The Career Ladder Lost Its Bottom Rung. The Screen Can't See the Next One.

There is a number that quietly broke a long-standing rule of the labour market, and almost nobody noticed the moment it flipped.

For most of the last forty years, a college degree was an unemployment shield. Recent graduates consistently had a lower jobless rate than the workforce as a whole. That was the entire deal: you took on the debt, you got the credential, and the credential bought you a softer landing into work.

By the end of 2025, the deal had inverted. The unemployment rate for recent college graduates climbed to roughly 5.7%, according to the Federal Reserve Bank of New York — while the overall rate sat near 4.2%. The people who did everything the system told them to do were now more likely to be out of work than the average person who didn't. The shield had become a liability.

And the place where that liability bites hardest is the one place every career has to begin: the first job. The bottom rung of the ladder is the part that's breaking. The uncomfortable part — the part this piece is about — is that AI automating junior tasks is only half the story. The other half is that the screening machinery built to fill those jobs was never able to see the candidates filling them, and now that blindness has consequences it didn't have before.

What actually broke

Start with the part that's real and measured, because it is.

In August 2025, a team at Stanford led by Erik Brynjolfsson published the first large-scale evidence on this, built from the payroll records of millions of American workers run through ADP — the country's largest payroll processor. The finding was stark and specific. Since generative AI went mainstream, employment for workers aged 22 to 25 in the most AI-exposed occupations — software development, customer service, and similar codified roles — fell by about 13% relative to less-exposed work. Employment for older, more experienced workers in those same occupations stayed flat or grew.

That is the signature of a very particular kind of disruption. It is not "AI is taking jobs." It is "AI is taking the jobs you get first." The tasks a junior developer used to cut their teeth on — the boilerplate, the debugging, the first-draft document, the routine research pull — are exactly the tasks a language model now does in seconds. Experience turned out to be the buffer. Inexperience turned out to be the exposure.

The venture-capital firm SignalFire put hard numbers on the hiring side of the same phenomenon. Across the largest public tech companies and VC-backed startups, the number of new roles going to people with less than a year of post-graduate experience fell by roughly 50% between 2019 and 2024. Big Tech cut new-graduate hiring by 25% in 2024 alone versus the year before; startups cut theirs by 11%. The share of new grads landing a role at one of the Magnificent Seven — Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, Tesla — has dropped by more than half since 2022.

Job-posting data tells the same story from a third angle. On Indeed, junior-level postings fell 7% year-over-year through 2025 while senior-level postings rose 4%. Companies didn't stop hiring. They stopped hiring at the bottom.

LinkedIn's chief economic opportunity officer, Aneesh Raman, named the pattern in a New York Times op-ed in May 2025 that went viral for a reason: "the bottom rung of the career ladder is breaking." He compared it to the hollowing-out of manufacturing in the 1980s — a structural removal of the entry point, not a temporary dip. Anthropic's CEO Dario Amodei went further the same month, telling Axios that AI could eliminate as much as half of all entry-level white-collar jobs within one to five years.

So that's the real, documented half: the tasks that made the first job worth offering are being automated, and employers are responding by offering fewer first jobs.

The part the headlines get wrong

Here's where the story usually stops, with a clean villain — the AI — and a tidy conclusion: the robots took the entry-level jobs.

It's not that simple, and the nuance matters, because it's the nuance that points at what's actually fixable.

By mid-2026 the apocalypse framing had cooled. Amodei's and Sam Altman's most dramatic predictions got walked back. The early aggregate data turned out messier than the warnings: employment for AI-exposed workers has stayed broadly stable since ChatGPT launched, and productivity has risen faster than unemployment. The New York Fed's Q1 2026 read showed the recent-grad jobless rate still elevated at around 5.7%, but underemployment edging down slightly to 41.5% — and Indeed's Hiring Lab reported in April 2026 that new grads were, on the whole, finding jobs somewhat faster than the gloomiest 2025 takes implied.

In other words: the entry-level market is genuinely strained, but it is not a smoking crater. Junior hiring didn't vanish. It got more selective. And "more selective" is not a story about AI replacing humans. It's a story about how companies decide which juniors are worth a bet — and that decision runs through the exact screening stack we've spent this entire series taking apart.

This is the pivot most coverage misses. When entry-level hiring tightens, the binding constraint stops being "are there jobs" and becomes "can you tell which inexperienced person is going to be good." That is a measurement problem. And it is the specific measurement problem that resume-and-keyword screening is worst on earth at solving.

Why the screen fails juniors first

Think about what an applicant tracking system actually measures, and then think about what an early-career candidate actually has.

The ATS reads the past. It parses a document for evidence of things already done — titles held, tools used, years accrued, keywords matched against a job description. The entire apparatus is a backward-looking pattern-matcher. Its core question is: does this person's history already contain the thing we're looking for?

