32% of managers cut a role for AI, then rehired for the same one

Sebastian Siemiatkowski went on record in May 2025 saying his AI chatbot "went too far." He'd replaced 700 customer service agents the year before, claimed $10 million in savings, and spent months as the poster child for AI-driven workforce reduction. Then quality dropped. Klarna started rehiring.

That reversal would be an anecdote. What makes it the anchor of a broader AI hiring boomerang is that five other organizations published hard numbers telling the exact same story in the same quarter.

The six sources

Robert Half surveyed 2,000 US hiring managers. 32% reported eliminating a role for AI, then rehiring for the exact same position. Finance departments led at 44%. Tech was at 32%. Their Canadian parallel survey of 1,365 managers returned 34% with the same result, and Deborah Bottineau (managing director, Robert Half Canada) named the top reason: "AI required more human oversight than expected." 38% of managers cited it. Not cost. Not regulation. The AI needed a babysitter.

Gartner's February prediction: 50% of companies that replaced staff with AI will rehire those roles by 2027. Their survey of 321 customer service and support leaders found only 20% actually reduced headcount after deploying AI. That's an 80% failure rate on the central promise.

Forrester put the cost to it. 55% of companies that made AI-driven layoffs regret the decision. 35.6% rehired more than half of the affected workers. And the number that should be in every CFO's inbox: companies paid 1.27 times the cost of their original layoffs in lost productivity and knowledge gaps. J.P. Gownder at Forrester called half of all AI-attributed layoffs destined to be "quietly reversed."

Block laid off over 4,000 people. Jack Dorsey framed it as AI-driven efficiency. Headcount went from 10,000 to roughly 6,000. At least four employees publicly confirmed they were rehired within weeks. Block's official explanation for the reversals: "clerical errors." Andrew Harvard, a design engineer caught in the cuts, documented his rehire publicly on LinkedIn. The euphemism is louder than the layoff.

Google's AI engineering hiring data for 2025 shows 20% of new AI hires were boomerang employees returning after the 2023 layoffs. The highest-profile return: Noam Shazeer left Google, founded Character.AI, and came back via a $2.7 billion licensing deal. John Casey, Google's head of compensation, confirmed the boomerang trend internally. The most expensive AI talent Google shed in 2023 was the same talent they paid a premium to reacquire two years later.

Careerminds added a sixth data point: 68.3% of companies that restructured for AI rehired some of the affected employees. 41% said they'd pursue alternatives to layoffs next time.

The gap nobody's tools could close

Six sources with no coordination and no shared incentive. Every one of them found the same structural problem.

Emily Potosky at Gartner put it directly: "AI simply isn't mature enough to fully replace the expertise, empathy, and judgment" companies eliminated. Julie Geller at Info-Tech Research Group used Klarna as the case study for what she called the trust degradation pattern. AI handles volume. It misses context. Customer satisfaction erodes. The rehire cost exceeds the savings.

The tools didn't fail completely. Klarna's chatbot handled millions of conversations. But companies treated AI systems as drop-in replacements for roles that required institutional knowledge, relationship context, and judgment built over years of doing the actual work. Those capabilities don't arrive in a model update. They accumulate. They're personal. And once you fire the person who held them, rebuilding that context from scratch costs more than keeping them on payroll ever did.

What Bottineau's "more human oversight than expected" finding reveals is the real architectural error. Companies assumed AI would eliminate the human in the loop. Instead it created a new loop where someone needs to understand both the domain and the AI system well enough to supervise the output. That's not fewer people. That's a different, more specialized role. Harvard Business Review found that most of these layoffs were based on anticipated future capabilities, not measured current performance. Companies cut first and evaluated second. The boomerang is what evaluation looks like when it finally arrives.

Where I called this

I wrote about the productivity panic coming for senior engineers in May, arguing that senior engineers are the load-bearing walls companies were ripping out. The boomerang data validates that thesis from the employment side. I run 30+ AI agents splitting planning from execution daily. They handle work I've already scoped and structured. They don't do the scoping. Every practitioner I know who ships with AI agents draws the same line.

The cost nobody is adding up

Forrester's 1.27x multiplier is the headline, but it understates the full damage. Megan Slabinski at Robert Half called it the position-recreation problem: the rehire isn't "getting the person back." It's getting them back at a higher salary band, with institutional context that takes months to rebuild, with team trust damaged by the layoff, and with onboarding costs stacked on top. One-third of companies surveyed spent more on restaffing than they saved from the original cuts.

Gartner's Kathy Ross pointed out that rehires often come "under different titles." Same work. New title. Higher salary. The job descriptions now include "AI oversight" or "human-in-the-loop QA" as core responsibilities. Companies didn't just rehire for the role they cut. They rehired for a more expensive version, because the AI system became a dependency that needs a dedicated operator. The operator role costs more than the original because it demands both domain expertise and fluency with AI systems that didn't exist when the original role was scoped.

Goldman Sachs estimates AI is eliminating between 5,000 and 10,000 US jobs per month nationally. That's the macro number. Block alone cut 4,000 in a single restructuring. If the Goldman number is right, Block's single layoff round represented weeks' worth of nationwide AI displacement concentrated in one company. The scale alone should have flagged the decision as something other than normal AI adoption. It was financial engineering dressed in AI language.

On the tool-cost side: Uber burned through its entire 2026 AI coding budget in four months and had to cap per-employee spending. The economics of AI tools shifted in the same quarter that the headcount thesis collapsed. Companies that banked on AI reducing both tool costs and labor costs are watching both bets come apart simultaneously. The playbook was: cut headcount, redeploy savings to AI tools, pocket the net. When both line items reverse, the net is negative.

Where replacement holds, where it doesn't

Block's case is the messiest and deserves its own caveat. Zachary Gunn at FTP argues the layoffs were pandemic-era bloat correction wrapped in AI narrative for investor positioning. Calling rehires "clerical errors" suggests a financial restructuring that borrowed AI language for the earnings call. Not every company claiming AI-driven layoffs actually displaced people with AI.

And there are domains where replacement genuinely works. Klarna saved $10 million before quality degraded. The pattern: high-volume, well-defined tasks with clear success metrics and low relationship dependency. Simple customer inquiries. Standardized document processing. Structured code generation where humans review every output before it ships.

Where it consistently breaks: anything where the knowledge lives in the person, not the process. Complex customer relationships. Engineering roles that require understanding why the rules exist, not just following them. Work where judgment compounds over years and can't be extracted into a training set. The companies that cut those roles fastest paid 1.27x to learn the lesson.

Forrester found 57% of executives now expect AI to increase headcount, not decrease it. That isn't a reversal of conviction. It's the measurement finally catching up. The tools are real. The capabilities are growing. But the workforce math that assumed you could subtract humans and add inference without a quality collapse was wrong. Six organizations with no reason to coordinate just published the proof.

What separates the two groups

The companies rehiring at 1.27x had one thing in common: they treated a tool like a replacement. The companies that aren't in any boomerang dataset treated it like infrastructure that changes how work gets done, not who does it. Same technology. Different assumption. Wildly different cost.