For generations, the career advice seemed straightforward: get educated, learn to think, and try to make a living with your brain rather than your back. Knowledge work was the safe end of the labour market. The plumber did something useful, certainly, but the accountant, analyst, lawyer or coder had the more future-proof career.
Artificial intelligence is beginning to scramble that logic. Generative AI lives in the environment we have spent decades making machine-readable: documents, databases, spreadsheets, code, email and screens. A plumber, nurse or electrician still has to navigate an unruly physical world full of awkward spaces, unpredictable people and endless edge cases. In some settings, it may be easier to automate the first draft of a legal memo than to replace a leaking pipe behind a wall.
That does not mean the white-collar job apocalypse has arrived. The evidence so far is much more complicated. But it does suggest we may be asking the wrong question when we obsess over which jobs AI will take. The more useful question is this: what becomes valuable when intelligence gets cheap?
Capability is not the same as displacement
There is no shortage of dramatic predictions. AI companies are scenario-planning for unemployment levels that would once have sounded unthinkable. Business leaders have warned of enormous disruption. At the other extreme, optimists point to every previous technological revolution and argue that new jobs will emerge as old ones disappear.
At this point, neither camp has enough evidence to declare victory.
One of the more useful pieces of research in the material I have been studying came from Massachusetts Institute of Technology researchers who tried to move beyond conventional AI benchmarks. They identified more than 11,500 workplace tasks, ran workplace-style versions of those tasks through more than 40 AI models, and had people in the relevant occupations judge whether the output was actually usable. By 2025, AI could complete roughly 65 per cent of text-based tasks to a minimally acceptable standard, up from around half the year before.
That is extraordinary progress. It is also easy to misread. ‘Good enough’ is not the same as reliable, error-free or ready to run unsupervised inside a real organisation. Legal work, for instance, performed relatively poorly because precision, judgement and strategic guidance matter. Management tasks could often be accelerated, but coordination and decision-making remained harder.
This is the distinction that gets lost in much of the jobs debate. A technology being capable of performing a task does not mean an organisation will immediately deploy it, that customers will accept it, regulators will permit it, or that the surrounding job disappears.
Jobs are bundles of tasks. AI can remove some, accelerate others and increase the value of the remainder. We may therefore see a great deal of task displacement before we see equivalent job displacement.
The safest jobs may not be where we thought
For years, automation anxiety focused heavily on factories, warehouses and transport. Robots were coming for physical labour first, or so the story went.
Physical automation is certainly advancing. Humanoid robots can now run, kick footballs and perform eye-catching demonstrations. Yet the World Humanoid Robot Games also test machines on wonderfully mundane activities such as connecting cables, handling warehouse goods and completing industrial assembly. The reason is revealing: the physical world is messy.
Molly Kinder at The Brookings Institution has described this as the ‘messy middle’. Hospitals, restaurants, construction sites, airports and nursing homes require dexterity, situational awareness, trust and constant adjustment to circumstances nobody anticipated in advance. Cognitive work conducted inside a computer is often much more structured.
There is an irony here. The economy spent decades telling young people that manual work was vulnerable and knowledge work was secure. AI may make that distinction considerably less useful.
NVIDIA chief Jensen Huang has even argued that the AI infrastructure boom will create huge demand for electricians, plumbers, carpenters and construction workers. Someone still has to build the data centres, factories, power systems and physical infrastructure behind all this supposedly immaterial intelligence.
None of this makes trades immune from automation. It simply reminds us that technological exposure is not distributed according to old assumptions about occupational status.
When abundance arrives, scarcity moves
Economics has a habit of rewarding scarcity. For much of modern working life, competent intellectual output was scarce enough to command a premium. Research took time. A polished first draft took skill. Code required specialists. Analysis depended on people who had spent years learning a field. AI is beginning to make some of those outputs abundant.
Organisational psychologist Adam Grant recently put the creative version of this idea neatly: when ideas become abundant, taste and tenacity matter more. Good judgement and execution become scarce. I think the same principle extends far beyond creativity.
