The displacement trap

Companies aggressively cutting jobs in the name of AI are making a strategic error, with 55% of those that made AI-linked redundancies later admitting the decisions were wrong. High-profile cases like Klarna and Salesforce illustrate the costs, including degraded customer satisfaction and the need to rehire talent after prematurely eliminating roles. The evidence suggests that durable advantage in the AI era will come from human-AI augmentation, not reflexive headcount reductions.

The companies cutting hardest may lose the AI decade. The CFO’s case for AI is the easiest in the board room. Identify the labor cost, apply the productivity multiplier, and book the saving. The business case writes itself, the board nods, and the press release follows within the quarter. This has been the dominant logic of enterprise AI adoption for the last few years. It is also, on its own terms, rational. Cost synergies are quantifiable. They appear cleanly on the P&L. They reassure investors that management is alive to the technological shift. And they offer protection against the lurking fear that some AI-native upstart, free of legacy headcount, will arrive and undercut the incumbent’s cost base. Unfortunately, this logic optimizes for the wrong variable. A growing body of evidence suggests that companies pursuing aggressive AI-driven displacement are paying costs they didn’t model. Customer satisfaction degrades. Institutional knowledge walks out the door, while competitive edge is quietly transferred to model providers. Trust within the surviving workforce erodes. And often the same talent has to be rehired anyway: the ground is shifting fast, and companies don’t yet know who they actually need. Our thesis is straightforward. The most durable advantage of the AI era will not come from reflexive headcount reductions, but from the deliberate, often uncomfortable work of human–AI augmentation: upskilling teams, redesigning workflows, and carrying people through genuine organizational change. The companies betting on displacement are settling for a cheaper present at the cost of a richer future. A note on framing is in order. Layoffs are a feature of capitalism, not a moral failing. They carry deep personal cost, but they can also be necessary — for renewal, for strategic refocus, for survival. This article is not a lament for the existence of redundancies. It is an argument that the rationale increasingly being offered for them, that AI has made human labor superfluous, is in many of its highest-profile cases both empirically wrong and strategically damaging. The data is striking. According to a 2025 Orgvue survey of more than 1,100 senior decision-makers, 39% of companies reported having made employees redundant as a direct result of deploying AI. Of those, 55% admitted the decisions were wrong 1 fn-1 . A separate Forrester report estimates that roughly half of all AI-attributed layoffs will be reversed in some form by the end of 2026 2 fn-2 . Twenty-three percent of companies that made AI-linked layoffs admitted those decisions were based on “broad expectations about automation rather than a task-level understanding of job responsibilities” 3 fn-3 . Companies cut roles before they validated that AI could reliably perform them. The cases that dominated the headlines were the same ones that became cautionary tales. Klarna , the Swedish fintech, is the case-study example. Between 2022 and 2024, the company cut its workforce from 5,500 to roughly 3,400, with CEO Sebastian Siemiatkowski publicly claiming the OpenAI-powered chatbot was performing the work of 700 customer service agents 4 fn-4 . By early 2025, customer satisfaction had dropped, complaints had risen, and the CEO was acknowledging publicly that “we focused too much on efficiency and cost” 5 fn-5 . By mid-2025, Klarna was rehiring human agents under an “Uber-style” flexible model targeting students and rural workers 6 fn-6 . The episode is now, in the words of one industry analyst, the “canonical enterprise cautionary tale for 2026: executives evaluating AI workforce strategies are increasingly required to explain how their plan avoids the Klarna outcome” 7 fn-7 . Salesforce went further, and more publicly. In September 2025, CEO Marc Benioff told podcaster Logan Bartlett that the company had reduced its customer support headcount from 9,000 to roughly 5,000 — 4,000 roles cut, “because I need less heads” 8 fn-8 . The comment landed weeks after Benioff had told an AI for Good Global Summit that AI would not wipe out white-collar jobs and would instead drive “radical augmentation” 9 fn-9 . The contradiction was not lost on observers. Standard Chartered offered the most recent and most cautionary case. On 19 May 2026, CEO Bill Winters told investors at a Hong Kong briefing that AI was “replacing, in some cases, lower-value human capital with the financial capital and the investment capital we’re putting in”, while announcing plans to eliminate roughly 7,800 back-office roles by 2030 10 fn-10 . The phrase “lower-value human capital” triggered immediate backlash. Hong Kong and Singapore regulators sought clarification. Winters issued a memo to employees the next day, then an apology on LinkedIn three days later 11 fn-11 . The damage was structural, not just reputational: the bank had publicly attached a four-year layoff plan to a phrase that re-categorized its own employees as a depreciating asset. The pattern across these cases is not the layoffs themselves. It is the communication. AI is being deployed as a blunt justification, frequently without any articulated vision for what the organization’s AI-enabled future actually looks like. There is rarely a “what comes next” narrative. There is rarely an explanation of where the freed capacity is being redeployed. The absence is telling, and it is, in many cases, a signal of the underlying strategic emptiness. This is not confined to the cases above. Meta is preparing to cut roughly 15,000 employees, 20% of its global workforce, while doubling its AI budget to $135 billion in 2026 12 fn-12 . Amazon announced 14,000 corporate layoffs in October 2025 and a further 16,000 in early 2026 13 fn-13 . Microsoft cut more than 15,000 positions through 2025, around 7% of its global headcount 14 fn-14 . The cumulative tech layoff count for 2025 alone exceeded 80,000, with industry analysts now anticipating a rehiring wave through 2026 as the practical limitations of current AI systems become apparent 15 fn-15 . The picture is not one of clean substitution. It is one of premature reduction. The displacement bias has a clean explanation: cost reduction is easy to model, and revenue uplift is hard to attribute. Cutting a salary is an unambiguous P&L event. The saving is immediate, quantifiable, and visible to investors within a quarter. Augmentation gains, by contrast, are diffuse. They show up as faster cycle times, better-quality output, customers retained rather than lost, products shipped that wouldn’t have existed otherwise. They are real, but they are difficult to isolate from the noise of normal business performance. The asymmetry of financial legibility tilts every boardroom conversation in the same direction. Clayton Christensen’s framework on disruptive innovation and jobs-to-be-done offers a sharper diagnosis. In Competing Against Luck , Christensen and his co-authors argue that companies fail when they define the “job” of a product or technology too narrowly. Asked what job AI does, the dominant answer in 2026 is: “it does the work a human used to do.” The framing is a substitution framing. It defines AI’s job as replacement. This is the wrong job. A more accurate framing puts the customer first: AI is a capability expander for the job they’re actually trying to get done. It helps organizations understand that job more precisely, and deliver against it faster and at greater scale than was previously possible. Defined that way, AI’s value is not in subtraction. It is in multiplication. In The Innovator’s Dilemma , Christensen’s central insight was that incumbents are most vulnerable to disruption when they optimize too aggressively for the metrics of their current business model. Companies pursuing AI-driven displacement are doing exactly this. They are using a transformative general-purpose technology to make their existing business cheaper to operate, rather than to build the business that comes next. The cost saving is real; the strategic positioning is regressive. There is a sharper version of this argument. If AI replaces the person, the AI vendor owns the capability. If AI augments the person, the company keeps the competitive advantage. Consider what happens when a company replaces 700 customer service agents with an AI assistant built on a third-party foundation model. The customer-facing function continues to operate, but the locus of competence has migrated. The institutional knowledge of how that company specifically handles edge cases, escalations, and brand-defining moments now lives, in any meaningful sense, inside a model owned by OpenAI, Google, or Anthropic. The company has not eliminated a cost; it has rented out a capability. Every time the underlying model improves, the company’s competitor with the same vendor contract gets the same improvement on the same day. The differentiation is gone. Contrast this with augmentation. A company that invests in training those same 700 employees, and gives them AI tools to handle three times the volume at twice the quality, retains the institutional context. The agents accumulate experience that feeds back into how the tool is deployed, how the workflow is structured, what edge cases get human attention. The competitive advantage compounds inside the company, not inside the vendor’s model. Not everyone will make the transition. Some will find the new way of working genuinely incompatible with their skills or inclinations, and natural turnover will do some of the work on its own. That is a legitimate and expected outcome — and it is very different from reflexive displacement. The