In February 2024, Klarna published an official press release that took over the tech news cycle. The buy-now-pay-later giant announced that in its first month, its AI customer service assistant did the work equivalent to 700 full-time agents. The reported metrics were eye-catching: customer satisfaction matched human reps, and the tool was projected to drive a $40 million annual profit improvement. The headline even came with an on-the-record endorsement from an OpenAI executive. At the time, Klarna became the ultimate poster child for AI replacing customer support.
Fourteen months later, at the very same company, the CEO vented to Bloomberg: “Cost unfortunately became too high of a weighting factor when putting this together, and the quality dropped as a result.” Shortly after offering that reflection, they reversed course and started recruiting human customer service agents again.
Both of these contradictory developments are completely real. There was no plot twist behind the curtain: the CEO stayed the same from start to finish, and the AI assistant never went offline for a single day. What actually flipped was the management calculation: they scrapped the scheduling logic that treated cost as the sole metric, putting real people back at desks to handle the messy problems machines couldn’t untangle. Looking across similar stories from 2024 to 2026, media headlines only give you scattered fragments; piece them together, and you get a complete closed loop: Step one, companies slash headcount aggressively on the promise of AI; Step two, once the cuts take effect and operations hit a wall, companies quietly hire people back.
Both steps are happening right now. Most people only notice the first step, because step one gets a press release, while step two only shows up in financial footnotes and job listings.
Step one did happen, and it is the part everyone can see. In January 2026, Meta mapped out an organizational overhaul called Project OT, where the most aggressive scenario planning contemplated cutting certain teams by 60%; on May 20, the first wave of 8,000 layoffs went ahead as planned. In its FY2026 10-K annual report disclosure, Oracle attributed job cuts to AI with SEC-filing precision: “The adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in workforce reductions,” marking a net reduction of 21,000 people in a single year, or about 13%. Amazon’s CEO sent an all-hands letter in June 2025 warning that “the total employee count will shrink over the next few years,” followed by two rounds of layoffs cutting around 30,000 people. Intuit slashed 17% in one stroke. Coinbase cut 14%, flattening its management hierarchy to within five layers and replacing dedicated managers with player-coaches.
Challenger’s industry tracking data showed that starting in March 2026, AI was the number one layoff reason cited by US employers for five consecutive months. The headlines weren’t lying.
Yet beside these noisy grand narratives, there is another number that needs to sit right next to them: in 2025, approximately 1.2 million people across the US lost their jobs, of which about 55,000—or 4.5%—were directly attributed to AI. Being the top cited reason reflects narrative momentum; 4.5% is the actual share of the workforce affected. A wide chasm separates the two numbers, paved at the bottom with an explanatory framework that investors much prefer to hear.
People in management consulting put this with brutal clarity: when a company claims it is cutting jobs because of AI, it delivers two pieces of good news simultaneously on its earnings call—it is embracing the frontier future, and it is cutting operating expenses. Admitting that business is deteriorating offers zero good news. So far, no regulatory agency has scrutinized the veracity of layoff rationales; regulatory attention remains trained on companies hyping AI capabilities in investment disclosures.
The finale of this play is decided by step two. No matter how large step one looks, it is only an invitation to step two.
Hitting the wall after headcount cuts: step two unfolded just as tangibly. The fastest mover was Commonwealth Bank of Australia. In July 2025, they eliminated 45 customer service roles, citing that a newly deployed voice bot made customer service easier.
A month later, they rescinded the layoff decision, apologized to employees, and admitted they “did not adequately assess all relevant business factors.” Union details revealed what happened: after the bot went live, call volumes surged instead of dropping, forcing management to add overtime shifts while pulling team leads to answer phones.
That was merely the fastest case. IBM announced a hiring freeze in 2023, saying 7,800 back-office roles would be replaced by AI over five years. Three years later, their CHRO announced at a summit that US entry-level hiring in 2026 would expand to three times its size, adding verbatim: “Yes, exactly the jobs people told us AI would do.” After its AI visual inspection system became overwhelmed, Ford rehired hundreds of experienced quality engineers. Media-cited Robert Half industry survey data showed that 32% of US hiring managers who eliminated a role due to AI rehired for the exact same position; a survey covering more than 2,500 enterprise decision-makers found that 74% of companies rolled back customer-facing AI that had already gone live; and media-cited Gartner projections indicate that by 2027, half of the companies that reduced customer service staff due to AI will hire them back.
