21,559 Firms, Five Years: Committed AI Adopters Grew Entry-Level Hiring 12%

21,559 US firms. Five years of spend records. The finding runs opposite to the headline most people expected from AI and jobs: the companies that committed hardest to AI grew their total workforce, and grew entry-level hiring fastest of all.

The Ramp Economics Lab, working with Revelio Labs, linked corporate AI spending to workforce records from January 2021 through February 2026, then ran a staggered difference-in-differences design (the Callaway-Sant’Anna estimator) to isolate what happened after a firm started spending on AI. High-intensity adopters added 10.2% to total headcount over the 24 months following adoption. Entry-level roles grew 12%, faster than the workforce as a whole. The much-forecast collapse in junior hiring did not appear in this panel. Its inverse did.

One number here carries a warning label, and the study is the one that attached it. The caveats end up mattering as much as the result, so I will get to them in full. First, what the panel actually measured.

What the panel measured

This is the first large causal panel to link AI spending to real workforce records, and the design is worth stating plainly because it determines how much weight the numbers can hold.

The unit of observation is a firm’s monthly AI spend, drawn from Ramp’s payment data, matched against that firm’s headcount from Revelio Labs. Adoption is a spending event: at least $100 per month in AI spend, sustained for three consecutive months. That threshold turns a vague question (“did they adopt AI?”) into a dated, observable fact, which is what a difference-in-differences design needs. The Callaway-Sant’Anna estimator handles the awkward reality that firms adopted at different times, comparing each firm’s trajectory after adoption against firms that had not yet adopted, rather than against a single fixed control group.

The output is an event study: headcount plotted relative to the month a firm crossed the adoption threshold. That framing is why the result reads as causal rather than correlational. It is not saying AI-using firms are bigger. It is tracking what happens to a firm’s hiring after it starts spending.

Intensity is the whole story

The headline number hides the variable that does the actual work. Adoption alone moved nothing. Sustained, heavy use moved everything.

Ramp splits its adopters into high and low intensity by spend per employee: roughly $33.67 per employee per month at the high end, against $2.78 at the low end. The low-intensity group shows no statistically significant change in headcount. A company can buy a handful of seats, call itself an AI adopter, and watch its workforce numbers stay flat, because on this data they do stay flat. The 10.2% and 12% figures belong entirely to firms spending at the committed end.

The high-intensity effect is also not a one-time step. It compounds. The event-study coefficients start small, near 3.66 log points four months after adoption, and climb to 45.24 log points by month 24. Whatever mechanism links heavy AI use to hiring, it accumulates across two years instead of landing as a single jump. That trajectory has a direct operational consequence: a firm reading its AI program at month three is reading it before the workforce signal exists. The effect that shows up at two years is invisible at one quarter.

The caveats the authors volunteer

Here is the number with the warning label, and the label is the study’s own.

Ramp states that its adopters were already growing faster than non-adopters before they adopted. The firms inclined to spend heavily on AI were, on average, the firms already hiring aggressively. Difference-in-differences is built to strip out fixed differences between groups, but it cannot fully rule out that a rising firm both hires and buys AI for the same underlying reason. The authors say so directly.

Three more limits sit in the open. The 24-month window captures headcount but can miss workforce-mix reallocation, the slower story of which roles get redefined rather than added or cut. The panel skews tech-forward and venture-backed, exactly the population most able to convert AI spend into growth, and least representative of a regional manufacturer or a hospital system. And this is a vendor study: Ramp sells payment infrastructure to these same firms.

That last point cuts two ways, and the direction it cuts is the reason this piece exists. A vendor with an incentive to publish a flattering number instead published its own selection bias in the methodology section. That is what a credible measurement looks like. The honesty about what the data cannot prove is a governance signal in itself, and it is the standard any internal AI-impact dashboard should be held to. Reporting on AI that never surfaces its own confounds is a sales asset in the costume of measurement.

The governance reading

Read together, the result and its caveats point at one operating principle: workforce outcomes track how intensely a firm sustains governed adoption. The binary question every board is asking, “are we adopting AI or not,” is the wrong instrument. It cannot distinguish the firm spending $2.78 per employee and seeing nothing from the firm spending $33.67 and compounding a 10% headcount effect. Both answer “yes, we adopted.” Only one of them shows up in the data two years later.

That reframes the entire headcount conversation. The fear that AI adoption cuts jobs treats adoption as a lever pulled once. On this panel it behaves like a commitment sustained across years, and the firms sustaining it hardest hired more people, including more juniors. The real workforce risk sits in adopting casually, declaring it done, and expecting an effect the spending was never intense enough to produce.

Do this now

Stop measuring AI adoption as a yes or no. Measure intensity. Pull your last twelve months of AI spend and divide it by headcount, per month. If the number sits closer to $2.78 per employee than $33.67, you are in the group this study found no effect for, and no amount of “we adopted AI” language changes that. Then set the review horizon honestly. The effect in this data took two years to reach full size and did not exist at month three. Judge the program on that clock, govern the intensity deliberately, and hold your own reporting to the same standard the study set: name the confound before you cite the number.

This analysis synthesizes AI Jobs Impact (Ramp Economics Lab + Revelio Labs, June 2026), a working paper by Kharazian, Simon, and Stevens whose own methodology section flags the selection bias in its sample.

Victorino Group helps organizations measure AI adoption by intensity and outcome, not by yes-or-no, and govern the difference. Let’s talk.