The Vanishing Entry Level
New Stanford/ADP dashboard data shows entry-level roles contracting 3.8% per year in AI-exposed occupations, while older cohorts (35-50) hold steady.
Once upon a time, junior designers earned their stripes on the slow and steady path. Depending on their date of entry, that might have been designing websites that didn’t convert, or building flows that confused users. Then they were shown how performance metrics suffered, or they sat behind the glass during usability testing while a senior designer traced the confusion back to decisions the fledgling designer made three sprints earlier.
Junior developers earned it in similar fashion. Maybe it was an accidental reveal of not-ready-for-release site content, or perhaps a 2 a.m. troubleshooting session that traced a production breakdown back to their front-end framework implementation choices. Different rooms, but the same mechanism. The lessons we learned before AI was on the scene? They cost us something, yes. But it’s that sting that made them stick, and with enough mistakes under our belt and the learning that lies therein, we develop judgment.
Of course, precisely how that pipeline was configured was never sacred. Just familiar and commonplace. You might even go so far as to call it common sense.
Calculators erased the requirement to do math in one’s head. Digital layout tools led to no more hand lettering. CAD erased the bottleneck of hand drafting.
Each time these changes landed on our doorstep, a chorus of practitioners insisted the new abstraction layer would produce generations that couldn’t really do the work, and could only operate the machines that were doing the work. But each time, judgment reconstituted itself one layer up. The “entry-level pipeline” was never a fixed pathway, it’s always been whatever the current domain, team and process kept at the bottom rung.
Why is this time different? I believe that AI advancements and adoption are so visible and widespread, this time we can watch the abstraction layer shifting around. We’ll either see the pipeline reform or see it fade away completely. We’ll see it clearly in the data, which no one’s arguing we have more of than say, when calculators first rolled out.
The Fork In the Data
In August 2025, a Stanford Digital Economy Lab team led by Erik Brynjolfsson published a finding that has since grown into a live-time dashboard built with ADP Research, covering roughly 4.6 million workers across more than 730 occupations.
The data’s headline is that employment for AI-exposed workers ages 22 to 25 is contracting at roughly 3.8% per year as of the most recent data, and that the decline has accelerated since the first year of measurement. Workers 35 to 40 in the same occupations are growing at about 2% annually. The squeeze reaches past new grads, as well. The 31-to-34 cohort is down about 1.7% year over year.
Across all workers, at any age, in any domain and context, AI-exposed occupations are down just 0.2%. Generally, jobs are holding on.
Within the same firm, entry-level hiring in the most AI-exposed roles fell about 13% relative to less-exposed roles.
It’s the On-Ramp That’s Collapsing
The Fortune piece covering the dashboard put it plainly: AI “absorbs tasks before it absorbs jobs.” It reaches first for retrieving, summarizing, scheduling, formatting. The tasks junior people have always been handed, because those tasks never required years in the seat.
It’s worth noting that the Stanford team thankfully ran some controls on the research, taking interest rates, pandemic overhiring corrections, and the pervading sense of global despair mostly off the table as competing explanations for this collapse of the entry level pipeline.
ADP’s chief economist, Nela Richardson, points at the operative variable: whether an organization deploys AI primarily as automation or as augmentation. Occupations where AI is brought in and augments human work show durable employment growth. Occupations where AI is introduced and automates tasks outright? Those are shrinking.
Same Tech, Very Different Outcome
The difference lives in an organization’s collective mindset and in the decisions that company’s leadership makes across several factors (talent training & retention, AI deployment and the usual build/rent/buy decisions). These decisions typically roll out without the “talent strategy” label, though that’s where it appears things are headed.
Hold that thought, because it’s the crux of everything that follows. Why? Because supervising AI output requires exactly the judgment we are quietly ceasing to train when AI is deployed primarily to automate entry level work vs augmenting those occupations in a sustainable fashion.
A Strong Counterargument
I want to hand over the mic for an optimistic read on this debate, because I believe it has strengths, and a potentially viable path forward.
A principal designer I respect at a mid-size HR-tech company pushed back on an earlier piece of mine, one arguing that AI adoption orchestration is becoming a dedicated role. Her stance on the ‘emerging role’ question, in her own words:
“I’m not entirely convinced this is a dedicated role. If the company creates a culture of knowledge sharing, it’s a great way to surface workflows that may benefit others… I think each person should be empowered to build what works for them specifically, share when appropriate, and adopt builds from others if useful.”
Her org runs on that model today. A dedicated channel, with people sharing AI workflows and skills proactively. No gatekeepers required. Having spent time in the very culture and org this approach is working in, I wasn’t surprised to hear it appeared to be effective in mitigating some of the worst aspects of AI Onboarding Debt.
But beyond a thoughtful, transparent culture where folks share with intention and strategy, there’s an even stronger case to be made for saving the entry-level judgment pipeline itself.
Judgment Comes from Decision Cycles, Not Keystrokes
The grunt work (design critiques, code reviews, support ticket triage) was never the point. Rather, it’s been the fee those early in their career had to pay to move up, to take on more responsibility and be given more power.
If AI waives that fee, a junior designer who once would have owned five flows across two years can now learn fifty lessons vs. five. They see dozens of cycles go from project to shipped to results in the same span of time. A junior developer sees the same shift, from five branches full of hard-won wisdom to fifty shipped-and-broken cycles. More repetitions, not fewer.
The traditional pipeline wasn’t the most ideal filter in the first place. It selected for who could (and was willing to) endure years of low-leverage work, and for who had the pedigree to get hired into the pipeline at all. Neither is a great predictor of judgment.
