Thirty dollars per employee per month. That's the spend threshold that separates the firms where AI grows headcount from the firms where it does nothing at all.
Not thirty thousand. Not three hundred. Thirty dollars — roughly a single seat of a coding agent, per head, in the first three months. Ramp's high-intensity adopters reportedly averaged $33.67 by one telling. Below that line, statistically, AI's effect on your workforce is zero.
That number is the whole story, and I'll spend the rest of this piece explaining why.
Cross the line, and the technology compounds into hiring, including entry-level hiring. Stay below it, and you get nothing. It's a threshold effect, not a labor switch.
Last week, this newsletter ran the automation-reckoning framing: AI as a labor switch, junior roles as the first casualties, the entry-level ladder pulled up behind Gen Z. That thesis leaned on real numbers, and I'm not walking them back. This week, a dataset arrived that forces a refinement — not a retraction.
Start with the contradiction, stated plainly, because both halves are true.
Through May 2026, companies tied close to 90,000 announced job cuts to AI. Goldman Sachs estimates AI erased about 16,000 net jobs per month over the past year, with Gen Z and entry-level workers taking the brunt.
Now the other half. Ramp and Revelio Labs, tracking actual AI vendor spend against workforce records for more than 21,000 U.S. firms, found that heavy adopters grew headcount 10.2% over the two years following adoption. Entry-level roles grew 12% — faster than the overall workforce.
Same technology. Same window — the study runs right through the period Goldman covers. Opposite signs.
The lazy move is to declare a winner. "AI isn't killing jobs," as Ramp's own economist titled his post. The honest move is to read the caveat, because the caveat is where the mechanism hides.
Those 10.2% gains are not spread across adopters. They are, in Ramp's words, "entirely driven by high-intensity adopters." Low-intensity adopters — the pilot-runners, the subscription-buyers — saw no statistically significant change in headcount. None.
So AI is not a labor switch that flips one way for everyone. It's a threshold effect. Cross the line, and the technology compounds into hiring, including entry-level hiring. Stay below it, and you get nothing.
This is not substitution versus creation. It's a capability tax — and it only pays out to firms with the capital, the engineers, and the networks to operationalize what they bought.
Look at who the adopters are before they ever adopt. Ramp is blunt about it: adopters are larger, faster-growing, more engineering-intensive, and more than three times as likely to be venture-backed. Median year-over-year headcount growth among adopters was 6.0%, against 1.6% for firms that never adopt.
These aren't average companies discovering a productivity tool. They may be already-winning firms adding a multiplier, though the data makes it difficult to say whether AI drives the hiring or just accompanies firms already expanding.
The most revealing detail in Ramp's write-up isn't even about spend. It's about networks. Who funded you predicts AI adoption better than what sector you're in. California tech firms adopt more than otherwise-identical New York firms. There is, the authors admit, "no inherent reason for that, other than the fact AI spreads through networks."
That's the capability tax made visible. The scarce input isn't the software — thirty dollars a head is not a moat. The scarce input is knowing how to install it, whom to hire, and which best practices to copy. And those things travel through founder networks, not open markets.
Which brings us to the part of the market that the substitution thesis simply cannot explain: the human-services boom sitting directly on top of the labor-replacing product.
On Tuesday, AWS launched a $1 billion internal organization for AI-focused forward-deployed engineers — humans who embed inside client companies to deploy purpose-built agents. It follows OpenAI's FDE venture, valued at $4 billion, and Anthropic's, at $1.5 billion. Per TechCrunch, both AI labs paired with private equity firms, which provided both the capital to launch and connections to portfolio companies that could become clients.
Read that again. The private equity partners were brought in for capital and for their connections to portfolio companies. The vendors are monetizing exactly the gap Ramp identified.
The FDE model, pioneered by Palantir, works like this: an engineer from the vendor works temporarily inside the client while the system gets stood up. AWS's Francessca Vasquez framed the deliverable as more than software — customers "gain lasting AI skills, workflows, and patterns they can use to innovate independently."
That is a description of selling operational capability. The product isn't the model. The product is the ability to use the model.
Now hold the two facts against each other. If AI were straightforwardly labor-replacing, why does the flagship AI product ship with a large human-services attachment? Why are the three biggest names in the field, per TechCrunch's reporting, tied to corps of embedded human engineers?
Because the substitution story was always missing a step. AI doesn't deploy itself. Someone has to redesign the workflow, and the firms that can't do that in-house will now rent the humans who can.
The corroborating evidence is everywhere once you look. Google is hiring hundreds of forward-deployed engineers to move customers from pilot to production. Box's Aaron Levie says AI created 13 new job categories at his company, including model evaluators — a role that exists because choosing between models "carries real operational weight."
IBM is the sharpest case. It's tripling U.S. entry-level hiring in 2026 — while admitting that work its junior developers did two or three years ago can now largely be done by AI. So it rewrote the job descriptions. Junior developers now spend less time on routine coding and more time working directly with customers.
That may be what the 12% entry-level growth looks like on the ground, though the IBM example and the Ramp/Revelio figure come from unrelated datasets. Not the old entry-level job, preserved. A new one, redefined around AI fluency.
Ramp's economist reached the same read: high-intensity firms are selecting for a genuinely new skill — people who know how to use AI well — and recent graduates are a natural place to look. The ladder isn't being pulled up. It's being rebuilt with different rungs, at the firms that can afford the reconstruction.
Which is why the reckoning isn't cancelled. It's relocated.
PwC's barometer, drawn from over a billion job postings, found AI-exposed junior roles are seven times more likely than less-exposed roles to require traditionally senior skills like leadership and judgment. Postings for those "seniorized" entry-level roles grew 35% since 2019; postings for ordinary entry-level roles shrank 10%. And 74% of AI's economic value, PwC found, is being captured by just 20% of organizations.
That is the same divide Ramp measured, from the opposite direction. The winners hire more, including at the bottom. The losers stagnate — and their workers may be among the 16,000 net jobs Goldman counts leaking out each month, though no evidence directly links the two datasets.
So here is the synthesis, and it's less comfortable than either headline.
Both datasets may be true because AI's real output could, arguably, be a widening capability divide, though no direct evidence establishes a single causal explanation. The gains are concentrated, threshold-gated, and network-distributed. The losses fall on everyone below the line — firms stuck in pilots, and the workers those firms don't hire or quietly cut.
Ramp knows this. Its own authors warn that firms without the channels "may fall behind," and the company promptly announced a partnership with Meta Small Business to close the gap. When the researcher and the vendor both start selling the solution, believe the divide is real.
The FDE armies may, arguably, be capital's clearest bet on where the scarcity sits, though that reading is speculative. Not in the model. In the humans who can install one.
That's the correction to last week's framing. AI isn't a switch that kills jobs. It's a tax that only the already-capable can pay to their own advantage — and an industry now exists to rent that capability to everyone else, one embedded engineer at a time.