Artificial Intelligence and the Future of Work

The decisive unit is the task, not the occupation

The International Labour Organization therefore studies generative AI at the task level. Its updated 2025 analysis covers almost 30,000 individual tasks and classifies occupations according to how much of their task content could be affected by generative AI. Under that methodology, roughly one in four workers worldwide is employed in an occupation with at least some GenAI exposure. The ILO nevertheless concludes that most occupations are more likely to be transformed than made redundant.^[1]

That distinction matters. If an accountant delegates parts of data entry, document checking or report preparation to AI systems, accounting does not automatically disappear as an occupation. The composition of the job changes instead: less time may go into standardised processing while review, interpretation, coordination and exceptional cases take a larger share.

The same pattern is likely across many knowledge occupations. Once some activities become cheaper or faster, value shifts toward the tasks that remain scarce. The important question is therefore not only what work AI can perform, but what work remains for people and how valuable those remaining tasks become.

1. Automation does not automatically eliminate a job

Public discussion often treats automation and job loss as if they were the same process. Organisationally and historically, they are not.

A company can automate part of a job and use the released capacity to provide the same output with fewer people. It can also keep the same workforce and serve more customers, introduce additional services or perform work that was previously uneconomic because it consumed too much time.

Which outcome occurs depends on more than the technology itself. Prices, demand, competition, strategy and organisational choices all matter.

The empirical evidence available by mid-2026 reflects that mixture. An ILO review finds real productivity gains from generative AI, but with substantial variation across tasks, firms and groups of workers. At the same time, large-scale employment displacement remains limited, and self-reported time savings have not yet translated automatically into correspondingly higher measured output, earnings or employment.^[2]

AI is already changing work, but the evidence does not support a simple equation in which one automated task equals one lost job.

2. Productivity is not the same as headcount reduction

Suppose a task that takes ten hours today can be completed in six with AI. Technically, four hours of productivity have been created; economically, what happens next remains open.

A company could reduce labour requirements. It could also use the same workforce to serve more customers, perform additional analysis, update products more frequently or offer services that were previously too expensive to provide.

Productivity first changes the cost structure of an activity. Only then do firms decide how to use that new cost structure.

This helps explain why technological progress can displace existing activity while also creating new demand. If a service becomes much cheaper, more customers may buy it. Work that was once affordable only to large organisations can become accessible to smaller firms, and activities previously outsourced may become economical to perform internally.

AI therefore does more than reduce the effort required for existing work. It can also change the amount of work that is economically worth doing.

3. AI can expand what an individual is able to do

One of the more interesting effects is already visible in usage data: AI can make the boundaries between traditional occupational roles more permeable.

In 2026, OpenAI analysed more than 800,000 work-related ChatGPT messages in the United States. It found that 16.8% of all work-related messages, and 43.5% of messages that could be assigned to a specific occupation, concerned tasks usually associated with a different occupation.^[5]

A salesperson can use AI to perform a basic customer-data analysis that might once have gone to an analyst. A small-business owner can structure a first-pass financial scenario, draft marketing material or review basic contract language without immediately involving a specialist at every step.

That does not turn a salesperson into a data scientist or an entrepreneur into a lawyer. Professional responsibility and the boundary between assistance and specialist expertise remain. What changes is the range of tasks one person can reasonably attempt.

AI can therefore alter work not only by automating tasks, but by extending individual capability into areas that were previously outside a worker's normal skill or cost boundary.

4. The value of work shifts toward the remaining bottlenecks

When technology makes certain activities easier, the skills that remain scarce become relatively more valuable. With generative AI, those often involve incomplete information, responsibility for consequences, social dynamics, organisational context or choosing between several plausible options.

An AI system may compare several contract structures, but management still has to decide which risks the organisation is willing to accept. It can generate sales arguments, but interpreting whether a customer is uncertain, angry or merely undecided — and how much a long-term relationship should matter — is not always reliably reducible to text.

AI can structure medical information, compare technical alternatives or draft project plans. Responsibility for the decision does not disappear simply because some analysis became easier.

The future division of labour may therefore run less between 'human' and 'machine' than between the parts of a task that can be made predictable and those where judgement, responsibility and context remain the bottleneck.