Researchers Roberto Rigobon and Isabella Loaiza at MIT Sloan School of Management have developed a useful framework for the human capabilities that remain particularly complementary to AI. They call it EPOCH: Empathy and emotional intelligence; Presence, networking and connectedness; Opinion, judgement and ethics; Creativity and imagination; and Hope, vision and leadership.
What makes their work interesting is not merely the list. In an analysis spanning 19,000 tasks across 950 occupations, they found that newly emerging workplace tasks showed higher levels of these human-intensive capabilities than tasks that had disappeared. In other words, work may already be evolving towards some of the things humans remain unusually good at.
This leads to what I think of as the Scarcity Shift. Whenever technology makes a capability abundant, look for where scarcity moves.
When answers become abundant, good questions appreciate. When content becomes abundant, taste appreciates. When synthetic communication becomes abundant, trust appreciates. When competent analysis becomes abundant, judgement appreciates. When execution gets easier, deciding what deserves to be executed becomes more important.
The point is not that these qualities are uniquely or permanently human. It is that their relative economic value can rise as other capabilities become easier to access.
The apprenticeship paradox
There is one part of this transition that worries me more than most: entry-level work.
Junior roles have always contained a fair amount of low-value activity. Graduates research, summarise, prepare first drafts, build presentations, check documents and perform routine analysis. Organisations tolerate the inefficiency partly because this is how people learn. You do the basic work, watch more experienced colleagues, make mistakes, develop judgement and gradually earn responsibility for harder decisions. AI is becoming very good at the bottom rungs of that ladder.
Randstad’s Gen Z Workplace Blueprint, drawing on more than 11,000 workers and an analysis of 126 million job postings globally, found postings for roles requiring zero to two years’ experience had fallen 29 percentage points since January 2024. Technology, logistics and finance recorded particularly large falls. Randstad does not attribute that decline solely to AI, and neither should we. Economic conditions and hiring cycles matter too. But the direction deserves attention.
It creates an apprenticeship paradox: if AI removes the work beginners used to do, how do beginners become experts?
Hybrid work adds another wrinkle. Research on remote work increasingly describes the office not simply as a place to complete tasks but as social infrastructure: a place where younger workers overhear conversations, observe how experienced colleagues handle ambiguity, build networks and absorb tacit knowledge that is difficult to put in a manual.
We could therefore end up removing junior tasks at the same time as we weaken some of the informal mechanisms through which juniors traditionally learned to move beyond them.
That is not an argument for dragging everybody back to a desk five days a week. It is an argument for redesigning early-career development deliberately. If AI does the first draft, a graduate may need earlier exposure to clients, decisions, critique and judgement rather than simply being given less work.
AI fluency will be baseline, not the differentiator
There is a danger in all this talk about human skills. It can quickly turn into comforting advice to focus on empathy and leave the technology to somebody else. That would be a mistake.
Randstad‘s 2026 Workmonitor surveyed 27,000 workers and 1,225 employers across 35 markets and analysed more than three million job postings. Vacancies asking for ‘AI agent’ skills had surged 1,587 per cent. Governments are responding too. The UK has announced an ambition to provide practical AI skills training to 10 million workers by 2030.
AI fluency is rapidly becoming part of basic professional literacy. The advantage will not come from choosing between technical and human capability. It will come from combining them.
The person who can use AI well and exercise sound judgement over its output is more useful than the person who does only one of those things. The manager who can automate routine analysis but still read a room has an advantage. The marketer who can generate 100 ideas in minutes but has the taste to discard 97 of them has an advantage. The adviser who can retrieve an answer instantly but knows which question the client is really asking has an advantage.
Being human is not, by itself, a career strategy. Human capabilities have to be developed with the same seriousness we once reserved for technical expertise.
Organisations will move more slowly than the technology
Another reason to resist precise predictions about mass unemployment is that organisations are stubborn things.
AI models can improve in months. Companies do not. They have legacy systems, budgets, risk committees, regulators, customers, employment laws, managers, incentives and habits. A technically possible automation can take years to become an operational reality.