A graduate's defining characteristic is that their history doesn't contain very much yet. That's not a flaw in the graduate. It's the definition of the category. They have a short trajectory, a thin keyword footprint, a vocabulary that hasn't been sanded into the industry's preferred terms, and accomplishments that live in coursework, projects, internships, and side work that no parser has a clean field for. Everything that makes someone early-career is everything the screen is structurally unable to read.

So the failure compounds in a way it doesn't for senior candidates. A senior engineer with a thin resume still trips enough keywords to survive the first cut. A new graduate with genuine ability and an un-optimised resume gets a low match score and disappears — not because the system judged their potential and found it wanting, but because the system has no mechanism for judging potential at all. It can only count what's already there, and by definition there isn't much there yet.

Now layer on the volume. Entry-level roles draw the largest application pools of any rung — there are simply more people qualified to apply for "associate" than for "director." The same AI that's compressing the number of openings is also being handed to every applicant to mass-produce applications, and a large share of submissions now carry AI-generated content; by early 2026 employers were openly asking applicants to stop, because the applications had all started to read the same. So the bottom of the funnel is where the flood is heaviest, the documents are most homogenised, and — by some estimates the majority of companies now let automated screening reject applicants with no human ever looking — the auto-reject is most aggressive.

Put it together and you get a brutal sorting machine pointed at exactly the people who can least afford it. The most volume, the thinnest track records, the most look-alike documents, and the least human oversight, all stacked on the one rung of the ladder where the candidate's value is potential rather than proof. The screen was built to find proof. It's standing guard over the rung where there isn't any yet.

This is the real reason "entry-level jobs that require three years of experience" became a punchline. It isn't a typo. It's the screen's logic made literal: a system that can only validate the past, applied to candidates who don't have one, will keep demanding a past until the category collapses into contradiction.

The pipeline bomb nobody is pricing

There's a second-order cost here that deserves to be said plainly, because most companies are making this trade without naming it.

If experience is the buffer against AI — and Stanford's data says it is, with senior workers untouched while juniors absorbed the hit — then a company that stops hiring juniors is optimising itself into a trap. Every senior you rely on was once a junior somebody chose to bet on. Cut the bottom rung for a few years to save on the roles AI can supposedly cover, and you wake up a decade later with no mid-career bench, no internal succession, and a generation of senior expertise you declined to grow.

You can't automate your way to a senior engineer. You can only grow one, slowly, from a junior who was given real work and real time. The firms treating entry-level hiring as the easiest line to cut are quietly defunding their own future org chart. The savings are immediate and legible. The cost shows up on someone else's quarter, which is exactly why it keeps getting made.

The companies that will look smart in 2032 are the ones that kept hiring early-career talent through the squeeze — but to do that without drowning, they need the one thing the current stack can't give them: a reliable way to tell which inexperienced person is worth the bet.

What you'd have to measure instead

If the problem is that the screen can only read the past and juniors don't have one, the fix isn't a better reader of the past. It's a different object of measurement entirely.

You stop scoring the contents of the document and start reading the shape of the trajectory — even a short one. An early career is not a flat absence of signal. It's a slope. Two graduates with identical thin resumes are not identical: one taught themselves a stack and shipped something real, took on harder problems each term, moved from following instructions to owning outcomes; the other accumulated the same nouns by sitting still. The keyword count is the same. The trajectory is completely different. Slope is visible long before scale is — if you're looking for slope.

That means evaluating the direction and rate of someone's growth, not just the current altitude. It means reading projects, the progression within an internship, the order in which someone took on responsibility, the evidence of someone repeatedly reaching past their level — the things that predict what a person will become, which is the only honest question you can ask about someone at the start. It means treating a non-standard path as data to interpret rather than a parsing error to discard. And it means being transparent about the logic, because "we bet on this junior and here's the pattern that convinced us" is a defensible decision in a way "the match score was 91%" never was.

This is not softer than keyword screening. It's harder, and it's more accurate, because it's measuring the thing that actually matters for an early-career hire instead of the thing that happens to be machine-readable. A keyword screen asks what you've already done. A trajectory read asks where you're going and how fast — which, for someone whose whole value proposition is the future, is the only question worth asking.

The bottom rung of the ladder broke partly because AI ate the tasks that used to sit on it. But it stays broken because the tools we use to refill it can only see candidates who've already climbed. The economy doesn't need a screen that's better at reading the past. It needs one that can finally see potential — because that's the only thing the people at the bottom of the ladder have ever had to offer, and right now nothing in the stack is built to look.

AgentR evaluates candidates on the trajectory of a real career — the slope, the evidence, the direction of travel — rather than the keyword density a resume happens to contain. That's the difference between a system that can only hire people who've already arrived and one that can spot the ones worth betting on before they do. If your funnel is auto-rejecting the future to protect the present, you're measuring the wrong thing. Let's talk.