5. Experience may become less valuable in one sense and more valuable in another

Part of professional experience consists of knowing information and performing routines faster than a beginner. AI can reduce the value of some of that advantage. A new employee can ask for process explanations, analyse documents and produce first drafts that previously required a much longer period of familiarisation.

That can narrow the gap between beginners and experienced workers in selected tasks. At the same time, another form of experience may become more valuable: recognising when a plausible answer is not good enough.

People who know a field well often recognise exceptions, unrealistic assumptions and implicit risks that do not appear in standard information. As AI makes average-quality output easier to produce, the ability to distinguish a sound result from one that merely sounds convincing may become more important.

Experience therefore need not lose importance. Its value may shift from possessing information toward evaluating information.

6. Entry-level work is especially sensitive

Many professions traditionally develop expertise through a hierarchy of tasks. Junior employees begin with comparatively standardised work, learn the field through repetition and gradually build the experience required for more complex judgement.

Those entry-level tasks are often among the easiest to automate. If AI takes over research, basic analysis, first drafts, routine coding or standard documentation, companies gain an immediate efficiency benefit.

The longer-term question is more difficult: how do experienced professionals emerge if some of the work through which beginners historically gained experience disappears?

A 2026 Anthropic survey of roughly 81,000 Claude users found that early-career respondents were more likely than senior workers to express concern about AI-driven displacement. The survey measures user perceptions rather than actual job losses, but it indicates where uncertainty is being felt particularly strongly.^[4]

At the same time, the OECD cautions against attributing current labour-market difficulties for young people too quickly to AI. Some deterioration in youth labour-market indicators began before the widespread arrival of large language models, so current evidence does not establish AI as the cause.^[7]

Both points can be true at once: today's weaker conditions for young workers need not have been caused by AI, while the automation of traditional entry-level tasks may still create an important organisational problem over time.

7. Training may have to move more deeply into the work itself

If this development continues, companies may need to rethink how expertise is built. Traditionally, competence often emerged because people completed large volumes of relatively simple work and learned patterns through repetition.

If machines take over more of that routine work, simply giving new employees access to AI will not be enough. Organisations may have to explain more explicitly why an answer is right or wrong, which criteria govern a decision and which exceptions deserve special attention.

Learning could therefore become more integrated into the workflow itself. A junior employee might process fewer similar cases from beginning to end, while spending more time analysing why an AI recommendation was accepted, modified or rejected.

That would be a different model of training: less repetition as the main learning mechanism and more explicit reflection on decisions. Whether companies make that transition well may help determine whether AI builds capability over time or merely reduces short-term labour cost.

8. The most important skill will not simply be prompting

Every technological wave temporarily creates skills that appear unusually scarce. In the early period of generative AI, one of those skills was prompt engineering.

Over time, a single prompting technique is likely to matter less than the ability to divide work sensibly between people and machines. That requires structuring a problem so that automatable components become visible, checking results, recognising uncertainty and deciding when a person needs to intervene.

An ILO report published in August 2026 describes rising demand for digital and AI-related skills alongside higher-order cognitive and socioemotional capabilities. It highlights AI literacy, adaptability, resilience and human agency among the skills becoming more important in the workplace.^[3]

That direction is intuitive. As systems take on more operational work, defining objectives, interpreting results and taking responsibility for decisions become more important.

AI literacy will therefore mean less than knowing every technical feature of a model. It will increasingly mean understanding where a system should be used, how its output should be checked and where its authority should end.

9. Organisations may change more than job descriptions

One of the largest uncertainties is that technological change does not stop at the level of the individual employee. If AI is used only to complete existing tasks somewhat faster, the organisation itself changes very little: the same report is produced, just more quickly.

Larger productivity gains often appear only when firms redesign processes themselves. A traditional workflow may contain five handoffs between departments because information historically lived in different places. If an AI system can assemble, check and prepare the same information, those handoffs may no longer be necessary.

At that point, AI becomes an organisational factor rather than merely a personal tool. Job descriptions, responsibilities and information flows begin to change.

That may be the deeper effect of the technology: not that every worker performs the same job with an AI assistant, but that companies recombine the work itself.

10. Agents move the boundary again

AI agents intensify this development. A classic assistant waits for a person to formulate a specific request, while an agentic system can, within defined boundaries, plan several steps, use tools and carry intermediate results forward.

