A new study of 21,559 US companies links actual AI spending to actual workforce records for the first time, and finds that intensive adopters grew headcount by 10%, including entry-level jobs. The finding is important. The fine print is just as important.
Two serious research teams have now looked at AI and employment using large-scale administrative data, and they have produced headlines that appear to point in opposite directions.
The first, from Stanford’s Digital Economy Lab, is the “Canaries in the Coal Mine” paper by Brynjolfsson, Chandar, and Chen. Using payroll records from ADP, it found that since the widespread adoption of generative AI, early-career workers aged 22 to 25 in the most AI-exposed occupations have experienced a roughly 16 percent relative decline in employment (a figure revised upward from 13 percent in the paper’s initial August 2025 version). It has become the most cited piece of evidence that AI is already displacing young workers.
The second was published on June 30, 2026 by economists at Ramp and Revelio Labs. “A New Look at AI’s Impact on Jobs” by Kharazian, Simon, and Stevens links actual firm-level AI spending to workforce records for 21,559 US firms and finds the opposite-sounding result: companies making the largest AI investments grew total employment by roughly 10 percent in the two years following adoption, and their entry-level headcount, the demographic the Stanford paper worries about, rose 12 percent.
Both papers are careful. Both use data most researchers never get access to. And, as this post will argue, both can be true at the same time. Understanding why requires looking at what the new paper actually did, what it found, and, most importantly, what its methodology can and cannot support. That last part gets its own detailed section at the end, because with research this consequential, the fine print is not a footnote. It is the story.
Measuring Adoption Instead of Guessing It
The Ramp and Revelio paper’s most important contribution may be methodological rather than empirical.
Until now, almost nobody studying AI’s labor effects could observe which firms actually adopted AI. The dominant workaround has been occupational “exposure scores”: estimates of how automatable a job’s tasks are in theory, derived from sources like GPT-4’s own assessment of O*NET task descriptions (Eloundou et al., published in Science), AI patent text matching (Webb), or capability benchmarks (Felten et al.). These scores vary across occupations but not across firms, and not over time. Two companies employing identical software developers get identical exposure scores even if one runs on coding agents and the other has banned ChatGPT.
The scores also disagree with each other. Research from the Budget Lab at Yale comparing seven exposure metrics found that computer programmers rank in the 99th percentile of AI exposure under one index and the 88th under another. A related advance came from the Anthropic Economic Index, where Massenkoff and McCrory measured which tasks workers actually delegate to Claude rather than which tasks are theoretically automatable, and found a striking gap between the two: only 33 percent of Computer and Mathematical tasks show meaningful observed usage despite 94 percent being theoretically feasible. But even that measure works at the occupation level, not the firm level.
The new paper cuts through this by observing money. Ramp is a corporate card and bill-pay platform, and the researchers can see every payment its customers make to AI vendors: foundation model subscriptions, API tokens, GPU cloud, coding agents, inference platforms. They define adoption as the first month of a sustained spell of at least $100 of AI spend for three consecutive months, which filters out one-off experiments. They then link these firms to Revelio Labs workforce records, built from millions of online professional profiles, to observe monthly headcount, seniority, roles, and estimated salaries at the same companies from January 2021 through February 2026.
For the first time at scale, the adoption decision and its workforce consequences are visible in the same dataset.
What They Found
The headline numbers come with a condition attached, and the condition is the finding.
The researchers split adopters by intensity, measured as AI spend per employee over the first three months after adoption. High-intensity adopters, the top tercile, spend an average of $33.67 per employee per month. Low-intensity adopters, the bottom two terciles, spend $2.78. That gap roughly separates firms buying serious infrastructure, coding agents, API usage, multiple models and vendors, from firms buying chat subscriptions.
