The Entry-Level Puzzle, Revisited

A firm-level study of 21,559 US firms, released June 30, 2026, links AI spending to workforce records for the first time at scale. High-intensity adopters grew total headcount about 10% and entry-level headcount 12%, and tilted the mix toward juniors. That corroborates the Issue 03 call. The economy-wide entry-level freeze is real, and it sits among the firms that have not adopted.

A research addendum to Issue 03, "The Entry-Level Puzzle" (June 16, 2026), prompted by Kharazian, Simon and Stevens (Ramp / Revelio Labs). This is a research addendum and stands outside the numbered series.

THE CLAIM

Since Issue 03 we have held that the entry-level crisis is a problem of frozen hiring flows, that the stock of exposed work kept growing while firms stopped hiring juniors into it, and that employer attribution of the freeze to AI is narrative and not measurement. A firm-level study released today, the first to link observed AI spending to workforce records at scale, tests that split with data we did not have when we wrote, and on the contested point it strengthens the call. The firms spending most on AI grew employment by about 10% and entry-level headcount by 12%, and tilted their workforce toward juniors. The firms adopting AI are hiring juniors, and the economy-wide freeze is concentrated among the firms that are not. Where we were too loose, in the body lines that let AI adoption itself stand as the cause of the frozen junior rung, we mark the miss and correct it below.

• +10.2% total headcount at high-intensity AI adopters over 24 months, measured against firms that have not yet adopted (Ramp / Revelio Labs).

• +12.0% entry-level headcount under the same design; low-intensity adopters show no change distinguishable from zero.

• +1.15ppt entry-level share of the workforce against -1.52ppt manager-plus share, so the mix tilts toward juniors.

• Gains are statistically significant only in Information, the sector that already does most of the adopting (selection, flagged).

• 3.1 hires rate in February 2026 marks the economy-wide freeze, a level unseen outside 2008 to 2013 (BLS JOLTS).

01 · THE PRIOR

The standing view we are marking to the tape

Issue 03 made a deliberately narrow claim, and the narrowness matters here. We argued that the entry-level crisis was real while the popular diagnosis was wrong, since AI-exposed occupations were not shrinking, growing 10.4% in the stock data since 2019, while the constraint sat in hiring flows, with the JOLTS hires rate frozen at a level last seen climbing out of the financial crisis. We separated the stock of jobs from the flow of new hires, placed the squeeze in the flow, and said in as many words that whether AI or remote work broke the ladder was less settled than the headlines implied. We also parked a hypothesis we could not test cleanly at the time, that the freeze was monetary rather than technological, and gave it a named falsifier. The call turned on composition, since two facts that look contradictory are both true once you separate the universe of jobs from the universe of new hires. That structure is why today's study, which looks at first like a rebuttal, reads to us as a confirmation.

02 · THE NEW TAPE

The firms buying AI are hiring, juniors included

The study, by Ara Kharazian and Ryan Stevens of Ramp and Lisa Simon of Revelio Labs, links observed AI spending from Ramp card and bill-pay records to Revelio workforce histories for 21,559 US firms, and is to our knowledge the first to pair real firm-level adoption spending with employment outcomes where the literature has leaned on occupational exposure proxies. Its preferred design compares firms that adopt against firms in the same spending-intensity tier that have not yet adopted, with sector controls. The result runs against the destruction narrative on every margin the paper measures. High-intensity adopters grew total headcount by about 10.2% over the two years after adoption and entry-level headcount by 12.0%, while low-intensity adopters showed no change distinguishable from zero. The composition moved against the seniorization story, since the entry-level share of the workforce rose 1.15 points while the manager-plus share fell 1.52 points, so adopters tilted their staffing toward the juniors as they automated. The gains were broad across functions, with sales, administration, engineering, customer service and entry-level engineering all higher. The authors reach our attribution conclusion from the firm side, observing that some chief executives blame layoffs on AI even though their firm-level evidence contradicts the claim.

03 · THE RECONCILIATION

Two universes, one composition

The instinct to read the study as a contradiction comes from treating its numbers and ours as measurements of the same thing. Ours are economy-wide, drawn from the youth unemployment gap, the aggregate hires rate, and the occupational stock in the BLS and Census household and establishment surveys. Theirs are firm-level and relative, tracking an adopting firm's headcount against an otherwise similar firm that has not adopted yet. The two can diverge without either being wrong, and the mechanism that reconciles them is the one Issue 03 already used. A selected set of fast-growing adopters expanding their junior ranks is consistent with an economy-wide hiring freeze, provided the adopters are a minority and the freeze sits among the firms that are not buying AI. The study supplies the reasons to read it that way. The authors describe their adopters as larger, more technical, faster-growing and more often venture-backed, their sample as skewed toward tech-forward and knowledge work, and their employment gains as significant only in Information. Their definition of entry-level is a seniority field inferred from public professional profiles, a construct distinct from the age-based recent-graduate population whose unemployment Issue 03 tracked, so the two entry-level findings measure different things. Set the firm-level result beside the economy-wide freeze and the two describe different locations, since the hiring is concentrated among heavy adopters in one sector while the freeze covers the rest of the market. That is the hypothesis we parked in June, now with firm-level evidence behind it. If the firms buying AI are the ones hiring, the economy-wide freeze is being carried by the firms that are not, which locates the cause of the freeze in the macro and sectoral story we could not pin down at the time.

