When exposure became insulation
On Wednesday, Axios reset the week’s jobs debate with something sturdier than vibes: economy-wide numbers. Instead of the oft-recycled script that AI is immediately hollowing out payrolls, fresh analysis pointed the other way. In roles most exposed to AI, paychecks got fatter and headcounts ticked up. The twist isn’t just that the sky didn’t fall—it’s that the rain seems to be cash.
The surprise tucked inside the data
Vanguard stitched together O*NET task descriptions with BLS and Census series, carving out roughly 140 occupations where current AI can handle a meaningful slice of tasks under moderate supervision—while discounting the parts of work that still hinge on in-person presence, managing people, or clinical judgment. Measured from Q2 2023 to Q2 2025, real wages in that “high exposure” group rose 3.8%, and employment grew 1.7%. In the less-exposed crowd, real wages crept up just 0.7% and employment 0.8% over the same window. If AI had already turned into an indiscriminate job shredder, the pain would show up here first. It doesn’t. Instead, the exposed group is outperforming. Axios framed the finding plainly and paired it with the mood from corner offices, which—perhaps surprisingly—now tilts toward hiring rather than trimming.
Methodologically, this matters. Vanguard’s approach stops short of gauging how many “jobs” AI could do in the abstract; it asks how much of what people are already paid to do is realistically within reach of current systems. It then watches what actually happened to those people’s wages and employment. That distinction—tasks versus jobs—turns out to be the hinge of the entire conversation.
How substitution becomes demand
There’s a reason exposure isn’t translating into cuts. When you reduce the unit cost of a task, you don’t just save; you usually do more of that task. That’s especially true when the demand side is elastic—more documentation because compliance is cheaper, more experiments because analysis cycles are faster, more product variations because content creation is trivial. Software and knowledge work are full of these multipliers. Make code, design, analysis, or outreach cheaper and you don’t simply bank the margin—you widen the scope of what’s worth doing.
Box CEO Aaron Levie put the point bluntly at a recent Axios event:
“The people who predicted mass job destruction will be proven wrong.”
His argument was less cheerleading than microeconomics: if software gets cheaper and faster to create, it shows up in more places, which increases the surface area for human work. That doesn’t mean every task survives; it means the bundle of tasks that constitute a job gets rebuilt, often around oversight, integration, and decision-making. The Vanguard figures suggest that, in aggregate, that rebundling increased both what workers in exposed roles are paid and how many of them are employed—at least so far.
The boardroom tells a similar story
If the numbers capture the recent past, the C‑suite is telegraphing the near future. A new Teneo survey, cited by Axios, finds most large-company CEOs expect AI to drive incremental hiring in 2026, with two-thirds anticipating an increase in entry-level headcount. That detail is the giveaway. When senior people become force multipliers, the constraint shifts to throughput—coordinating more projects, launching more pilots, covering longer tails of customer needs—so leaders staff at the base of the pyramid to capture the upside. It’s the inverse of a hiring freeze; it’s a throughput expansion plan.
Crucially, this isn’t the euphoric phase of an innovation cycle where everyone hires indiscriminately. Much of today’s AI spend is still plumbing—data centers, tooling, data engineering. Yet the intention to grow junior roles implies executives see near-term returns from pairing low-cost task automation with human judgment and institutional context. In other words, we are not waiting for artificial generality to transform work; the transformation comes from making ordinary tasks cheap and orchestrating them at scale.
What the numbers don’t say—yet
None of this is a guarantee about the next five years. The period Vanguard analyzed—Q2 2023 to Q2 2025—captures the first full sweep of the modern generative wave, not the endgame. Macro forces unrelated to AI still matter. Some tech and professional services sectors have cooled hiring for reasons ranging from rates to geopolitical risk. Composition effects can also mislead: high-exposure roles often sit inside firms that are already winning, which can lift wages independent of technology. And measurement is inherently laggy; once organizations complete their infrastructure buildout and process reengineering, second-order automation could compress headcount in specific niches.
But the burden of proof has flipped. You can no longer claim that AI’s first-order impact is obvious, negative, and already visible in the most exposed jobs. The opposite shows up in the data. If a broad labor shock were underway, it would be hard to hide in economy-wide series. Instead, we see a pattern that looks like short-run complementarity: productivity goes up, wages follow, and employment nudges higher in the roles closest to the tools.
What this means for people who actually do the work
For operators, the signal is to optimize for orchestration and rate of learning. The valuable edge is not writing the line of code the model can write; it’s deciding which ten lines matter and how to route their output into systems, teams, and customers. That’s why entry-level hiring reappears on CEO roadmaps: apprenticeships need throughput. In practice, careers in exposed roles will be built on three habits—treating models as teammates, instrumenting workflows to make decisions auditable, and packaging outcomes in ways that align with real constraints like risk, privacy, and regulation.
For companies, the wage premium in exposed roles reads as a price for capability, not a tax on efficiency. Pay goes up where output scales with the tool. Leaders who imagine savings only as headcount reduction will leave money on the table; those who combine automation with redistribution of human attention—toward supervision, integration, and customer contact—will capture the demand expansion that the numbers imply.
The baseline for 2026
Take Axios’s piece as a baseline, not a victory lap. The present tense is clear: between mid‑2023 and mid‑2025, the workers closest to AI saw faster real wage gains and slightly stronger job growth than peers, and most big-company CEOs are planning to hire into that momentum next year. If the cycle turns, we’ll see it. For now, the economy’s most exposed seats aren’t emptying. They’re being rewired—and, tellingly, refilled.
Sources: Axios; Vanguard’s December 2025 labor analysis (PDF); Teneo’s CEO/investor survey (press release).
