Researchers at Ramp Economics Lab and Revelio Labs linked observed corporate AI spending โ actual card and bill-pay transactions, not survey answers โ to workforce records across 21,559 U.S. firms. Companies in the highest AI-investment tier grew total headcount 10.2% in the two years after adoption, and entry-level headcount 12%. Companies in the low-intensity tier saw no statistically significant change at all (Ramp Economics Lab, 2026).
Not a smaller gain. Zero.
That contrast is the finding worth your attention, and it is not the one the headlines picked up. The interesting result is not that AI adoption and entry-level hiring can move together. It is that adoption itself explained nothing. Intensity did all the work.
The Study Measured Spending, Not Sentiment
Most of what a Head of Operations reads about AI adoption is self-reported. Someone surveys executives, executives describe their own maturity, and the resulting curve reflects confidence as much as deployment. The measurement problem is well established at this point โ you have probably seen your own leaders rate the company's AI capability well above what the deployment data supports.
The RampโRevelio design sidesteps it. AI spend is taken from Ramp's corporate card and bill-pay records โ money that actually left the company โ and joined to Revelio Labs' workforce records for the same firms. Firms were then classified as high- or low-intensity adopters based on AI spending per employee in the first three months after adoption, and their employment tracked forward (Revelio Labs, 2026).
That construction matters for how much weight the result carries. Nobody in the sample was asked how transformative their AI program felt. The independent variable is a transaction record.
It is, as far as the authors claim, the first study to link firm-level AI spending to workforce data at this scale (Ramp Economics Lab, 2026). For an operating leader, that is the difference between a data point you can plan against and one you can only quote.
Intensity Is the Variable, and It Behaves Like a Threshold
Run the two tiers side by side and the shape of the result is unusual.
High-intensity adopters: +10.2% total headcount over two years, +12% entry-level, with gains appearing gradually and spread broadly across job functions โ engineering, sales, administration, and customer service (Ramp Economics Lab, 2026). Low-intensity adopters: nothing statistically distinguishable from no adoption at all.
If AI investment produced returns roughly in proportion to what you spend, you would expect the low tier to show a small positive effect โ muted, noisy, but present. It does not. The relationship is not linear, and half-measures do not land halfway.
Read that as an operating constraint. A modest allocation spread thinly across eleven departments is not a smaller version of a serious program. On this evidence, it is a different category of activity, and it does not appear in the workforce data at all.
That should reframe how the budget conversation gets run. The default mid-market posture โ a little AI in every function, so no one feels excluded and no single bet can embarrass anyone โ is precisely the profile the low-intensity tier describes. It is a portfolio built to minimise regret, and it returns nothing measurable.
The Entry-Level Number Cuts Against the Canary
The 12% figure lands in the middle of an active argument, and it is worth being precise about what it does and does not settle.
Stanford's Digital Economy Lab, working with ADP payroll data, documented a roughly 13% relative decline in employment for early-career workers in the occupations most exposed to AI โ the "canaries in the coal mine" result that shaped much of the 2025โ2026 entry-level narrative (Stanford Digital Economy Lab, 2025).
Both findings can hold simultaneously, because they are not measuring the same object. Stanford tracks occupations: within AI-exposed job categories, young workers lost ground. Ramp tracks firms: within companies investing heavily in AI, entry-level headcount rose. Some roles are shrinking across the economy while some firms are hiring more junior people than before. Aggregate occupational exposure and firm-level investment intensity are different axes, and an operator sits on the second one.
Which is the practically useful reading. You do not control whether your industry's task mix is AI-exposed. You do control your firm's investment concentration โ and that is the variable the firm-level data says moved.
The temptation is to pick whichever study confirms the plan you already have. Resist that. The honest synthesis is narrower and more useful: exposure sets the pressure on a role, intensity of investment appears to set whether your organisation grows into that pressure or contracts under it.
The Counter-Argument: This Might Be Selection
The strongest objection to the Ramp result comes from the paper itself, and it deserves to be stated plainly rather than buried.
