The narrative is appealing in its simplicity: AI makes workers more productive, therefore you need fewer workers, therefore headcount can be reduced. It is also, in most cases, wrong — and the companies that have built strategies around this logic are beginning to discover the difference between what sounds plausible in an investor presentation and what actually happens in practice.
This article makes a specific argument. It is not that AI doesn't increase productivity — it does, in documented and meaningful ways. The argument is that productivity gains in specific tasks are categorically different from the elimination of whole jobs, and that treating them as equivalent produces decisions that damage organisations while resting on a logical error.
What the productivity research actually shows
The strongest evidence for AI's productivity effects comes from a small number of carefully designed field studies. Noy and Zhang (2023), in a pre-registered experiment with 453 college-educated professionals, found that access to ChatGPT reduced the average time taken on professional writing tasks by 40% and raised output quality ratings by 18% (Noy & Zhang, 2023, Science). Brynjolfsson, Li, and Raymond (2023), studying 5,172 customer support agents at a Fortune 500 firm over 3 million real customer interactions, found a 14% increase in issues resolved per hour, with the largest gains — 34 to 35% — concentrated among the least experienced workers (Brynjolfsson et al., 2023, Quarterly Journal of Economics). Peng et al. (2023) found that GitHub Copilot roughly halved the time required for a specific software development task.
These are real, significant, reproducible effects. They are also effects on specific tasks within specific jobs — writing, customer chat responses, targeted coding problems. They tell us precisely nothing about whether the jobs containing those tasks can be eliminated.
The task-job gap: why this distinction matters
Every job is a bundle of tasks. AI can accelerate, assist with, or automate some of those tasks. But even where AI can handle a significant share of a job's task components, it does not follow that the job as a whole becomes redundant.
Nobel Prize–winning MIT economist Daron Acemoglu, in a widely cited 2024 analysis applying a task-based macroeconomic model, estimated that only approximately 5% of economy-wide tasks are both exposed to AI and can be profitably automated within the next decade (Acemoglu, 2024, "The Simple Macroeconomics of AI"). This is not a claim that AI is unimportant. It is a precise claim about the gap between what AI can do in a controlled experiment and what it can cost-effectively replace across the real-economy range of messy, varied, context-dependent work situations.
The OECD's 2024 analysis of AI's impact on productivity, distribution and growth reaches a similar conclusion: AI's productivity effects, while real, are concentrated in specific knowledge-intensive tasks and are not yet evident in aggregate productivity statistics — which as of 2024 showed only modest growth of around 0.4% across OECD countries, with no detectable AI signal at the macro level (Filippucci et al., 2024, OECD AI Papers, No. 15).
The explanation for why micro-level task productivity gains don't produce macro-level employment reductions is structural. When a worker completes a task 40% faster, the time saved is typically absorbed into other tasks within the same job, into higher quality output, into managing the AI's errors and edge cases, or into the expanded scope that often follows demonstrated productivity improvement. The job changes. It does not disappear.
The case of Klarna: what full replacement actually produces
Klarna became, in 2023 and 2024, one of the most cited examples of AI replacing human labour at scale. The Swedish fintech company deployed an AI assistant — developed with OpenAI — that it claimed was performing the work of 700 customer service agents, handling approximately two-thirds of all customer interactions. The AI resolved issues in under 2 minutes compared to 11 minutes for human agents. Headcount fell from over 5,500 to around 3,500. The story was covered globally as evidence that AI-driven workforce reduction had arrived.
By mid-2025, Klarna was rehiring. CEO Sebastian Siemiatkowski publicly acknowledged what the data had shown: customer satisfaction had deteriorated, service quality had declined on complex interactions, and the cost savings that had been projected had not fully materialised. "We focused too much on efficiency and cost," Siemiatkowski told reporters. "The result was lower quality, and that's not sustainable." The AI had been effective at handling routine queries. The interactions it could not handle well — emotionally charged, multi-step, requiring genuine empathy or nuanced judgment — had accumulated, gone unresolved, and eroded the customer relationship.
The Klarna case is not an argument that AI tools have no place in customer service. It is an argument about what happens when task-level capability is treated as role-level replaceability. The AI did well at the easy queries. The hard queries — the ones that most matter to customer retention and brand perception — remained a human problem.
The Amazon Just Walk Out case: the hidden workforce
A different illustration of the same underlying problem emerged from Amazon's "Just Walk Out" technology — the AI-powered, cashier-free checkout system rolled out across its Fresh grocery stores. The technology was presented as a showcase of autonomous AI: customers pick up items and leave, with AI handling the entire checkout process through computer vision and sensor data.
In 2024, The Information reported that the system had been relying on more than 1,000 workers in India to manually review shopping footage, with roughly 7 out of 10 transactions requiring human verification. Amazon maintained that these workers were labelling data and providing backup for the AI rather than operating it — but the delayed receipts that customers received hours after leaving the stores were the visible evidence that human review was embedded in what had been presented as an automated process. Amazon subsequently wound down Just Walk Out at its own grocery stores.
The case illustrates a pattern that appears repeatedly across AI deployments: the tasks that AI can handle autonomously are a subset of what the full workflow actually requires. The remainder stays human — sometimes visibly, sometimes not.
What the layoff data shows about AI-motivated cuts
Between 2022 and 2025, the technology sector saw over 685,000 cumulative job cuts — the largest correction since the dot-com collapse (Layoffs.fyi tracker data). The causes were mixed: pandemic-era overhiring, rising interest rates, and genuine AI-driven restructuring. But the claims made at the time of announcement are worth examining against subsequent performance.