Looking at this long string of rollbacks together, two details are more telling than the rollbacks themselves. One is about where they hit the wall; the other is about the technology itself.
The first detail: every single rollback was triggered by a collapse in quality signals, not a single one by failing to meet cost targets. Commonwealth Bank of Australia ran into a rebound in call volumes. IBM’s internal HR system ran smoothly on 94% of routine requests, but choked on the remaining 6% of hard cases—the very ethical judgment calls where errors were least tolerable. Ford collided with long-tail defects that automated systems failed to identify. Not a single company came out saying AI was too expensive to justify; the cost math worked out every time. What collapsed was the other ledger: failing to deliver acceptable quality.
The second detail: the technology itself was never what got pulled. After Klarna resumed hiring customer service agents, the workload handled by AI according to official company statements actually rose rather than fell, climbing from the equivalent of 700 full-time employees to 853. The cost-saving goal remained intact, just achieved through a hybrid model instead. The technology settled into its rightful place: fielding simple queries, while leaving a human channel open for complex problems.
There is another pattern to these rollbacks: the speed of error correction is directly proportional to how close the function is to the customer. Customer service went from layoffs to reversals in as little as a month, because call volumes speak immediately. For back-office functions far from customers, correction cycles take years. When Duolingo’s CEO admitted over a year later that they scrapped AI evaluation criteria, the phrasing used was “we’ve walked backward”—a phrase that reached the public exactly one year after the famous AI-first all-hands letter. Looking back across these reversals, companies never pulled back the technology itself; what they rescinded was an organizational plan that treated replacing humans as its goal.
At this point you might ask: these companies aren’t foolish, so why is step one always launched with utter confidence, while hitting the wall always arrives on a delay? The root cause is that the acceptance metrics used to justify headcount cuts and the places where damage actually happens do not live in the same spot.
Executives mostly look at average metrics in their status reports. Customer service looks at average response times and mean satisfaction scores; engineering looks at code change volumes and commit output. Averages always look pristine in weekly reports, yet damage occurs entirely in the tails: the few most complex customer escalations, the trickiest defect categories, the 6% of cases requiring human ethical judgment. Average metrics are inherently blind to the tail. An average response time can look fantastic even while nobody is around to catch the thorniest 6% of problems.
Meta’s internal data quantified this mechanism at Big Tech scale. According to a Reuters 2026-08-26 investigative report, after cutting middle management and unleashing agents, internal code changes jumped 220% in a year, while new features actually reaching users rose only 36%. Major incidents spiked by 40%, and the time employees spent firefighting surged by 70%. Tools for writing code multiplied and accelerated, but the people reviewing and cleaning up the mess did not. Code piled up in the pipeline until it reached users as outages. A dataset later released by a Stanford team showed the exact same pattern: a mid-sized software firm mandated that per-capita output double; two years later, individual throughput did reach 2.09 times the baseline, but each engineer’s review load roughly doubled as well. Human review coverage with substantive comments shrank from nearly 90% to under 70%, with automated review tools filling the gap. Elongated review wait times became the everyday symptom of that clogged pipeline.
There is another side to the same dynamic. Management consultancy BCG quantified it in a 2026 survey of more than 11,000 professionals: 47% reported spending more time directing and managing AI than doing their primary job.
Connecting this evidence makes the logic between step one and step two clear: average metrics are all green when layoff decisions are made because averages don’t measure the tail; tail damage takes time to accumulate, and by the time it surfaces in call volumes, incident rates, and management exhaustion, the layoffs have already been carried out. The few companies that avoided hitting the wall took a different approach precisely here: automation software company Zapier ran a company-wide one-week AI experiment in 2023 to measure genuine capacity before restructuring. Its founder later noted that while the number of agents inside the company had surpassed employee headcount, they were still hiring.
Hitting the wall wasn’t bad luck. The acceptance metrics were blind to the cliff from day one—that was the root of it.
