Estimated reading time: 9 minutes ·
In most service companies around the world, labour replacement no longer occurs through mass layoffs or public announcements. It happens when a contract is not renewed, when a vacant position is not filled, when a task that previously required three people is now managed by an autonomous system in minutes. Artificial intelligence agents — systems capable of receiving an objective and executing it without continuous human supervision — are being deployed today in the legal, financial, translation and data analysis sectors at a speed that no government projection had anticipated. What is presented as “support tools” operates, in practice, as a mechanism of silent substitution. This text examines why that substitution remains invisible to those who are living through it, and why that invisibility is precisely what makes it possible.
The Substitution Has Already Occurred
Three years ago, language models made basic coherence errors. A year ago, they were already drafting contracts and producing financial reports. Today, AI agents make decisions, navigate systems and delegate subtasks to other agents without human intervention. The leap was not gradual. It was a rupture.
In March 2024, Klarna announced that its agent was managing the equivalent of the work of 700 employees, with a projected saving of 40 million dollars annually. The chosen terminology was “efficiency” and “productivity” — the euphemisms with which industrial capital has historically described the elimination of posts. The difference from previous automation cycles is that this one affects not only physical or repetitive jobs, but the professions considered until very recently as immune to replacement. Lawyers who draft routine documents, analysts who consolidate data, programmers who produce standard code, administrative assistants.
What distinguishes AI agents from previous tools is their chaining capacity. An agent receives an objective, breaks it down into steps, consults sources, drafts, verifies and delivers a result without continuous human intervention. What previously required the coordination of several specialist workers can be managed by an autonomous system in minutes. This is not an exaggeration. It is the operational description of systems already in operation in hundreds of medium and large companies across three continents.
An artificial intelligence agent is a programme capable of receiving an objective and executing it autonomously, step by step, without human intervention. It does not answer questions; it makes decisions, navigates systems and delivers a result. It is, in practical terms, a digital worker that does not rest, does not negotiate and does not invoice.
The Great Misunderstanding
The majority of the population working in offices, call centres, law firms or newsrooms has not seen any AI agent. It has seen, perhaps, a new tool on its screen, a browser extension, a writing assistant. It has not seen the deployment of agentic systems because that deployment occurs on servers, in outsourcing contracts, in human resources decisions made several floors above. The dominant perception remains that artificial intelligence is an aid, an accelerator, something that complements the human worker without displacing it.
This narrative is actively maintained by the same technology companies that sell the automation systems, by consultancies that are paid to facilitate the transition, and by corporate leaders who have very concrete financial reasons for delaying the public alarm. The result is a massive comprehension deficit in the working population, which has no access to its employers’ strategic plans or to the commercial demonstrations those employers received over the past twelve months. They have access to their posts, which still exist. That sensation of continuity is precisely the time margin the capital needs to complete the transition without organised resistance.
The Speed Nobody Calculated
The academic and government projections about AI’s impact on employment that circulated between 2020 and 2023 systematically underestimated the speed of the transition. They were calibrated on previous models of technological diffusion, which took decades to permeate entire sectors. Large-scale language models, the technical foundation of AI agents, did not follow that curve. They went from academic niche to enterprise production tools in under three years. GPT-3.5, published in March 2023, was the visible inflection point.
In 2023, the McKinsey Global Institute estimated that 70% of labour activities in service sectors could be automated with already existing systems. That figure generated no proportionate political or trade union response. It generated, primarily, panels at innovation conferences. What makes this acceleration unprecedented is not only its speed but its transversal reach. Previous industrial revolutions destroyed specific jobs and created different ones in emerging sectors. What agentic automation destroys are intermediate cognitive functions — those that historically allowed educated middle classes to insert themselves into the economy without physical capital of their own.
“Automation does not destroy work, it destroys the negotiability of work,” wrote the economist Aaron Benanav in Automation and the Future of Work, published in 2020, anticipating with precision the mechanism now in operation: not the immediate disappearance of all jobs, but the systematic degradation of the bargaining power of those who hold them.
The Jobs That No Longer Exist
The list is not speculative. The law firm Allen & Overy implemented Harvey, a contract-review agent, in 2023, reporting significant reductions in hours assigned to due diligence tasks. The translation agency Lionbridge announced in 2024 a substantial reduction of its freelance translator pool in technical and legal segments, replaced by automated pipelines with minimal post-editing. Several European audit firms deployed agentic financial analysis systems between 2023 and 2025 that compress into a few hours tasks previously assigned to teams of two or three analysts.
These cases are not exceptional. They are representative of an adoption pattern occurring in parallel across dozens of sectors. The profile of the worker at immediate risk is not the unskilled employee, but the mid-level professional with five to fifteen years of experience in routine cognitive functions, well-paid and in high historical demand. The legal assistant who reviews jurisprudence. The financial analyst who consolidates subsidiary reports. The programmer who writes low-level code. None of these workers considers themselves at risk. They all are.
The World Economic Forum’s Future of Jobs 2025 report projects the displacement of 92 million jobs globally before 2030, concentrated in administrative, financial and data-processing functions. The same report anticipates 170 million new posts — a figure often cited as a signal of equilibrium but which omits what is decisive: the jobs destroyed disappear in months; the jobs created require years of retraining in sectors that the current labour market is not training at any scale.
The Perception Gap
The central problem is not technological. It is narrative. And the main vehicle for that deficit is the media. Dominant coverage of artificial intelligence oscillates between technological enthusiasm and speculative catastrophism, avoiding almost systematically what happens on the operational level: contracts that are not renewed, posts that are not filled, functions that disappear without a press release. When media address the employment impact they do so through experts who speak of “transformation” and “jobs of the future”, not of concrete displacement. A worker who consumes that coverage has no structural reason to be alarmed. Technology journalism, funded in large part by the same companies that sell the automation solutions, does not have the instruments to make visible what happens far from innovation conferences.
Trade unions, when they exist, negotiate over jobs that still exist, not over entire categories being eliminated. Governments speak of “reskilling” and “adaptation”, terms that presuppose that whoever loses a specialised technical job can retrain into something equivalent with a few months of training. That presumption has no empirical support in any previous automation cycle. What does have empirical support are the queues for unemployment benefits, the social downgradings, the early retirements disguised as voluntary departures. And in that gap, several sectors have already completed their transition without anyone having taken a collective decision about it.
What Cannot Be Undone
There are sectors in which the debate over whether AI will affect employment is anachronistic because the effect has already occurred. The professional translation of standard technical and legal documents is one. The production of digital content in volume is another. The preliminary analysis of medical images for diagnostic triage is a third. In all these cases, the economic model that sustained human employment is no longer competitive under current market conditions, and the return to that model is practically impossible without regulatory intervention that no state is seriously considering.
The breaking point is not the moment a technology appears. It is the moment when its deployment cost falls below the alternative labour cost and when the quality of its result reaches a threshold sufficient for commercial use. That threshold was reached across multiple sectors between 2023 and 2025, simultaneously and without deliberate coordination, simply because the improvement curve of foundation models was faster than projected. There was no central decision. There was no announcement. There were contracts that were not renewed, independent workers who stopped receiving commissions, posts that went unfilled when someone retired. Substitution works best when it is invisible.
The Mechanism Is Already Running
The question that habitually opens this kind of analysis is whether artificial intelligence will destroy jobs. That question supposes there is still time to respond, to regulate, to prepare. The available data suggest something more uncomfortable: that the destruction is already underway, that its speed exceeds the response capacity of any state with normal legislative processes, and that the most affected populations are precisely those with least access to the debates where the pace of that transition is decided.
What is at stake is not only employment. It is the material base of the urban middle classes in service economies, which constitute the majority of contemporary economies. When that base erodes, what follows is not a new knowledge economy accessible to all. It is a low-value service economy, a growing mass of workers deskilled by the market that trained them, and an even greater concentration of the benefits of productivity in the hands of those who already control the technological capital. The mechanism is not new. What has changed is that the speed of substitution has exceeded the speed of comprehension, and that gap is, for those who control the technological capital, the optimal condition of operation.
G.S.
Sources
- “Klarna AI assistant handles two-thirds of customer service chats”, Financial Times, February 2024
- “Klarna cuts workforce from 5,000 to 3,800 as AI takes on more work”, Reuters, September 2024
- “Harvey, the AI lawyer for Allen & Overy”, The American Lawyer, December 2023
- “The economic potential of generative AI: The next productivity frontier”, McKinsey Global Institute, June 2023
- “Future of Jobs Report 2025”, World Economic Forum, January 2025
- “Lionbridge reduces freelance pool as machine translation scales”, Slator Language Industry Report, 2024
- Automation and the Future of Work, Aaron Benanav, Verso Books, 2020
Actualizado el 19 de April de 2026
