handles well.
On that task, consultants using AI were 19 percentage points less likely to reach the correct answer than consultants working without it. Experienced consultants were fooled too. The tool didn’t fail loudly. It produced fluent, confident, wrong work, and people shipped it.
Nobody publishes that boundary. It moves, and it doesn’t announce itself in the output. Which means somebody has to be answerable for checking.
The rung that’s thinning
Stanford’s Digital Economy Lab has tracked the employment side. Its August 2026 update reports that employment for workers aged 22 to 25 in highly AI-exposed occupations now runs about 19 percent behind comparable workers of the same age in less-exposed occupations, a gap that has widened from 15 percent a year earlier. The gap appears to come primarily from reduced hiring rather than layoffs. The door opens less often.
Two secondary findings matter more than the headline. The declines concentrate in work built on codified knowledge, the layer the tools supply. And employment rose among experienced workers in occupations that run on tacit knowledge.
The researchers are careful, and anyone citing them should be: they report no widespread economy-wide displacement, and they describe patterns rather than proving causes.
The case that this has nothing to do with AI
There is a serious counterargument, and skipping it would be dishonest.
The Economic Innovation Group makes a timing argument. Job postings in those same occupations peaked in the spring of 2022 and fell through that year, before ChatGPT was public. The job ladder is procyclical: when firms stop hiring and internal moves freeze, the youngest cohort takes it hardest, and that pattern is older than any model in this discussion.
That argument may be right. It doesn’t change what follows, because the question isn’t only what thinned the entry rung. It’s what that rung was doing.
What those jobs were for
Entry-level work was never just output. Any manager who has carried a first-year employee knows the arithmetic doesn’t close on production alone. For a while they cost more than they return.
What you were buying was transfer. The junior sat near people who knew things. He owned small pieces of real work, with his name on them, reviewed by someone who caught what he missed. He was in the room for the failures. He learned which vendor promises meant nothing, which alarm was always false, and which quiet remark in a meeting was the one that mattered.
Some of that is codified, and the tools now carry it. Some of it isn’t, and what makes the difference is not the technology. It’s whether the work is arranged so that a person still owns something real, gets corrected on it, and is present when things break.
That arrangement used to come bundled with the job. It doesn’t come bundled with the tool.
The cheap part and the expensive part
Free courses from major universities have been available since 2012, and Reich and Ruipérez-Valiente, tracking 5.63 million learners across MIT and Harvard offerings on edX, found completion around 3 percent by 2017-18. Making the material available did not, by itself, ensure people learned it.
Explanations are cheap now, and so is feedback: an answer when you’re stuck at eleven at night, correction at the moment you’re wrong rather than three weeks later. Though as the consultants discovered, immediate correction is not the same as correct correction. What remains expensive is supervised responsibility for work that matters, and no subscription sells it.
Where judgment gets spent
If the tools supply much of the codified layer, what an organization still needs is someone answerable for the rest: checking the fluent answer, recognizing the exception, diagnosing a failure nobody has seen before, and carrying the consequences of the call.
In my experience that responsibility concentrates where it always has, in the people an organization phones when something is strange. Sometimes that’s a team lead or a director. Just as often it’s the senior technician nobody outranks on the subject, the nurse who has seen this presentation before, the engineer who knows why the workaround exists. Rank is where judgment often ends up. It is not what judgment is.
And here is the position that creates. An organization can become more dependent on experienced judgment at the same time it hires fewer of the people who might develop it.
The bench nobody built
That organization is drawing on a stock it has stopped replenishing.
None of it shows up in a quarter. The work ships. The dashboards stay green, because the people making the calls are the ones who have always made them, and they’re very good at it.
If entry opportunities keep contracting and nothing deliberately replaces what they did, the shortage arrives later, quietly, in the form of a bench that was never built. And the replacement isn’t a subscription. It’s the unglamorous thing the old rung made room for: real ownership of small work, reviewed by someone who knows better, close enough to the failures to watch one happen.
The capability will still exist. It will just belong to whoever thought to keep building it.
Discover more from At Ground Level
Subscribe to get the latest posts sent to your email.
