The AI boomerang: cutting first and measuring later is costing twice

Here's a simple test for any "AI efficiency" announcement: did the company measure what the people it cut actually did before it cut them? Judging by the growing wave of rehiring, many didn't.

The pattern is now hard to ignore. Outlets such as RT's Actualidad are running with the "AI boomerang" label: companies bringing back employees they had replaced with AI. Klarna is the best-known cautionary tale. As Mundiario recounts, the fintech bet heavily on automating customer service and later resumed hiring after acknowledging the limits of that approach. The same piece cites a Gartner forecast that by 2027 half of the companies that cut staff and blamed AI will rehire for similar roles.

Treat that last number as what it is: a forecast, not a result. Some of the surveys floating around, like the claim that up to 55% of employers regret AI-driven layoffs, come from undated pages we can't verify well. The direction, though, is consistent across sources.

The mistake wasn't using AI, it was subtracting people first

We don't think this story shows that AI doesn't work. It shows a sequencing error. A sensible rollout goes: map the work, automate the parts that hold up, measure the quality, and only then decide how many people you need. Plenty of companies seem to have run it backwards. They booked the savings in the press release and left the evidence for later.

The tell is what tends to disappear along with the headcount. Mundiario's analysis points to the loss of accumulated knowledge: experienced employees know how customers behave, spot problems before they escalate and handle the odd cases. A model can answer hundreds of queries in seconds, but it can also invent information or repeat a bad decision at scale. And when the people who would catch those errors have been let go, an error stops being a ticket and becomes a financial loss, a complaint or a reputational hit.

That is the double payment in the thesis. First you pay for the cut and the disruption it causes. Then, in our view, you pay again for recruiting, retraining and the customer goodwill you lost along the way.

The business case was shakier than the boardroom assumed

There is a reason the cut-first logic was seductive. According to a second Mundiario analysis, 77% of boards expect AI to reduce costs, while 38% of companies have already gone over their implementation budget. Integration, upkeep, infrastructure and human oversight can absorb a good part of the promised savings. If you book the payroll cut up front but the AI bill arrives later and bigger, the math flips.

Our reading: layoffs were the only line of the AI business case that was immediate and certain, so they became the case. Everything else, quality and error rates among them, was assumed.

Be skeptical in both directions

Two cautions. First, not every layoff attributed to AI is really about AI. Mundiario itself notes that costs, restructuring and economic conditions also play a part, and that the expectation of what AI will do tomorrow is often used to justify cutting jobs that still matter today. "AI-washing" a routine cost cut is convenient, and the boomerang may partly be that story unwinding.

Second, rehiring is not a victory lap for workers. A Stanford study cited in the same article found employment for 22-to-25-year-olds in highly AI-exposed occupations 19% below the trajectory it would otherwise have followed, though the researchers found no evidence of economy-wide job destruction. So the damage looks concentrated at the entry level, exactly where people build the judgment that companies now say they miss. Rehiring veterans while closing the on-ramp for juniors only postpones the same problem.

Our verdict

If you run a company, treat headcount as the last variable you change, not the first. Pilot, measure errors and customer outcomes, keep the people who know where the model fails, and cut only what the data has already proven redundant. If you're an employee, don't read the boomerang as job security; read it as a sign of where your value lies. The people who can verify AI output are, we think, the ones companies will discover they can't do without.

The technology will keep improving, and some of those roles probably will vanish for good. But "we'll figure out the quality later" turned out to be an expensive way to find out which ones.