The results split along exactly that line. High-intensity adopters grew total headcount by 10.2 percent over the 24 months following adoption, relative to comparable firms that had not yet adopted. Entry-level headcount rose 12.0 percent, statistically significant at the 1 percent level. The growth was broad across functions: sales headcount up 10.3 percent, engineering up 7.3 percent, administrative roles up 7.8 percent, customer service up 6.3 percent, entry-level engineering up 6.3 percent. Operations was the only category with no detectable growth.
Low-intensity adopters, meanwhile, showed no statistically significant change in anything. Not total headcount, not entry-level jobs, not any role category. In the authors’ words, enterprise chat subscriptions do not appear to be enough. The employment gains appear only among firms making sustained, material investments in the technology.
The Learning Curve
The time path of the gains is as informative as their size. The effect on total headcount is essentially zero in the month of adoption (0.003 log points), still barely visible at month three (0.020), then climbs steadily: 0.071 by month six, 0.188 by month twelve, 0.277 by month eighteen, roughly a 32 percent gain. The month-24 estimate is larger still, but the authors are appropriately cautious about it because fewer firms are observed that far out and the confidence intervals widen considerably.
Nothing about this profile looks like a technology that pays off at purchase. It looks like a technology whose benefits require complementary investment: months of identifying use cases, integrating tools into workflows, establishing practices, and only then converting productivity into expansion and hiring. That reading aligns with a theme that has appeared repeatedly on this blog, most recently in the Logicalis CIO report findings: AI adoption is an organizational capability, not a procurement decision. Firms that buy subscriptions, run pilots, and stop are, on this evidence, spending money to stand still.
Where the Gains Live, and Where They Don’t
The sector breakdown deserves more attention than it will get in most coverage. The employment gains are statistically significant in exactly one sector group: Information, which covers software, internet, and media firms, where high-intensity adopters grew headcount by about 13.4 percent with clean diagnostics. Professional and technical services shows a positive but insignificant estimate. Every other sector group is close to zero, imprecise, or both.
Adoption itself is just as uneven. As of December 2025, 53.7 percent of Information firms in the sample had crossed the sustained-adoption threshold, against 14.3 percent in health care, 11.5 percent in accommodation and food services, and 11.3 percent in arts and entertainment. Larger firms adopt more (41 to 43 percent above 250 employees versus 12 percent below 10), and adoption peaks among firms where engineers are 30 to 50 percent of the workforce.
This is where the apparent contradiction with the Stanford paper starts to dissolve. The two studies measure different things. The Stanford work measures occupation-level displacement across the whole economy: whether young workers in AI-exposed jobs are losing ground relative to others, wherever they work. The Ramp and Revelio work measures firm-level growth among adopters: whether companies investing in AI expand relative to comparable companies that have not yet done so. Both can be true simultaneously. AI-adopting firms can be hiring aggressively while aggregate demand for certain entry-level occupations shrinks, because the firms doing the hiring are a selected, fast-growing minority, and because displacement can happen at the thousands of firms that adopted lightly or not at all. Growth concentrating inside intensive adopters is not the same as growth for everyone.
What the Two Papers Together Suggest
Read jointly, the emerging picture is redistribution rather than simple destruction or creation. Employment appears to be shifting toward firms that adopt AI intensively and know how to convert it into growth, and away from exposed occupations elsewhere. That is a meaningfully different diagnosis than either “AI is killing jobs” or “AI is creating jobs,” and it has different implications: the relevant question for an individual worker is less “is my occupation exposed?” and more “is my employer on the right side of this shift?”
The authors’ own closing advice is unusually direct for an economics paper and worth quoting in substance: if you are a young person choosing between otherwise similar firms, choose the one using AI. If you are an engineer worried AI will eliminate engineering jobs, the firms adopting AI are hiring engineers faster, not slower. And if you are reading headlines where CEOs blame layoffs on AI, be skeptical, because the firm-level evidence says the opposite.
That skepticism, however, must run in both directions, which brings us to the section this article has been building toward.
