The Most-Hired Casualty
In May, tech-sector layoffs hit their highest monthly level in years — and the reason companies gave, over and over, was AI. So here’s the number that should stop you cold: in that same stretch, the very job the headlines had marked for death wasn’t shrinking. It was the most-hired role in the whole industry.
Start with SignalFire. The venture firm tracks hiring across the major tech companies, and its latest data found something that cuts against the entire narrative: software engineers — the job everyone named first when they talked about AI wiping out work — made up about fifty-five percent of all new hires at major tech firms in the most recent hiring data. Not a shrinking slice. More than half. The single most-predicted casualty of this technology is, by the actual hiring numbers, the most resilient role on the board.
And it isn’t just engineers. Take entry-level cybersecurity — another category people had already written the obituary for, on the theory that AI would automate the junior analyst out of existence. Dark Reading, working from ISC2’s cybersecurity workforce survey, says the opposite is happening. Those jobs aren’t disappearing; they’re being reshaped — the routine alert-triage work shifts to automation, and the demand moves toward the judgment and the strategic, harder-to-automate skills. The data backs the reshaping: AI skills are now the field’s most-cited skill need, and a meaningful share of teams expect automation to create new entry-level roles even as it absorbs the old tasks. The entry rung isn’t vanishing. It’s being redrawn. Same pattern, different profession.
Now widen the lens. Axios Macro reported that the after-the-fact revisions to the official U.S. jobs numbers may be about to start adding jobs back rather than subtracting them. No single number proves anything about AI on its own — plenty of forces move the macro data — but a labor market quietly revising upward is not the picture you’d expect if this technology were the wholesale job-killer it was sold as.
And here’s the tell from inside the building. OpenAI — the company at the center of all of it — says ninety-seven point nine percent of its own employees now use AI agents in their work, up from around forty percent not long before. Near-universal, in under a year. The people closest to this technology didn’t get replaced by it. They picked it up.
So that’s the picture, before we’ve said a word about why. The jobs marked for deletion are holding. And the most AI-saturated workforce on earth is still a workforce of people — there, at their desks, using the tools.
Which Half It Eats
The reason the doomed jobs held — and the reason this matters far beyond a hiring chart — comes down to one distinction the predictions never made. AI doesn’t replace work. It replaces a kind of work: the undifferentiated, repeatable execution that runs the same way every time. And it leaves almost untouched the thing sitting on top of that execution — the judgment about what to do, when, and whether the output is right.
Watch where that line falls, because the IT services world is drawing it in real time. The reporting on why traditional IT outsourcing firms are struggling in the AI era lands on a single phrase: agentic outsourcing. The model is shifting so that the AI performs the operational work — the ticket, the migration step, the config change — while the human role narrows to oversight and decision. Read what that actually does to a business. If your firm’s product was a pile of billable hours doing predictable tasks, the machine just learned to do the pile. If your product was the judgment about which tasks mattered and whether they were done right, the machine made you more valuable, because now you can render more of that judgment without drowning in the execution underneath it. Same technology. Opposite outcome. The only variable is which side of that line your business was selling. And notice the second-order effect, because it’s the part the layoff headlines miss: making judgment cheaper to apply doesn’t shrink the demand for it — it expands it. That’s the real reason engineers didn’t just hold their ground but became a bigger share of hiring, not a smaller one. The same logic runs straight to the MSP. Sell judgment amplified by the machine, and you’re not defending a shrinking lane — you’re widening the one you’re in.
And the reason this is landing now, hard, is that the execution layer turns out to be enormous. Splashtop went and measured where IT teams actually spend their day, and the number is the whole story: more than half — fifty-three percent — of IT capacity goes to endpoint maintenance and reactive, repetitive tasks. Sit with that. The majority of the work isn’t the judgment. It’s the grind. It’s exactly the undifferentiated execution AI is built to absorb. So when the technology arrives, it doesn’t nibble at the edge of the job — it walks straight into the half of the work that was already the most automatable, and it does it at scale.