distinction is not whether headcount eventually falls. It is whether the company has first defined what the new way of working actually looks like, invested in giving people a real chance to adapt, and made deliberate choices about where human judgment remains irreplaceable. The argument that a powerful new technology will eliminate human work is older than industrial capitalism. The argument has been wrong every time. The question worth asking is why it has been wrong, and what the answer suggests about the present moment. The Luddite movement of 1811–1816 is remembered, unfairly, as a parable about resisting progress. What’s more interesting is what happened after the displacement panic subsided. Mechanized textile production did not eliminate textile workers. It dramatically expanded the market for textiles, which in turn dramatically expanded the demand for skilled labor to operate, maintain, and improve the new machinery. Productivity per worker rose; total employment rose alongside it. The factories that won the Industrial Revolution were not those that cut the most labor. They were those that reimagined what was producible. Wedgwood didn’t dominate British ceramics by minimizing headcount; it dominated by industrializing design, branding, and distribution in ways its competitors could not match. A more useful pair of examples for the present moment comes from companies that built competitive advantage by going against the prevailing cost-cutting logic of their era. Cadbury and Bournville. In 1879, the Quaker chocolatiers George and Richard Cadbury made what their peers regarded as an eccentric decision: they moved their factory four miles outside Birmingham, into the countryside, and began building what they called a “factory in a garden”. By 1900, the Bournville estate included 314 cottages and houses set across 330 acres, with sports fields, gardens, schools, swimming pools, and a pension scheme for workers 16 fn-16 . The Cadburys provided medical treatment, paid holidays, and a quality of life that was, by Victorian standards, radical 17 fn-17 . George Cadbury then did something genuinely unusual: in 1900 he transferred ownership of the entire estate to the newly established Bournville Village Trust, ensuring the community would outlast any individual benefactor 18 fn-18 . The conventional wisdom of the era was that the way to compete on industrial scale was to minimize labor cost. The Cadburys did the opposite, and built one of the most enduring brands in British retail. The Bournville estate became the architectural blueprint for the garden city movement 19 fn-19 . John Lewis Partnership. In 1920, John Spedan Lewis, then running the Peter Jones department store in Sloane Square, introduced a formal profit-sharing scheme for his staff. From that point onwards, staff were referred to as Partners. In 1929, following his father’s death, Spedan Lewis signed the First Trust Settlement, transferring his shares to a board of trustees on behalf of all employees 20 fn-20 . The full transition to employee ownership was completed in 1950 21 fn-21 . Lewis’s stated principle was that profits should not flow solely to shareholders; capital deserved a reasonable but limited return, and “labor should be the recipient of the excess” 22 fn-22 . The partnership model carried the John Lewis business through the Great Depression and the Second World War, and remains in place today. Like Bournville, it is an existence proof: when the prevailing logic is cold-blooded cost reduction, going the other way — investing in trust, ownership, and long-term alignment with the workforce — can be the more durable strategic bet. There is a quiet pattern in both cases. Productivity gains from transformative technologies arrive over the long term. Pre-empting them with aggressive cost reductions in the short term often means eliminating capabilities the business does not yet know it needs. The ground is shifting; optionality has value; trust is the cheapest form of optionality available to a company that wants to keep its workforce engaged through a transition. Another useful piece of historical evidence for the present moment is an economic puzzle from a century ago, identified and explained by the Stanford economist Paul David in his 1990 paper The Dynamo and the Computer . The puzzle is this: electricity was a transformative general-purpose technology. The dynamo was invented in the 1870s. By the early 1900s, electric motors were widely available. And yet, for roughly thirty years after electricity entered the factory, measured productivity barely moved. Economists who looked at the period concluded the technology was overhyped. They were wrong; they were just looking too early. What David showed was that the productivity gains were waiting on a redesign that took decades. Factories of the late nineteenth century used “group drive” systems: a single large steam engine drove an enormous central shaft, which in turn drove dozens of machines through systems of belts and pulleys. The