The Fork Itself Is an Optimist’s Best Evidence
The Stanford/ADP finding doesn’t say “AI destroys entry-level work, full stop”. It says entry-level work contracts where AI is deployed as automation, vs how entry-level work holds on just fine where AI gets rolled out as augmentation. That’s proof that a viable, extensible path is live right now, and visible in the same data-driven dashboard. An org that deliberately pairs their junior staff with AI as a force multiplier (versus cutting staff and trusting that shiny new agent) is running the augmentation condition, even if they don’t frame it that way.
Where the Counter Doesn’t Hold Up
The verification bootstrap
Today’s design and development veterans catch bad AI-generated decisions because they made enough bad decisions of their own to spot them before they cascade. The next cohort will inherit that verification mantle, but sans the specific quality of cycles that make recognition possible for those who’ve been around the block a few times. Those fifty watched cycles, blissfully absent of grunt-work and guaranteed sting-free? They don’t have the same stressors, and don’t dish out the pain that spurs judgment the way that five lived-cycles do, at least not in most domains.
A newbie pilot’s required flight-simulation hours compile into judgment because the feedback is fast and unambiguous. Either the plane crashed or it didn’t.
Design work rarely offers that, and neither does most engineering above the unit-test level. When a junior designer’s flow gets shipped and the confusion it causes surfaces as a slow bleed in adoption metrics three quarters later, attribution is hard because it’s tangled with someone else’s redesign and a rebrand marketing swore was crucial for their Q4 goals. A junior dev’s architecture calls look fine until it’s load-bearing under features they hadn’t scoped and didn’t have the experience to proactively account for.
Delayed, diffuse, and hard-to-attribute feedback doesn’t write to long-term judgment the same way a simulated stall warning does for that newbie pilot in training.
What the Data Reveals About Choice
Even if we grant the augmentation model everything it appears to offer, that path still has to be chosen. The ADP Research dashboard is the record of what’s actually being chosen at scale: role contraction is concentrated in the same place as automation-mode AI deployment, with the net effect growing month over month by roughly half a percentage point.
It’s not a data artifact, it’s the result of deployment decisions by thousands of firms, made mostly by default. A junior IC navigating cycles with judgment-producing stressors (the augmentation model) built in is a cost center without a plainly visible, near-term payoff.
Nobody is budgeting for the learner, and four years of data have revealed the preference for the vast majority: nobody is about to.
Even If Augmentation Wins, There’s a Cost
Say augmentation wins outright and judgment keeps forming. It’s still worth naming what that victory actually costs, because the maturation process isn’t instant.
The advancement of digital layout tools produced designers whose judgment formed differently, yes. However, it formed no less rigorously than those who did hand lettering and used rubylith. It also resulted in a period where the in-between cohort absorbed the digital disruption, without a playbook and with a dual learning-and-practice load. They carried both digital and analog knowledge and techniques, without any extra for the trouble.
The 2023–2030 workplace entrants are in that awkward phase right now, whether you fall on the pessimistic or optimistic side of this topic. And while I don’t have a confident estimate for how long this phase will run, 5-7 years would be my guess based on prior precedents and the degree to which AI seems to confound precedent entirely.
The Category Error
Here’s where the real question and the designer’s sharpest observation come together. Her sharpest observation wasn’t the culture argument, but a question she raised in passing about the hard part of AI deployment and adoption.
“The socialization, governance & trust that comes along with it… whose responsibility is it?”
Organic knowledge-sharing culture solves for process and tooling diffusion. In an org doing things right, relevant and effective workflows built by one person quickly reach the rest of the org, even without a dedicated role driving that outcome. That’s a real problem, a real source of pain I’ve had several seniors convey to me, and one that the model her org practices addresses well.
That said, formation of judgment is a different problem wearing similar clothes. A shared channel of AI workflows may tell a junior how to build something, or perhaps even what to build.
But it doesn’t put them through the cycle of thinking, building and shipping, then watching the consequences land. It doesn’t knock them down, make them wince and then clamber up and step back to the drawing board.
We need to find new ways of building in the reps the verification bootstrap shows us are nonnegotiable. While conscious diffusion can move artifacts between people who already have judgment, it doesn’t manufacture judgment in people who don’t yet have it. An org can run a flawless sharing culture and still have zero judgment formation happening underneath it.
In this scenario the collapse just happens quietly, even inside companies with excellent knowledge sharing. A few years on, there won’t be any juniors capable of exercising judgment and independently making calls that require it.
I’m not claiming my designer friend’s model is wrong, and it answers the diffusion question well. But her own “whose responsibility is it?” still sits on the stack one layer down, where the artifacts stop and the judgment has to kick in.
Who funds the learner
If the fork is real, and the data says it is, simply naming your org’s AI deployment strategy as “augmentative” doesn’t mean the judgment-formation pathway will build itself, even inside the best sharing cultures in the industry. Someone has to figure out what a junior must do across those fifty watched cycles so that watching turns into judgment.
The apprenticeship cannot consist solely of watching fifty data points of exposure. There must be real stakes. There have to be consequences a junior human gets to feel, and real accountability for catching what the AI gets wrong.
Someone has to build the structured apprenticeship interlaced with using AI to speed up and automate what doesn’t contribute to building judgment. That’s a role missing from most headcount plans, and the revealed-preference data says it’s losing the budget argument every quarter the dashboard reports.
So the age-old mechanism survives: entry-level to senior decision maker. Judgment can still form on the far side of this AI transition, the same way it formed on the far side of every transition before it. But it won’t form by default, and the question every org is answering right now, whether by strategy or by default, is the one this piece can’t answer for them:
Who funds the learner?