Let’s begin with generative Artificial Intelligence (GenAI) – a technology that has firmly embedded itself in our daily lives over the past three years and is already reshaping the world of work. According to the latest report from the International Labour Organisation (ILO, 2025), around a quarter of jobs worldwide – more than 600 million roles – are potentially exposed to the effects of generative AI.
A joint study on Latin America by the ILO and the World Bank estimates that between 26 and 38 per cent of jobs in the region – around 88 million roles – could be affected by GenAI over the coming years. These projections include both the partial transformation of existing tasks, as well as the complete replacement of roles in sectors such as administration, communications, software, manufacturing, and finance.
But beyond generative AI and its well-documented impacts, we are now encountering – and, to a large extent, being sold on – a new wave of systems that, if they deliver on their promises, could accelerate large-scale job displacement even further. These so-called agentic AI systems and AI agents demand close and critical scrutiny.
AI agents are algorithmic systems characterised by a larger degree of autonomy and independence in decision-making than traditional AI assistants – such as large language models (LLMs) – although they are still designed to carry out tasks aligned with pre-defined goals.
Agentic AI refers to the orchestrated interconnection of various systems – whether agents or assistants such as chatbots, robotics, or other automated platforms. This integrated approach enables such systems to adapt in order to carry out complex tasks, such as keeping a factory operational with minimal human oversight, or managing logistics and supply chains in sectors like industry or agriculture.
While the debate around AI and employment has gained traction in international organisations and multilateral forums, a significant gap remains: there has been little serious discussion about autonomous AI – its various forms, associated risks, and potential.
It’s worth noting that the evolution of these algorithmic systems is also driving the expansion of generative AI, which, as we’ve seen, is already reshaping the service sector. However, without a clear understanding of how these new forms of artificial intelligence are being developed and deployed, it remains difficult to assess whether they will significantly amplify the impact of automation on employment – or if many are, in reality, still limited prototypes, far from achieving the full autonomy so often touted in Silicon Valley narratives.
From AI agents to agentic AI
As previously noted, an AI agent is defined as an autonomous system – either physical or purely digital – capable of perceiving its environment, processing information, making decisions, and taking action to achieve a given objective. Unlike AI assistants, such as chatbots, agents are equipped with real-time learning and adaptive capabilities, which in theory allow them to operate dynamically in complex environments, gradually improving their performance over time.
Examples include AutoGPT, AgentGPT, BabyAGI and CrewAI – systems that claim to be able to research a topic, consult sources, write articles, and adapt to new instructions, all without direct human intervention.
Agentic AI, meanwhile, is described as a step forward in both autonomy and complexity – capable of receiving a broad objective, devising a strategy, breaking it down into concrete tasks, and coordinating other agents or systems (including robotic ones) to carry them out.
Within this spectrum, tools like Claude (developed by Anthropic) and Manus sit somewhere in the middle. While Claude can operate as an assistant or content generator, it also has the potential to be integrated into more complex architectures as part of an agentic system. Manus, designed for collaborative workflows, can assume agent-like roles when used to coordinate tasks across a network.
In essence, agentic AI can be understood as an overarching architecture designed to tackle complex problems autonomously. Operating within this framework are AI agents – individual software units tasked with carrying out specific functions, each with its own goals and resources. Put differently, agentic AI provides the structural foundation for autonomy, while AI agents act as the building blocks which, through interaction, contribute to achieving larger, more complex objectives.
Between the promises of autonomy and the processes of automation
Autonomy is one of the most heavily promoted features of both AI agents and agentic AI – with the prevailing message being that these advances will allow systems to make decisions entirely independently.
However, the reality today is that such decisions still rely heavily on highly specific programming frameworks – and the risks of errors, biases, and misinterpretations remain significant. (This article doesn’t even touch on other important concerns, such as psychosocial risks, security, or privacy.)
A closer look at how existing systems are being applied – regardless of their specific type – reveals a clear gap between promise and practice. Recent experiments suggest that, rather than seamlessly reorganising production processes without human input, both AI agents and agentic systems often fall into cycles of inefficiency.
A recent study by researchers at Carnegie Mellon University and Stanford’s Institute for Human-Centred Artificial Intelligence created a fictional company, TheAgentCompany, staffed entirely by AI systems built on language models similar to GPT-4. Each system was assigned a specific organisational role, working together to simulate the development of a new software product.
Although the AI agents and assistants showed some ability to organise and communicate, their collaboration was ultimately inefficient – marked by repeated tasks, drifting objectives, and a lack of strategic alignment – ultimately preventing the launch of a functional product.
The experiment descended into an unproductive cycle of meetings with no tangible outcomes, mirroring the very same coordination challenges faced by real-world companies – a dynamic that was, in the end, all too human.
AI-powered robotics reshaping industry and logistics
The integration of AI systems with robotics is being rapidly adopted across the industrial and logistics sectors. This trend is accelerating the automation of production processes and, in turn, having tangible effects on the world of work.
In China, for instance, we are witnessing the rise “dark factories” – facilities designed to optimise automation processes in key sectors such as electronics and electric vehicles. Foxconn, the world’s largest iPhone manufacturer, plans to automate 90 per cent of its assembly operations, while companies like Haier, Midea and Siemens already operate factories run entirely by robots and AI. They even promote their ability to manufacture phones without any human involvement – as if it were something to celebrate.
UPS – the American company behind one of the world’s largest distribution networks – introduced artificial intelligence systems at one of its logistics facilities. Designed to optimise delivery routes, implement dynamic pricing, and manage loads, these new systems led to the elimination of 20,000 jobs in 2025, along with the closure of 73 facilities worldwide.
Salesforce CEO Marc Benioff confirmed that his company cut 4,000 customer service positions – reducing the team from 9,000 to around 5,000 – following the integration of AI agents now handling roughly 50 per cent of customer interactions.
Autodesk, a San Francisco-based software company, also announced a wave of layoffs in 2025, cutting nearly 1,350 jobs – around 9 per cent of its global workforce. The company justified the move as part of a “larger restructuring aimed at strengthening its AI-driven products and digital platforms”.
Another example comes from the Indian firm Tata Consultancy Services (TCS), which announced in July the elimination of 12,261 jobs – its largest workforce reduction to date. The company attributed the cuts in part to disruptions driven by artificial intelligence and automation, as well as shifts in technology service delivery models.
Amazon, which operates distribution centres in Brazilian cities such as São Paulo and Betim, has implemented intelligent robots and algorithmic systems for sorting and inventory management. This led to an estimated 10 per cent reduction in staffing between 2022 and 2025 at its most highly automated facilities.
Finally, in the United Kingdom, Ocado cut 500 jobs across its technology and finance divisions following the introduction of AI systems at its automated warehouses.
Technological acceleration and the growing urgency for collective response
Although agentic AI remains in its early stages and riddled with limitations, the prevailing narratives of its imminent impact are already influencing investment flows and reshaping how work is organised.
In other words, even though it has yet to fully materialise, the future of work is already being shaped as though it had.
That’s why, rather than accepting the narrative of inevitability – which so often goes hand in hand with technological determinism – the future of work will depend on the ability of trade unions, social movements, governments, and multilateral institutions to critically assess how much of what is promised about AI agents and agentic AI is, in fact, part of a narrative pushed by Big Tech to attract speculative investment.
At the same time, it’s crucial not to lose sight of the tangible and already visible risks these technologies pose – particularly when it comes to the accelerating pace of automation and its impact on the world of work.