architecture of the factory, typically multi-story, organized around the location of the central power source, was a direct consequence of how steam power worked. When factory owners first introduced electricity, they did the obvious thing. They replaced the steam engine with a dynamo, kept the central shaft, kept the belts, kept the building. The technology was new; the workflow was old. Productivity barely moved 23 fn-23 . The breakthrough came when a new generation of factory owners abandoned the old architecture entirely. The “unit drive” approach put a small electric motor on each individual machine, which meant the central shaft was no longer required, which meant the building no longer had to be multi-story and centralized. The single-story assembly line, the defining production architecture of the twentieth century, became possible. Productivity exploded 24 fn-24 . David’s insight was that the lag was not in the technology. It was in the imagination required to redesign around it. The parallel to the present moment is instructive. Putting a chatbot on top of an unchanged customer service workflow is just swapping steam for electricity. The technology is new; the workflow is old. The displacement of a few customer service roles registers as a productivity gain in the short term, but the deeper transformation — the rethinking of what the workflow itself is for, what the team is now capable of producing, what new products and services become possible — is what compounds. Erik Brynjolfsson, who has built on David’s framework for the current AI era, has called this the “productivity J-curve”: real gains arrive only after a substantial period of investment in workflow redesign, not after the initial deployment of the technology itself 25 fn-25 . Companies that cut headcount before they have redesigned the workflow are extracting a one-time saving while forfeiting the much larger gain that comes from the redesign. A final historical parallel, narrower in scope but vivid in its implications. When Automated Teller Machines were introduced widely in the 1970s and 1980s, the conventional expectation was that bank teller employment would collapse. The opposite happened. The economist James Bessen documented the pattern in detail. The number of tellers required to run an average urban bank branch fell from twenty to thirteen between 1988 and 2004. But the cost of running a branch fell along with it, which meant banks could afford to open more branches in pursuit of market share. The number of urban bank branches increased by 43% over the same period. Total teller employment didn’t fall; in some accounts it nearly doubled, from roughly 300,000 in 1970 to over 600,000 by the early 2000s 26 fn-26 . The nature of the job changed. With cash-handling automated, the remaining work shifted to relationship management, cross-selling, and the financial advisory functions that customers genuinely valued. Tellers became part of what Bessen calls the “relationship banking team” 27 fn-27 . Automation of a task did not equal elimination of a job; the work re-formed around the new capability. The point is not that AI will replicate this dynamic in every sector. It plainly won’t. But the assumption that headcount must fall in proportion to capability gains is, historically, a recurring error. If displacement is the default, augmentation is the option requiring imagination. And it is the option being publicly endorsed by an unusually small number of senior leaders. The most visible exception is Jensen Huang. At Nvidia’s GTC conference in March 2026, Huang was asked by CNBC’s Jim Cramer why so many tech companies were citing AI as the reason for layoffs. His answer was direct: Because you’re out of imagination. For companies with imagination, you will do more with more. For companies where the leadership is just out of ideas, they have nothing else to do, they have no reason to imagine greater than they are; then when they have more capability, they don’t do more. Jensen Huang · CNBC, March 2026 Huang has returned to the same theme repeatedly 28 fn-28 . In a May interview with Channel NewsAsia, he called the narrative connecting AI to job losses “lazy” and “irresponsible”, arguing that some executives attribute layoffs to AI “to sound smart” 29 fn-29 . Demis Hassabis, CEO of Google DeepMind, made a near-identical argument the same month, calling AI-driven developer layoffs a reflection of “a lack of imagination” from employers 30 fn-30 . The competitive logic Huang is pointing to is asymmetric. A headcount cut is a measurable saving that shows up cleanly in the next quarter’s P&L. Augmentation is a compounding capability that builds an organizational moat over years. One of these protects the current business model; the other builds the next one. This is not theoretical. IBM offers the clearest publicly available illustration. In 2026, the company’s AskHR agent was automating 94% of routine HR tasks — genuine substitution at the task level, for roles that were genuinely automatable. But IBM’s total