Which means the question was never whether AI would take work. It was always which half of the work you built your business on — and whether what’s left is something a client will still pay a premium to have a human own.
The Rent-a-Team Trap
So aim that distinction at the MSP, because it forces a decision most owners are making by default instead of on purpose — and the market is making the stakes obvious right now.
Start with the clients. The NFIB’s latest small-business jobs report put hiring plans at a six-year low: a net nine percent of owners plan to add jobs in the coming months, and a record share — fourteen percent — name labor cost as their single biggest problem. Read what that’s actually telling you. Your clients have decided they are not going to solve their next problem by hiring a person. They can’t afford the headcount and they don’t want the risk. So what they’re in the market for is no longer a body to do the work — it’s the outcome the work was supposed to produce, delivered without another seat on their payroll. The client who used to ask you to help them staff up is now asking you to make the staffing question go away. That is a demand for judgment-as-a-service, whether or not anyone calls it that.
Now the trap, because there’s a tempting way to answer that demand that quietly puts you on the wrong side of everything we’ve described. There’s a whole pitch aimed at MSPs right now — managed support teams, pre-trained and certified, that you can stand up in thirty days, sold as the answer to the technician shortage. And on a spreadsheet it looks great: rent the bodies, beat the staffing problem, scale. But look at what you’ve actually become if that’s your offer. You’ve made your product a pile of cheaper hours doing predictable work — which is the exact thing the machine learned to do, and the exact tier of provider the agentic shift is hollowing out. You didn’t escape the labor problem. You repackaged it and put your name on it.
So the fork is this. You can build your firm around the layer AI makes scarcer and more valuable — fewer, more senior people whose judgment you sell as the outcome, amplified by the machine doing the grind underneath them. Or you can keep selling bodies-by-the-hour, dressed up as managed services, and compete on price in the one lane the technology is actively draining.
Why Do We Care?
The fair objection is that labor arbitrage is old news — MSPs have rented cheaper hours, onshore and off, for decades and made good money doing it, and a new tool doesn’t repeal that. But every prior version of that trade was your cheaper bodies against a competitor’s more expensive bodies — the rival was always another human, so there was a floor under the price. Agentic execution removes the floor: the predictable work now competes against a machine doing it at almost no marginal cost, which means the body-shop isn’t being undercut by a cheaper labor pool this time — it’s being replaced by no labor pool at all. And that’s the one race you can’t win by standing up the next managed team thirty days faster.
What to Consider
Sort your own service lines by which side of the execution line they sell. Take every recurring service you offer and mark each one as either predictable execution — the work that runs the same way every time — or judgment a client pays a human to own. The execution column is your exposure: it’s where the agentic shift competes against you at near-zero marginal cost, and where pricing on hours stops being defensible. You can’t reposition what you haven’t separated, so this sort is the work that comes before any pricing or staffing decision.
Pressure-test any “rent-a-team” expansion against the floor that just disappeared. Before you sign onto a managed-support-team or staff-augmentation model to solve a capacity problem, ask the question the old arbitrage math never had to: who are you actually competing against on that work in two years — a more expensive human, or a machine doing it for almost nothing? If the honest answer is the machine, you’re not buying a growth lever, you’re buying into the exact tier that gets hollowed out — and the thirty-day stand-up speed the pitch sells you is irrelevant to a race against zero marginal cost.
Reposition your bench toward the judgment the machine makes scarcer, not the execution it absorbs. Look at where your people spend their time and move your senior judgment to the front of what you sell, with the machine carrying the repeatable volume underneath them. That means hiring and promoting for the decisions — what to do, when, and whether the output is right — and stopping the reflex to add headcount for the predictable grind. The firm that makes that shift on purpose keeps the half of the work clients still pay a premium for; the firm that doesn’t is left selling the half that’s draining.
If this trend continues, within the next twelve to eighteen months the MSPs still selling predictable execution by the hour will be losing renewals not to a cheaper competitor but to no competitor — to clients who moved that work to an agent — while the firms that repositioned around judgment will be quoting outcomes the machine can’t own, in the one lane the technology keeps widening for them.