Cheaper, then worse: what the AI rehiring wave reveals
Robert Half, Forrester, Gartner and MIT data all point the same way: companies are rehiring after AI-driven cuts. The cause isn't that AI fails — it's automating the documented work while shedding the judgment that was never written down. Here's how to deploy AI without walking it back.
Written for CXOs & boards
Something in the data on AI and jobs complicates the headline story: companies that cut roles to deploy AI are quietly hiring people back. Robert Half finds 32% of managers who eliminated positions after adopting AI are now rehiring for the same or similar work. Forrester expects the reversal to be widespread — it predicts half of all AI-attributed layoffs will be quietly reversed in 2026, with the work often returning offshore or at lower pay. And Gartner expects it to reach the very function AI was meant to own outright: half of the companies that cut customer-service staff for AI will be rehiring by 2027, often under a new title.
The most public version is Klarna. In early 2024 the fintech said an AI assistant was doing the work of 700 service agents and fielding two-thirds of its chats. A year on, its chief executive conceded the shortfall — the system handled the volume but not the hard parts, quality came out "lower," and the company was bringing humans back for the interactions AI couldn't carry.
The most expensive conclusion
The easy reading is that AI was oversold and the cuts were a mistake. It's the most expensive conclusion a leader can draw. Decide the technology simply doesn't work and you either swear off it while competitors who didn't pull ahead, or run the same play again in two years under a new name.
It also doesn't fit the evidence. MIT's 2025 study of enterprise AI — the one that found 95% of corporate gen-AI pilots deliver no measurable return (the same gap between AI spend and AI return we examined in Hong Kong) — was explicit that the shortfall "does not seem to be driven by model quality." The failure is in the approach, not the technology.
What the rehiring wave actually shows
Look at why companies say they're rehiring and the picture sharpens. In Robert Half's data the leading reasons aren't "the model broke." They're that the work needed more oversight and quality control than expected (38%) and relationship skills the technology couldn't replicate (37%). Klarna is the same story in miniature: the bots absorbed the routine queries and stumbled on the complex, the emotional, the multi-step.
That points to a precise mistake — one that has nothing to do with whether AI works. These companies automated the part of the job that was written down — the documented, repeatable, easy-to-measure majority — and in doing so quietly deleted the part that wasn't: the exception nobody logged because it never looked important, the judgment that told an experienced hand this one is different, the relationship that kept an account from leaving. None of it appeared on a process map, which is exactly why it was invisible when the cuts were modelled, and exactly why it went first.
The trap is in the arithmetic. Automate the visible work and sever the unwritten work, and you don't keep most of the value — you keep the cheap, easy majority and lose the part that made the service trustworthy. The saving lands now and shows up in next quarter's numbers; the damage lands later, when customers meet the gaps and quietly leave. By the time the quality line bends, the people who held the unwritten work are gone, and rehiring is the only repair left. Cheaper now, worse later.
Illustrative — the shape of the pattern, not plotted data. The saving registers at once; the quality cost surfaces months later, once automation meets the work that was never written down. ASTRA’s framing, consistent with the Robert Half and Forrester findings cited above.
Beneath it sits a simpler error. Most "AI transformations" redesign nothing. They wrap a model around the process the company already had — same approvals, same handoffs, same decade-old assumptions, only faster. Automating a flawed process doesn't repair it; it just runs the flaw at speed. The productivity gain is real and the strategic gain is near zero, because the work being optimised should never have existed in that shape.
The decision nobody owns
Why does this keep happening to well-run companies? Because they staffed the build and left the decision unowned. Every serious firm now hires for delivery — AI architects, forward-deployed engineers, solutions engineers — and files them, almost without exception, under engineering or IT. Read those job descriptions and they describe one verb: build. Ship the prototype, wire it into the stack, stand up the monitoring. Every line assumes the thing should be built. But whether to build this, in this shape, and whether the full cost is worth it, are settled before anyone writes code — and that is the decision no one is given to own.
Two questions belong upstream of the build, and almost nobody holds them. The first is about redesign: forget the inherited process — what are we actually trying to achieve, and what's the leanest shape that gets there? The second is about honest cost: not can we build it, but should we — counting the lines that never make the business case, like the oversight the system will demand, the integration and upkeep behind it, and the quality control you quietly hand to your own customers when the bot can't cope.
Inside owner, outside partner
That decision can't be outsourced wholesale either. The unwritten work lives inside your people, not in a consultant's deck, so the owner has to sit inside the business, close to the real workflow, with authority to change it. But they shouldn't decide blind — and the data agrees: MIT found AI efforts built with a specialised outside partner reached production about 67% of the time, against a third as often for purely internal builds. A good partner supplies what an insider can't generate alone — the pattern from running the same change across many companies, a view across the whole organisation, and benchmarks to decide against reference points instead of in the dark. The insider owns the decision and lives with the result; the partner brings the evidence to make it well. Two roles, not one technical hire asked to be both.
What this means in practice
For a leader standing up or repairing an AI function, the corrections are concrete:
- Put the build behind the decision. Answer the redesign and true-cost questions before approving anything. If the only person in the room is the one who builds it, neither gets asked.
- Redesign the work before you automate it — decide what deserves to exist, then automate that, not the org chart you inherited.
- Find and capture the unwritten work before cutting the people who hold it. Oversight, judgment, the handled exception, the kept relationship don't transfer through a job description.
- Cost the whole iceberg, not the headcount line — including the quality control you'd otherwise push onto customers.
The real lesson
The research is consistent on the punchline: this was never a referendum on whether AI works. It's a verdict on how it was deployed. The companies that pull ahead over the next two years won't be the ones that automated first — they'll be the ones that understood what they were automating before they touched it.
That's the discipline ASTRA is built around — and why our forward-deployed engineers sit inside the business problem rather than beside the server, why we run what we build instead of handing over a deck, and why we price against a result rather than the hours or tokens behind it. You own the decision; we bring the evidence to make it well.