headcount had gone up, not down. “What AI does is it gives you more investment to put into other areas,” CEO Arvind Krishna told the Wall Street Journal — those areas being software engineering, marketing, sales, and roles requiring critical thinking against other humans 31 fn-31 . The substitution at the task level funded expansion at the capability level. The measurement gap remains the practical obstacle. Many executives instinctively understand the augmentation argument but struggle to model the ROI of a 10x-output employee against the clean arithmetic of an eliminated salary. This is a solvable problem; it requires investment in the right metrics, in the redesign of workflows, in the patient measurement of capability expansion over multiple quarters. The fact that it is harder than displacement modelling is precisely why companies that solve it first will hold a durable advantage. The change management challenge is real, and worth naming directly. Not everyone will make the transition, and that is fine. What matters more is how leaders frame the journey for those who will. A leader who leads with curiosity and possibility — what becomes possible, what new capabilities open up, what the team can now do that it couldn’t before — will retain dramatically more talent than one who leads with threat or ambiguity. The best people are paying attention. They are deciding, right now, which organizations have a serious answer to the question of what the augmented future looks like — and which are using AI as cover for a strategy they haven’t bothered to articulate. The shift required is in the question being asked. From: What can AI replace? To: What does AI make possible? There are three questions every board should be asking before broaching headcount: This is the augmentation diagnostic. The most expensive talent in any organization is its senior expertise, and that expertise is almost always under-deployed — consumed by reporting, scheduling, document preparation, and the administrative tax that comes with seniority. The AI question is not “who can we replace?” It is “which of our best people could we make twice as productive by removing the bottom 30% of their workload?” The output of an augmented expert is not the same expert doing the same job faster. It is the expert pursuing the next thing they wouldn’t otherwise have time for: a new product line, a customer segment, a strategic initiative the company couldn’t previously staff. The right answer to this question is uncomfortable; it reveals capabilities the organization is currently failing to build. This is the disruptive-innovation question. The companies that win the next decade will be those that use AI to enter markets and offer products that were not viable at their existing cost structure. The Klarna case is instructive in reverse: had Klarna deployed AI to enable its existing agents to offer financial advisory services to customers who couldn’t previously access them, it would have built a new capability rather than corroded an existing one. The implications for hiring, talent strategy, and organizational design are direct. The job description changes before the headcount does. The traits that matter most in an AI-augmented organization — judgment, taste, the ability to direct AI tools toward valuable work, the willingness to learn fast — are not the traits that current hiring funnels are calibrated for. The companies that win the talent market of the next decade will be the ones that have publicly committed to augmentation as a strategy. The best people are choosing accordingly. The displacement bias in enterprise AI adoption is understandable. It is financially legible, easy to model, easy to defend to a board, and consistent with the prevailing logic of cost discipline. It is also, on the evidence so far, strategically mistaken. The empirical record of 2024–2026 is a record of cases where the displacement logic underperformed its own business case. Klarna walked back. Salesforce contradicted itself within weeks. Standard Chartered apologized. The companies that committed publicly to the substitution thesis are now, on average, the ones explaining why the gains did not materialize as predicted. Meanwhile the companies pursuing augmentation, quieter in their announcements, slower in their headline metrics, are building the capability stacks that will define the next decade. The window of differentiation remains open. The question now is no longer whether AI will change every organization — that is settled — but how, and who decides. The leaders who get this right will look obvious in hindsight. The leaders who get it wrong will have spent the most consequential technological transition of their professional lifetimes optimizing for the wrong variable. Like most traps, the displacement trap springs by default — along the path of least resistance, taken without notice. The way out is not to slow down on AI. It is to recognize the unfamiliar terrain, ask better questions, and find the imagination to answer them.