A consistent pattern has emerged: companies whose layoffs were specifically framed around AI replacement — rather than overhiring correction — have disproportionately either underperformed on the promised efficiency gains or begun quietly rehiring in adjacent roles. Research from The Burning Glass Institute found that in AI-exposed fields, between 2018 and 2024, entry-level job postings fell sharply (software development junior roles from 43% to 28% of total postings; data analysis from 35% to 22%) while total job postings in those fields stayed flat and senior hiring remained stable. Companies were not hiring fewer people overall. They were eliminating entry-level positions and expecting senior staff to absorb the work that AI was supposed to have replaced.
The International Center for Law & Economics review of empirical AI and labour market evidence (2025-2026) found that "aggregate labour market effects through 2024-2025 remain limited in most datasets" and that studies "find no evidence of immediate economy-wide labour displacement." The observed effects have been concentrated in entry-level roles in specifically AI-exposed occupations, with younger workers carrying a disproportionate share of the adjustment cost.
The positive case: what AI productivity gains actually enable
The more useful question for leaders is not "how many jobs can we cut" but "what does the team now become capable of that it wasn't before?" This is not a rhetorical reframe. It has material implications for business outcomes.
When AI tools raise individual productivity by 14-40% on specific task categories, the people in those roles can do one of several things with the released capacity: produce more output of the same kind, produce higher quality output, take on more complex or judgment-intensive work that was previously too time-consuming, or develop skills that make the organisation more capable over time.
The organisations that have captured the most value from AI productivity gains are, consistently, those that have directed the capacity into the third and fourth categories — using AI as a capability expander rather than a headcount justifier. A consulting team that uses AI to handle research synthesis and first-draft production doesn't need to be smaller. It can serve more clients, tackle more complex problems, and produce better work than a larger team did before. A software team using AI coding assistants doesn't need fewer engineers. It can build faster, ship more features, and reduce technical debt in ways that weren't previously affordable.
The Brynjolfsson et al. study is instructive here: the largest productivity gains from AI went to the least experienced workers, effectively compressing the performance distribution. This is the augmentation case in practice — AI making the whole team better, not replacing part of it. The experienced workers' judgment remained essential; what changed is how effectively the full team could apply it.
What good leadership looks like here
For leaders navigating AI adoption, the evidence suggests a few practical conclusions.
Do the task-level analysis, not the headcount analysis. The right question is: which specific tasks within which specific roles can AI assist with or accelerate? Not: how many roles can AI replace? The first question produces useful decisions. The second produces the Klarna outcome.
Treat productivity gains as capacity, not as cost reduction. When AI tools free up time within existing roles, the value is in what that capacity is directed toward — not in how quickly it can be converted to a smaller payroll. The organisations that have done this have built compounding capability advantages. The ones that have converted it to immediate headcount reduction have often found themselves with the same workload, fewer people, and service quality problems that their competitors have capitalised on.
Recognise the tasks AI cannot do. Empathy in difficult conversations. Judgment in genuinely novel situations. Accountability for decisions that affect people. Trust-based relationships with clients and colleagues. These are not temporary gaps in AI capability waiting to be closed. They are structural features of what makes human professional judgement valuable — and they are disproportionately present in the high-stakes interactions that most determine whether clients stay and whether teams perform.
Invest in development rather than elimination at the junior level. The pattern of eliminating entry-level roles is eroding the pipelines through which senior capability is built. The organisations cutting junior positions today are creating capability gaps that will surface in five to eight years, when the people who would have become their experienced mid-career professionals were never developed. The short-term payroll saving is real. The long-term capability cost is real too, and larger.
The trap
The AI productivity trap is the gap between what the evidence says and what certain business narratives claim. The evidence says: AI can meaningfully raise productivity on specific tasks, for specific workers, in specific contexts. That is genuinely valuable, and leaders who ignore it are leaving real capability on the table.
What the evidence does not say — and what the Klarna rehirings, the Amazon revelations, the flat aggregate productivity statistics, and the Acemoglu analysis all point toward — is that task-level productivity gains justify workforce reductions at scale. They don't, for structural reasons that are unlikely to change in the near term: because jobs contain more than their AI-automatable tasks, because the tasks AI handles worst are often the most consequential, and because the capacity freed by AI has more value directed toward higher-quality work than toward a smaller headcount.
The leaders who understand this distinction will build organisations that are genuinely more capable. The ones who don't will cut their way into the same problems Klarna is now hiring its way back out of.
References
- Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.
- Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–953. (Working paper originally 2023.)
- Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv:2302.06590.
- Acemoglu, D. (2024). The simple macroeconomics of AI. Economics Letters, 261, 111625. MIT Working Paper version available via NBER.
- Filippucci, F., et al. (2024). The impact of artificial intelligence on productivity, distribution and growth: Key mechanisms, initial evidence and policy challenges. OECD Artificial Intelligence Papers, No. 15. https://doi.org/10.1787/8d900037-en
- OECD (2025). OECD Compendium of Productivity Indicators 2025. OECD Publishing, Paris.
- Dell'Acqua, F., et al. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper 24-013.
- International Center for Law & Economics (2026). AI, productivity, and labor markets: A review of the empirical evidence.
Want to explore this further?
Start with a free 30-minute discovery session — no commitment required.
Book a Discovery Call