The authors are explicit that AI adopters are not a random sample. They are already larger, more engineering-intensive, more likely to be venture-backed, and faster-growing than non-adopters. Sector-level gains concentrate in Information (Ramp Economics Lab, 2026).
That is a real problem for the naive causal reading. Fast-growing, well-capitalised software companies both spend heavily on AI and hire heavily, and the common driver may simply be that they are winning. Under that interpretation, heavy AI spend is a symptom of expansion rather than a cause of it โ and a 220-person industrial services firm buying more licenses would not inherit the 12%.
Two things keep the finding useful even after that discount. First, the study tracks changes within firms before and after adoption, not a raw comparison between adopters and non-adopters โ the growth arrives after the spend, on a gradual curve. Second, and more decisive for your purposes: the null result in the low-intensity tier is not explained away by selection. Those firms also chose to adopt. They also had budget, sponsorship, and intent. They spent less per employee, and got a statistically indistinguishable zero.
So treat the 12% as directional and context-bound. Treat the zero as the transferable lesson.
What Plausibly Separates a 12% From a Zero
Here the evidence gets thinner, and I want to be clear about where the data ends.
The Ramp paper measures spending intensity and employment outcomes. It does not identify the mechanism โ it cannot tell you what high-intensity firms did with the money that low-intensity firms failed to do. Anyone claiming otherwise is reading in.
The best available evidence on mechanism comes from elsewhere. McKinsey's State of AI 2025 found that while a large majority of organisations report using AI, only a minority report meaningful EBIT impact at the enterprise level โ and that the practice most associated with the organisations seeing returns is fundamental redesign of workflows around the technology, rather than deployment on top of existing process (McKinsey, 2025). That is a correlation from a self-reported survey, not a causal estimate, and it should be weighted accordingly.
Put the two together with appropriate caution and a plausible story emerges: sustained high spend per employee is what workflow redesign looks like in the ledger. Redesigning how work moves through a function costs integration effort, data work, and process time โ it shows up as a large, concentrated bill. Buying seats and leaving the workflow untouched is cheap, which is exactly why it produces a low-intensity signature and a null result.
I would hold that as a working hypothesis, not a finding. What is measured is this: concentration of spend predicts workforce growth, and dispersion predicts nothing.
Three Moves Before the Next Budget Cycle
Measure your intensity, not your adoption
Most mid-market AI reporting counts licenses, active users, and departments touched. None of those is the variable that moved in this study. Compute AI spend per employee, and compute it per function rather than across the company โ a firm-wide average conceals whether you are running one serious program or nine cosmetic ones. You will usually find the total looks respectable while every individual function sits in the low-intensity tier.
Concentrate rather than distribute
If dispersion returns nothing, the implication for allocation is uncomfortable but clear: the same budget concentrated in one or two functions is a materially different bet from the same budget spread across nine. Pick the functions where the surrounding process can actually be rebuilt โ where you own the workflow end to end and can change how work is routed, not just how it is drafted. Fund those properly and starve the rest deliberately.
This will be politically harder than the spreadsheet suggests. Every function head has an AI line item now, and withdrawing one reads as a judgement about that function. Make the argument on evidence: a half-funded pilot is not a smaller win, it is a zero with a budget line attached.
Tie the AI case to a hiring case, explicitly
The reflexive framing โ AI investment as headcount avoidance โ is not what the highest-intensity firms in this dataset produced. They grew, including at entry level (Ramp Economics Lab, 2026). Whether that generalises to your sector is genuinely open, given the Information-sector concentration. But it is worth naming the assumption in your own business case, because "this replaces three hires" and "this lets us absorb more work with the team we are building" produce different implementation choices, different training investment, and different internal politics from day one.
One Decision
Pull your AI spend for the last four quarters and divide it by function and headcount. Then ask one question: is any single function in this company actually in the high-intensity tier, or have we distributed a serious budget into nine unserious programs?
If it is the second โ and for most mid-market operators it will be โ the fix is not more money. It is fewer bets.
The firms in this dataset that grew headcount, entry-level hiring included, were not the ones that adopted AI earliest or most enthusiastically. They were the ones that concentrated. The firms that got nothing were not the ones that stayed out โ they were the ones that spread it thin.