Australian middle-market research from Pitcher Partners illustrates the gap. In its 2025 Business Radar survey, 72 per cent of leaders said they were actively using AI tools, yet only 13 per cent had made AI a genuine strategic priority with dedicated budgets and plans to scale. Adoption is widespread; transformation is not. That gap buys people time. But it would be foolish to confuse slow organisational change with low eventual impact.
Walmart offers a more plausible picture of the next phase than either utopia or apocalypse. CEO Doug McMillon has said AI will change essentially every job in the company. Walmart has been automating warehouse and back-of-store work while creating roles such as ‘agent builders’ and continuing to need people in high-touch roles. The company has indicated its overall workforce could remain broadly flat while the business grows, meaning the composition of work changes even when total employment does not collapse.
That may be the pattern to watch: not jobs vanishing in neat occupational blocks, but organisations progressively unbundling work and redistributing tasks between people and machines.
The repricing of human capability
Nobody can responsibly tell you today whether AI will eventually create more jobs than it destroys. The range of credible possibilities is simply too wide. Anthropic is openly scenario-planning for everything from modest disruption to a world in which demand for human labour is persistently lower. Other researchers point out that we have not yet seen anything resembling that level of displacement in aggregate labour-market data.
Both facts can be true. The job apocalypse is not here, and preparing for significant disruption is still sensible.
What we can see more clearly is a repricing already underway. Some cognitive tasks that once required scarce expertise are becoming cheaper. Some physical skills look more resilient than expected. Early-career pathways need redesign. AI literacy is becoming baseline. And capabilities such as judgement, trust, taste, curiosity, leadership and human connection are becoming more important precisely because information and competent output are becoming easier to produce.
For leaders, that suggests a useful exercise. Don’t only ask which jobs AI could automate. Ask three questions instead:
- What just became cheap?
- What remains scarce?
- What becomes more valuable because of the change?
Those questions will produce different answers in a law firm, hospital, mine, school, retailer or construction company. They will also change as the technology improves.
That is why I suspect the future of work will be less about protecting a fixed list of ‘human jobs’ and more about continually noticing where value has moved.
The future of work won’t simply be shaped by what machines learn to do. It will be shaped by what remains valuable once they can.
Michael McQueen is a globally recognised trend forecaster, change strategist and keynote speaker.
A bestselling author of 10 books, Michael’s latest release was named by Malcolm Gladwell and Adam Grant as one of the top five must-read new leadership books. He is a sought-after media commentator, with his insights regularly featured in Forbes, The Guardian, and CNN.
To find out more about Michael and his work, click here.
NOTES
MIT Sloan School of Management. “Choose the Human Path for AI.” December 16, 2025. https://mitsloan.mit.edu/ideas-made-to-matter/choose-human-path-ai.
Axios. “MIT Study Challenges AI Job Apocalypse Narrative.” April 2, 2026. Reporting on MIT Computer Science and Artificial Intelligence Laboratory research testing more than 11,500 workplace tasks across more than 40 AI models.
Randstad. The Gen Z Workplace Blueprint: Future-Focused, Fast-Moving. 2025. Based on a survey of 11,250 workers globally and analysis of more than 126 million job postings worldwide.
Reuters. “Young Workers Most Worried About AI Affecting Jobs, Randstad Survey Shows.” January 20, 2026. Reporting on Randstad Workmonitor research covering 27,000 workers, 1,225 employers and more than three million job postings across 35 markets.
Pitcher Partners. Business Radar 2025: Understanding the Businesses That Drive Australia’s Economy. October 2025.
UK Government. “Government Expands Free AI Training for 10M Workers.” January 28, 2026. Announcement of expanded AI foundations training and workforce initiatives.
Rigobon, Roberto, and Isabella Loaiza. EPOCH research on human capabilities and artificial intelligence, as summarised by MIT Sloan School of Management in “Choose the Human Path for AI,” December 16, 2025.
The Guardian. “‘We Are Teaching the Machine to Take Our Job’: Banks Bet Big on AI as Thousands of Jobs Cut.” September 9, 2025. Used as contextual reporting on the difficulty of separating AI-driven displacement from broader organisational restructuring.
