Research update · 15 September 2026
The first AI-spending test says heavy adopters added staff. It is not the last word on entry-level risk.
The loudest AI-jobs claims often start with a task-exposure score. A new firm-data working-paper presentation starts somewhere closer to actual rollout: corporate spending on AI vendors. Its result points in the other direction for a selected set of U.S. firms.
What this study measured
Ramp Economics Lab and Revelio Labs say they link firms’ Ramp-recorded spending with AI vendors to Revelio workforce data. Their analysis covers more than 21,000 U.S. firms. They classify high-intensity adopters as the top third of AI spend per employee in the first three months after adoption; Ramp says this is roughly $30 per employee per month, not a universal definition of adoption.
What it found
Ramp reports that high-intensity adopters grew total headcount 10.2% over two years after adoption and entry-level headcount 12%. Low-intensity adopters had no statistically significant headcount change. The reported gains appeared only after six to 12 months, which the authors interpret as an organizational learning period.
Why the result cannot prove AI created those jobs
The authors say adopters were already larger, more engineering-intensive, more likely to be venture-backed and faster growing. The data describe Ramp-linked U.S. firms, and the accessible report is published by a company whose product is corporate-spend management. That makes the observed spending measure valuable, but it does not remove selection or make the outcome representative of every employer, occupation or worker.
Why it does not erase the early-career warning
Stanford’s 19% payroll gap compares workers aged 22–25 across occupations with different AI-exposure scores. The Census papers compare exposed industry-state cells and college majors. Ramp compares firms with different spending intensity and reports headcount, including an entry-level category. These are not the same treatment, people, comparison group or outcome. A growing AI adopter can coexist with a tighter route into particular exposed occupations.
The bottom line
The result blocks an easy claim: exposure is not evidence that actual AI adopters are broadly cutting staff. It does not license the reverse claim that adoption makes a reader safe. The next decisive studies must observe adoption and rival changes such as remote work, then track comparable workers’ hiring, employment and pay.