Placing this two-step cycle in the broader labor market yields a third layer of observation. Headlines about AI wiping out jobs and statistical shifts in employment tell two very different stories.
The team at the Stanford Digital Economy Lab updated their research in an August 2026 revised paper, drawing on monthly data from ADP, the largest payroll processor in the US, covering millions of workers. The authors noted these are descriptive metrics rather than causal estimates. Their primary finding was clear: they found no evidence of economy-wide, large-scale job displacement. The real structural squeeze is concentrated in a specific demographic: young workers aged 22 to 25 in roles with the highest AI exposure saw employment levels 19% below what they would have been had they kept pace with their low-exposure peers—a gap that stood at 13% a year earlier and continues to widen. Two details are even more critical. First, the gap primarily stems from companies hiring fewer young workers, not laying them off: the entry door is narrowing, rather than the existing stock being purged. Second, jobs face two diverging fates: roles where AI replaces humans experienced declining employment, while roles where AI assists humans remained flat or even grew, particularly for senior workers.
Hiring platform data confirms this divergence right at the front door: job postings for senior roles rose 14.7% over the year, while mid-level postings dropped 6.7%. Tracking by workforce analytics firm Revelio showed that companies aggressively adopting AI grew overall headcount 27% faster than peers that didn’t, with growth concentrated in senior positions. Salesforce is the most direct example: they eliminated 4,000 customer support roles, letting AI handle half of customer conversations, while in that same year total headcount reached an all-time high—the additional hires were all sales reps selling AI products.
Putting these figures together leads to a counterintuitive conclusion: AI adoption intensity is not correlated with headcount reductions—it is actually negatively correlated. Companies using AI as a capacity tool are expanding headcount; companies using AI as a layoff narrative are walking it back. For job seekers, this distinction matters far more than whether AI will “take away jobs.” What is genuinely shrinking is a specific niche: entry-level roles with well-defined boundaries where AI can directly substitute for humans—writing boilerplate code, tier-one customer service, routine content creation. What is expanding is the other end: evaluating whether AI output is actually good, defining acceptance criteria for agents, and cleaning up messes when automation goes off the rails. The generation end is getting cheaper; the absorption and verification end is getting more expensive and facing greater talent shortages. At the aggregate level, AI hasn’t taken away the rice bowl; it has simply moved where the bowl sits.
Back to the opening question: Will AI actually take your job? By mid-2026, with the first two acts played out, the outline of the answer has emerged: it isn’t a yes-or-no multiple-choice question, but a series of far more specific inquiries.
First question: Does your work sit on the generation side or the absorption side? Tasks on the generation side—having AI write code, draft first passes, answer straightforward questions—are getting cheaper fast. Tasks on the absorption side—judging whether output is correct, setting acceptance criteria, handling edge cases and incidents—are getting more expensive and running short of people. These two ends followed opposite trajectories within the very same company: Salesforce scaled back customer support in 2025 while expanding AI product sales roles in the exact same year.
Second question for your organization: Does it have an absorption and verification layer? When a company lays off dedicated managers and leaves one person saddled with 15 direct reports, it saves a single line on the payroll while potentially tossing away the final human checkpoint that intercepts agent misbehavior. Meta’s numbers prove just how valuable that checkpoint is, and show how quickly outages and firefighting surge without it. In the agent era, middle managers act as verification nodes in the system, rather than mere bureaucratic friction.
Third question to keep in mind for the next round of layoff headlines: Has the company published internal data, is the sequence right, and did it hire people back a year later? Apply very different credibility discounts to the two types of companies: those that published internal metrics and adjusted course accordingly, versus those running purely on narrative. Meta’s own data convinced it to tap the brakes—and to date, it remains the only company to do so.
Including ourselves, many of us spent the past two years predicting agents would take over jobs. Connecting what unfolded across 2026, that prediction was only half right: AI is taking over the simpler half of the task list, while making the harder half far more expensive and critical. The cycle keeps turning; with every lap, companies fall into the same ditch, and with every lap, jobs are redefined.
Rather than asking where the finish line is, take a close look at which side of the loop you stand on. Only when you see your position clearly do you know which skills to build—and whose headlines to believe.