That means automation can begin to cover not only individual tasks but chains of connected tasks. A worker no longer has to trigger every intermediate step and can instead define a goal, supervise the process and make decisions at important points.

The human role consequently shifts further from execution toward supervision. That should not be confused with complete autonomy: errors, permissions, exceptional cases and accountability remain.

In sensitive processes, people are therefore unlikely simply to disappear. They are more likely to move to different points in the workflow.

11. Not everyone will benefit equally

Technology does not distribute productivity gains automatically or evenly. An employee in a company with strong AI systems, good data and well-designed processes works under very different conditions from someone using the same models inside a poorly digitised organisation.

Countries differ as well. The ILO notes that GenAI exposure is higher in high-income economies because a larger share of employment is concentrated in clerical and knowledge-intensive occupations. Those economies also often have stronger conditions for integrating the technology productively.^[1]

The same technology can therefore produce different outcomes. In one organisation it may increase productivity and wages; in another it may reduce labour demand; elsewhere little may change because infrastructure, skills or capital are missing.

The future of work will not be determined by AI capability alone. Access to that capability, and the quality of the organisation around it, will matter just as much.

12. Forecasts of millions of jobs gained or lost should be read carefully

Few subjects generate more spectacular numbers than the future labour market.

The World Economic Forum's Future of Jobs Report 2025 estimates that a combination of technological, economic, demographic, geoeconomic and environmental macrotrends could create roughly 170 million jobs worldwide by 2030 while displacing about 92 million existing roles, producing a net increase of around 78 million jobs.^[6]

Those figures are often repeated as if they were a forecast of artificial intelligence alone. That is misleading. The report covers several macrotrends, not only AI, and a substantial part of its analysis is based on the expectations of more than one thousand large employers. It is better read as a structured assessment of possible change than as a precise labour-market prediction.^[6]

Such studies are useful for understanding expected direction and scale. They should not create the impression that we can already calculate exactly how many jobs a single technology will create or destroy by 2030.

13. The occupation of the future may become broader rather than narrower

Industrialisation and modern corporate organisation produced increasing specialisation in many fields. Complex work was divided among different occupations because no individual could master every required discipline.

AI may reverse part of that pattern. If software supports research, translation, analysis, programming, visualisation and administrative work, one person can cover a larger portion of a process.

A marketer may need less external help for basic data analysis. An entrepreneur can structure first-pass financial models. A developer can create documentation and interfaces more quickly, while an analyst can build simple automations with less specialist support.

OpenAI's data on cross-occupation AI use provides an early indication of this direction.^[5]

Rather than merely making each occupation more efficient, AI may make some roles broader. People will not become experts in every neighbouring discipline, but they may be able to work further into adjacent domains before specialised expertise becomes necessary.

14. The real competition may be between different models of work

The debate often asks whether humans or AI will be more capable. For companies, a different comparison may matter more: a person without AI, a person using AI, and a redesigned human-AI system.

The first stage is to give workers tools. The second is to teach them to use those tools well. The third begins only when processes, responsibilities and information flows are redesigned around what those tools make possible.

The largest differences may emerge there. A company that uses AI only to produce the same documents faster may gain moderate productivity. A competitor that redesigns entire workflows around the same technology may achieve very different cost structures and response times.

The future of work will therefore depend not only on which models are available, but on which organisations learn to redesign work around them.

Work does not simply disappear — its value is redistributed

The debate about AI and work is often framed between two extremes. One predicts imminent mass unemployment; the other argues that technology has historically created enough new work and therefore there is little reason for concern. Both positions miss much of the transformation taking place between those endpoints.

By 2026, large-scale employment displacement from generative AI is not yet clearly visible in the empirical evidence, while changes in tasks, productivity and work organisation are already measurable.^[2]

The technology is still developing rapidly, so today's evidence cannot simply be projected ten years forward. Much of the near-term change is likely to occur inside existing occupations: some activities become cheaper, faster or fully automatable, shifting human effort toward other parts of the process.

Some jobs will disappear, others will be created, and many may retain the same title while meaning something different in day-to-day practice. What matters is not only what AI can technically do, but how firms respond, which new services become economical, how training is organised and which tasks people continue to perform better or more responsibly.

The future of work is therefore unlikely to be defined by one moment in which AI suddenly replaces millions of people. It will emerge incrementally, each time a task is redistributed between person, machine and organisation.