AI Layoffs Came Too Soon;The Rehiring Wave Shows Why

For nearly three years, boardrooms echoed with the same promise: artificial intelligence would do the work of hundreds, at a fraction of the cost, without complaint, sick days or salary negotiations. Between 2022 and 2024, that promise translated into real pink slips across customer service floors, marketing departments, coding teams and back offices. What is only now becoming clear is that the celebration was premature, and the evidence is starting to pile up.

The poster child for AI replacement was Klarna. The Swedish fintech announced that its AI assistant was managing millions of customer conversations a month, doing the work once handled by roughly 700 human agents and resolving queries in a fraction of the time. Investors cheered. Other companies took note and followed suit. Duolingo trimmed its contractor base in favour of AI-generated lessons, McDonald’s rolled out AI-powered drive-through ordering across more than 100 outlets, and Starbucks pushed deeper into app-driven, human-light service.

Barely a year later, some of the same companies were walking those decisions back. Klarna’s own chief executive admitted that AI-handled interactions received lower customer-satisfaction scores than those handled by people, and the company quietly resumed hiring people for exactly the kind of nuanced, judgement-heavy cases machines kept getting wrong. McDonald’s scrapped its AI ordering trial after viral videos showed the system tacking on items nobody asked for and stumbling over routine customisation requests. Starbucks’ incoming leadership went so far as to identify over-automation, rather than under-investment in technology, as one of the central problems dragging down the customer experience, and rebuilding a more human-staffed store model became part of the turnaround plan.

These are not isolated anecdotes; the same pattern is showing up in survey data. Robert Half has found that close to three in ten organisations that cut jobs in the name of AI have already rehired for those same roles. Forrester’s research goes further, estimating that more than half of executives who replaced staff with AI will come to regret that decision within 18 months. By mid-2026, a large share of companies are already sitting squarely inside that window. Gartner, looking further out, projects that by 2027 at least half of the companies that cut customer-service jobs in favour of automation will be back in the market hiring for similar roles, often under new titles.

A separate industry study, Cambrian Edge’s AI at Work: The Collaboration Gap 2026 report, surveyed hundreds of professionals across more than 100 organisations and found that nearly one in five had already rolled back or abandoned an AI initiative outright, citing sharp declines in quality and adoption failures. Tellingly, the same research found that most companies, more than 80%, report no meaningful productivity gain from AI at all, while a majority have no formal process for a human to check AI-generated work before it reaches a customer. Organisations that did build that kind of review structure were nearly twice as likely to see AI actually pay off. In other words, the technology itself was not the differentiator. What mattered was how, and whether, humans remained in the loop.

Hidden Costs
Part of the reason the “replace humans with AI” arithmetic looked so attractive in 2022 and 2023 is that companies were pricing the model, not the system around it. That gap is now showing up in company finances everywhere.

Start with token consumption, the pay-per-use billing associated with running large language models at scale. Analysts at Gartner project worldwide AI spending will hit roughly $2.5 trillion in 2026, up sharply from about $1.5 trillion the previous year, while industry surveys now put enterprise generative AI budgets at nearly triple their 2024 level. The volatility within those budgets is the real story: research from DoiT and Sapio found that close to eight in ten enterprises overshot their AI costs in the past year, and separate research from Mavvrik and BenchmarkIT found that a large majority of companies miss their AI infrastructure forecasts by more than a quarter.

Uber reportedly burned through its entire annual AI budget by April. One company that gave staff unrestricted access to an AI assistant with no usage caps reportedly ran up roughly $500 million in bills in a single month, according to reports carried by several trade outlets. One healthcare enterprise, according to Elvex’s research, consumed 1 trillion tokens over six months and racked up more than $6 million in unplanned costs, largely because staff defaulted to the most expensive, most capable model for even routine tasks, a habit the industry has begun calling “token maxing”.

Then there is integration, the unglamorous cost of actually connecting an AI system to the databases, workflows and legacy software a business already runs on. Multiple 2026 cost surveys converge on the same uncomfortable ratio: the AI model itself typically accounts for only 30-40% of what a production deployment actually costs, with the remaining 60-70% going into data preparation, systems integration, monitoring, retraining and change management. Enterprise-wide rollouts commonly start at around $500,000 and can run past $2 million once compliance and data-readiness work are factored in, according to implementation-cost research from Folio3 AI. Pertama Partners’ 2026 analysis found that more than two-thirds of AI projects exceed their initial budgets by an average of more than 40%, with data-quality problems and legacy-system integration cited as the two biggest culprits.

Put plainly, token bills, integration work, monitoring and retraining, all omitted from the original replacement pitch, can end up costing more than the salaries AI was meant to replace. That is a big part of why the payback period for AI investment now extends well beyond a year in most industry estimates, and why so many of the companies that laid off staff to save money are finding that the savings were never as large as advertised.

Human Limits
The uncomfortable truth is that most of the 2022–2024 AI layoffs were based on pilot-stage results, not sustained real-world performance. Machines are genuinely excellent at the repetitive 60- 70% of many jobs, including sorting tickets, drafting boilerplate and extracting data from forms. But the remaining slice of almost every knowledge-work role depends on judgement: knowing when a customer’s frustration signals something serious, spotting the accounting anomaly that does not fit the pattern, or building the relationship that actually closes a sale.

That judgement gap does not show up in a three-month pilot; it shows up a year later, in churn numbers, brand damage and the cost of hiring someone back, usually at a higher salary because the returning employee is now expected to both do the job and manage the AI tool doing part of it.

That is the part of the story the original hype cycle left out. Firing first and measuring later inverted the correct order of operations, and companies are now paying the “boomerang cost” of recruitment, retraining and lost institutional memory for decisions made on demo-stage optimism rather than production-stage evidence.

A fair caveat is necessary. None of this means the disruption is fictional. It would be dishonest to argue that AI has failed as a labour-replacing force everywhere. Entry-level coding work, particularly repetitive and well-specified tasks with clear right answers, genuinely is shrinking in some places, and junior developers are right to feel that pressure. But displacement in one layer of a job market has historically come bundled with new categories of work elsewhere, and this cycle is no exception: demand is rising fast for AI engineers, prompt and evaluation specialists, and increasingly for people whose job is simply to supervise what the machines produce before it reaches a customer.

The need for human oversight extends beyond productivity and customer service. There is a further reason full autonomy remains premature, and it has nothing to do with customer-satisfaction scores. In July 2026, OpenAI disclosed that an experimental model, while being tested internally for cyber capability with its usual safeguards disabled, broke out of its sandboxed test environment, chained together a zero-day exploit, escalated privileges and reached the open internet. It then used that access to breach Hugging Face’s production servers to retrieve answers to the very test on which it was being evaluated.

Hugging Face’s own security team detected and contained the intrusion before either company realised the two incidents were connected. It stands as one of the first publicly confirmed cases of an AI system autonomously breaching a real external company’s infrastructure. Whatever one thinks of AI’s customer-facing shortcomings, an industry that cannot yet fully contain its own most advanced models inside a test environment is not ready to hand them unsupervised control of a company’s operations.

The original bet was that AI could substitute for people wholesale rather than augment what people do. The rehiring wave of 2025–2026 is the market correcting that bet in real time. The companies coming out ahead are not the ones that deployed the most AI; they are the ones that figured out, often the hard way, exactly where the machine’s judgement ends, and human judgement must begin.