04 · THE CORRECTION

Where we were too loose, and what survives

Issue 03 was not vindicated whole. The literal scorecard claims hold, since exposed occupations are not shrinking, the market is safe for incumbents and shut to entrants, and attribution remains a choice. The body carried a causal insinuation our own data did not support, that AI adoption itself was the thing freezing the junior rung, and the firm-level evidence now locates the freeze elsewhere. The line that managers assume one senior worker with AI does the work of a senior plus two juniors, and that firms are already staffing as though it were true, reads differently once the firms staffing with the most AI are adding juniors faster than seniors. We should have held that mechanism as a hypothesis and written it as one, and we are marking it. The structure survives, and the error was letting AI carry a causal role in the freeze that the adopter data does not give it.

SECOND-ORDER EFFECT · THE ENTRY RUNG SPLITS BY FIRM AI-INTENSITY

The divergence Issue 03 measured, young workers against everyone else, now looks partly like a second divergence hiding inside the first, adopters against laggards. If the heavy adopters are expanding their junior ranks while the rest of the market freezes, then the entry-level labor market is bifurcating along the same line Issue 02 drew through compute budgets and Issue 04 drew through the model stack, the line between the firms compounding an AI advantage and the firms watching it happen. The young worker's problem becomes a sorting problem, since the jobs still exist but have concentrated in a particular kind of firm. The operators who rebuild the junior pipeline now are buying, at a graduate wage, the mid-level staff the laggards will be repricing in three years. For capital, the read is that entry-level dislocation is becoming a marker of which side of the adoption divide a firm sits on, and that the talent cost of being a laggard is deferred rather than absent.

WHAT WE'D DO · THIS QUARTER

1. If you are adopting AI, treat juniors as cheap capacity you keep. The firms furthest along in the data are doing exactly this, adding entry-level faster than they add managers.

2. Separate your freeze from the market's. A firm that is buying AI and freezing junior hiring is an outlier against the adopter evidence, and is likely cutting seed corn for a reason its own numbers will not support.

3. Recruit from the laggards. The buyer's market in trained juniors that Issue 03 flagged is, on this evidence, created mostly by the firms not adopting; the graduates are cheap because the laggards are not hiring them.

4. Discount the concentration before you generalize. The measured gains are real but significant only in Information; outside software and media, treat the adopter-hiring result as a leading indicator for your sector and confirm it before you act on it.

5. Keep auditing attribution. The new study lands where we did, that AI is named in cuts more often than it is measured in them; before you cite it, internally or externally, confirm your own numbers show it.

WATCHLIST · WHAT WOULD CHANGE OUR MIND

A representative replication. The Ramp / Revelio sample skews tech-forward by construction. A firm-level study on a representative panel that found adopters not hiring juniors would weaken the reconciliation, and we will weight it heavily when one appears.

The Information question. If the employment gains broaden beyond Information to professional services and other sectors in later adoption cohorts, the adopter-hiring story strengthens and the seed-corn risk for laggards sharpens.

The Stanford canaries, still. The Issue 03 falsifier stands. If the youth employment decline spreads from ages 22 to 25 into 26 to 30, the AI-causation story gains ground against our framing and we will say so.

The reconciliation's own falsifier. If better data shows the aggregate freeze is concentrated among adopters and not laggards, our composition read is wrong, and AI adoption moves back toward the center of the entry-level story.

SELECTED SOURCES

Kharazian, Simon and Stevens, "A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment" (Ramp and Revelio Labs, June 30, 2026; ramp.com/data/ai-jobs-impact) · Second Order Issue 03, "The Entry-Level Puzzle" (June 16, 2026) · BLS JOLTS hires rate via FRED (JTSHIR, through April 2026) · BLS OEWS with Pew Research exposure groups (Issue 03 calculations) · Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab.

Reported point estimates from the Ramp / Revelio study are plotted as such and are not Second Order calculations; the JOLTS series is computed from filed data (data/SOURCES.md). Second Order is an independent research briefing, provided for discussion purposes only; it is not investment, legal, tax, or accounting advice. © 2026.