What This Study Can and Cannot Prove
The paper is the best firm-level evidence we have on AI adoption and employment. It is also a working paper by industry economists using proprietary data from a selected sample, with diagnostics the authors themselves flag. Here is the honest audit, point by point, with the paper’s own numbers.
Selection is enormous, and the authors show it. Firms that adopt AI were already different before adopting: 40 more employees on average, 4.4 percentage points faster year-over-year headcount growth, $16,499 higher mean salaries, 34 percent VC-backed versus 9.5 percent of never-adopters, and 54.2 percent in tech-adjacent sectors versus 25.9 percent. The paper’s Figure 4 (page 19) shows adopters on a visibly steeper growth path before adoption. Any simple comparison of adopters to non-adopters confounds AI with being the kind of firm that adopts AI.
The fix is clever but changes the question. The preferred design compares early adopters to later adopters in the same intensity group, before those later firms adopt, using the Callaway-Sant’Anna staggered-adoption estimator. Pre-adoption trends align closely under this design, which supports the identification. But the estimate then answers a narrower question: what happened to firms that chose intensive adoption, relative to similar firms that would choose it slightly later. It cannot tell us what would happen if a randomly chosen firm, or a reluctant one, adopted AI. The never-adopter comparison in Table 6 (page 36) shows why this matters: it produces larger estimates but fails its own diagnostics spectacularly, with 11 of 11 pre-periods flagged, and the authors correctly demote it to descriptive evidence.
Even the preferred diagnostics are not spotless. The high-intensity total-headcount estimate carries 3 flagged pre-periods out of 11, meaning high-intensity adopters showed some differential movement even before adopting. The education outcomes are weaker still (5 of 11 flagged for MBA headcount, 4 of 11 for bachelor’s). And the eye-catching month-24 estimate of roughly 57 percent comes from the thinnest slice of the panel with the widest intervals; the month-18 figure of about 32 percent is the more defensible number, and the paper says so.
The sample is not the economy. Every firm in the study is a Ramp customer with at least $5,000 in monthly spend over six consecutive months and at least five employees. Ramp customers skew tech-forward and business-spend-active. The paper’s own benchmarking makes the selection visible: roughly 42 percent of its panel qualifies as AI adopters, versus 17 to 20 percent of US businesses in the Census Bureau’s nationally representative survey. Revelio’s workforce data, built from online professional profiles, has known coverage gaps across occupations and industries that reweighting only partly corrects. The results describe a selected slice of corporate America, and the paper is upfront that sector-level gains outside Information may simply not be measurable yet in this sample.
The mechanism is a black box. The authors state it plainly: we do not know which operational practices generate the growth. Product acceleration, engineering leverage, sales productivity, support automation, new business lines, all plausible, none identified. Nor can stories adjacent to reverse causality be fully excluded: a firm anticipating expansion may simultaneously ramp AI spending and hiring, and clean pre-trends reduce but do not eliminate that concern, especially with 3 of 11 pre-period flags in the key group.
Consider the source, then read it anyway. This is a working paper, not yet peer-reviewed, written by economists at Ramp and Revelio Labs studying Ramp and Revelio data. That is exactly why the dataset exists at all, and the paper’s transparency about its own weaknesses is well above the standard of most industry research. But the incentive structure belongs in the ledger: a fintech platform publishing evidence that its customers’ AI spending correlates with growth is not a disinterested party, in the same way this blog flags when AI labs publish favorable findings about their own models.
None of these caveats overturns the finding. Together, they bound it: among tech-forward, spend-active firms that chose intensive AI adoption, employment grew substantially and broadly in the two years that followed, with the gains so far concentrated in the Information sector. That is what the data supports. “AI creates jobs” is not, and neither is “AI adoption would grow your company.” The paper’s real contribution is raising the evidentiary bar in a debate that has been running on proxies, surveys, and CEO anecdotes, and by that standard it deserves both the attention it will get and the scrutiny this section has tried to apply.
Key sources: