Automation has a way of turning every conversation about the future into a binary one: either machines replace people, or they don’t. The more provocative versions of that story claim that AI will eventually erase most human labour and leave society with a strange new class of “artisans” and “hipsters”—people who survive by doing the last bits of work that can’t be automated, or by selling authenticity in a world where everything else is generated on demand.
It’s an appealing narrative because it’s simple. But it’s also misleading in a specific way: it treats “automation” as if it were a single force that moves in one direction—toward eliminating human tasks—rather than a set of technologies that can substitute, complement, reorganise, and sometimes even expand the demand for human effort. There is good reason to be dubious about the idea that automation will supplant all demand for human labour. The more accurate question is not whether humans will disappear from the economy, but how AI will change what humans do, where they do it, and which parts of the value chain become more valuable.
The first thing to understand is that labour markets don’t respond to “technology” in the abstract. They respond to tasks, workflows, incentives, and business models. When AI arrives, it rarely replaces an entire job in one clean sweep. Instead, it attacks particular components: the routine part of decision-making, the repetitive part of documentation, the pattern-recognition part of triage, the drafting part of writing, the translation part of communication, the monitoring part of operations. Those are often the easiest targets because they are measurable and scalable. But once those pieces are automated, the remaining work doesn’t vanish. It changes shape.
Consider what happens when a company deploys AI to handle customer support tickets. The obvious outcome is fewer tickets handled by humans. But the less obvious outcome is that the company now needs humans for escalation handling, policy exceptions, relationship management, and quality assurance. In other words, the work shifts from “answering” to “deciding,” from “producing” to “verifying,” from “doing” to “owning outcomes.” That shift can reduce headcount in some roles while increasing demand in others. It can also raise the bar for the remaining positions, making them harder to fill and more expensive to staff.
This is why the “total replacement” framing tends to overstate the speed and completeness of substitution. AI can be astonishingly capable at generating text, images, code, and predictions. Yet capability is not the same as economic adoption. Businesses adopt technology when it improves margins, reduces risk, or unlocks new revenue. And those improvements often depend on human judgement, human accountability, and human relationships—especially in sectors where errors are costly or where customers care about trust.
There’s also a second reason to be sceptical: productivity gains don’t automatically translate into fewer workers. Sometimes they do. But sometimes they create new demand. When something becomes cheaper, faster, or more accessible, consumption expands. That expansion can absorb displaced labour, at least partially. The classic example is how automation in manufacturing reduced the cost of goods and increased overall production, which then required logistics, maintenance, sales, and new kinds of service work. AI could follow a similar pattern, though the mechanism may be less visible because the “goods” are often information services rather than physical products.
If AI makes marketing content cheaper, companies may produce more campaigns, test more variants, and target more niches. If AI makes software development faster, startups may build features sooner and launch more products. If AI makes legal research quicker, firms may take on more cases or offer new pricing models. In each case, the demand curve shifts. Even if some tasks are automated, the total volume of work can rise enough to keep humans employed—though not necessarily in the same occupations or at the same wages.
The third reason the hipster-artisan story doesn’t fully land is that it assumes the only remaining human work is “creative” or “authentic.” But many of the tasks that remain after automation are not romantic. They are administrative, operational, compliance-related, and interpersonal. They involve coordinating across systems, managing exceptions, handling ambiguity, and taking responsibility when the model is wrong. These are not the kinds of tasks that make for a charming cultural image. They are the kinds of tasks that keep organisations running.
AI systems are powerful, but they are not omniscient. They can hallucinate, misinterpret context, fail silently, or produce outputs that look plausible while being incorrect. In high-stakes environments—healthcare, finance, legal, critical infrastructure—humans remain essential not because the work is inherently human, but because accountability is inherently human. Someone has to sign off. Someone has to explain. Someone has to bear the consequences. That someone is usually a person with authority, training, and liability.
Even in lower-stakes settings, businesses often prefer human oversight because it reduces reputational risk. A company can tolerate occasional errors in a draft. It struggles to tolerate errors in a final decision that affects a customer’s money, safety, or access. As a result, AI adoption frequently leads to a new layer of human review rather than a clean elimination of human labour. That review layer can be smaller than before, but it can also become more specialised and more expensive.
Then there is the question of “customisation.” The hipster-artisan narrative implies that people will survive by making bespoke products. But customisation is not always a luxury; it can be a necessity. Many industries require tailored solutions because customers have different constraints, regulations, and preferences. AI can help generate custom outputs quickly, but it still requires humans to define requirements, interpret constraints, and ensure the output fits the real world. In practice, the most valuable human work may be the work of translating between business goals and technical capabilities.
This is where the unique take emerges: AI may not reduce human labour so much as it changes the unit of work. Instead of paying for hours of manual production, organisations increasingly pay for outcomes, throughput, and risk-managed decisions. That shift favours people who can orchestrate systems—people who understand both the domain and the toolchain. It also favours organisations that can integrate AI into existing processes without breaking them.
In other words, the future may be less about everyone becoming an artisan and more about everyone becoming a manager of complexity. Not necessarily a manager in title, but a manager in function: someone who supervises AI outputs, corrects them, and decides what to do next. That kind of work is cognitively demanding. It requires judgement, domain knowledge, and the ability to detect when a model is drifting away from reality.
The labour market impact, therefore, is likely to be uneven. Some roles will shrink dramatically. Others will grow. Many will be transformed. The most vulnerable workers are those whose tasks are both routine and tightly coupled to a specific workflow—jobs where the output is easily measured and where there is little room for discretion. The most resilient workers are those whose tasks involve discretion, relationship-building, or accountability, or those who can combine domain expertise with AI-assisted production.
But resilience is not the same as safety. Even roles that remain can change quickly. A job that used to require a certain level of expertise may become easier to perform with AI tools, compressing wages and reducing bargaining power. Meanwhile, the jobs that require higher-level oversight and system design may become more scarce and better paid. This creates a polarisation dynamic: fewer high-skill roles, fewer low-skill roles, and a shrinking middle—though the exact shape depends on policy, education, and how quickly firms restructure.
Education and training are central here, and they’re often discussed too vaguely. The key issue is not whether people can learn to use AI. Most can. The key issue is whether education systems can teach the skills that matter in an AI-augmented economy: critical thinking, domain understanding, data literacy, and the ability to evaluate outputs. People need to know how to ask good questions, how to verify results, and how to understand failure modes. They also need to learn how to work with tools rather than simply operate them.
There is also a structural issue: AI adoption can outpace the ability of labour markets to adjust. Even if new jobs are created, they may appear in different locations, require different credentials, or demand different schedules. That mismatch can produce unemployment or underemployment even in a growing economy. The hipster-artisan story glosses over this friction by imagining a smooth transition into new forms of work. Real transitions are messy. They involve retraining, relocation, and time. They also involve social safety nets—or their absence.
Policy will shape how painful the transition is. If governments treat AI-driven productivity as a reason to cut taxes and deregulate, the benefits may concentrate quickly while adjustment costs fall on workers. If governments treat AI as a reason to invest in training, wage insurance, and labour mobility, the transition can be less brutal. If governments impose strong rules around transparency, auditing, and accountability, they may slow some adoption but also preserve trust and create compliance-related employment. None of these outcomes are guaranteed. But they are decisive.
Another factor often overlooked is that AI can increase the demand for human labour in unexpected ways. When AI automates a task, it can also increase the number of tasks that organisations attempt. For example, if AI makes it easy to generate drafts, companies may produce more documents, more versions, more experiments, and more iterations. That increases the need for editing, governance, and archiving. It also increases the need for people who can manage information quality and ensure that outputs align with brand and policy.
Similarly, AI can create new categories of work around data stewardship. Models are only as good as the data they are trained on and the data they are fed. Organisations need people to curate datasets, monitor drift, manage consent, and ensure compliance with privacy laws. These are not glamorous jobs, but they are essential. They can grow as AI usage grows.
There is also the “