When AI Changes the Labour Market, Who Captures the Gains?

The wrong question: will AI take our jobs?

Most occupations are bundles of tasks rather than single activities. A software engineer writes code, investigates incidents, explains trade-offs, reviews work and accepts responsibility for a release. A financial analyst collects data, tests assumptions and communicates a decision. Generative AI can perform some of these tasks quickly, while the surrounding work remains dependent on context, verification and accountability.

This is why exposure to AI is not the same as replacement by AI. The International Labour Organization's 2025 global index estimates that one in four workers is in an occupation with some degree of generative-AI exposure. Its central finding is that transformation is more likely than complete elimination because most jobs still contain tasks requiring human input. Transformation can nevertheless be disruptive when it removes the routine tasks through which junior workers traditionally learn.

What the job forecasts actually measure

The World Economic Forum's Future of Jobs Report 2025 projects 170 million roles created and 92 million roles displaced by 2030, producing a net increase of 78 million roles. This is a forecast based on employer expectations and structural trends, not a promise to an individual worker.

The net figure hides substantial churn. Creation and displacement do not necessarily occur in the same occupations, regions or skill groups. The forecast also includes technology, economic conditions, demographic change and the green transition, so it should not be presented as proof that AI alone will create or destroy a specific number of jobs.

A new data, AI or cybersecurity role is not automatically a replacement role for someone whose routine work disappeared. The relevant question is how quickly workers can move into new roles without losing income, status and access to opportunity.

• Net employment growth can coexist with large-scale occupational churn.
• New roles may require preparation that displaced workers have not had time to acquire.
• Technology forecasts should be read as scenarios, not guarantees.

The productivity paradox: more output, weaker demand?

At company level, the business case for AI is straightforward. If a team produces the same output with fewer hours, the firm may reduce costs, increase speed or serve more customers. That is a real productivity gain.

At economy level, the outcome depends on what happens to the gains. Savings can return through lower prices, higher wages for complementary workers, shorter working hours, new investment, public revenue or employee ownership. In that scenario, productivity can broaden purchasing power and create new markets.

The alternative is concentration. If displaced workers lose wage income while the gains flow mainly to owners of models, platforms, chips and data centres, the firm may be more efficient while households have less money to spend. Machines can expand supply; they do not automatically become customers. This is not an argument against automation. It is a warning not to confuse technical efficiency with a complete economic strategy.

The IMF's analysis makes the distributional issue explicit: AI may raise productivity and average income, but labour-income inequality can increase when AI complements higher-income workers more strongly than it substitutes for lower-complementarity tasks. Capital returns can also increase wealth inequality.

Three plausible labour-market scenarios

In a shared-productivity scenario, AI lowers the cost of useful services, helps small teams compete and creates demand for new products. Workers receive better tools rather than simply fewer hours, and companies invest gains in expansion, training and improved working conditions.

In a concentrated-productivity scenario, adoption is fastest in firms with capital, data and distribution. Routine and entry-level work shrinks while new opportunities require credentials, networks or geographic access. Output rises, but bargaining power weakens.

In a delayed-adaptation scenario, long-term jobs eventually appear but education, regulation and company practices move slowly. Workers are told to reskill, yet training is disconnected from actual vacancies. The transition cost is pushed onto individuals, particularly contractors and people entering the workforce.

Why ‘learn AI’ is necessary but incomplete

Tool fluency matters, but it is only one part of AI-era capability. Workers also need domain judgement, output verification, communication of uncertainty and accountability for decisions. An AI-generated answer is cheap; a decision that survives scrutiny is valuable.

OECD research on AI and changing skill demand indicates that AI-exposed occupations continue to require problem-solving, management, communication and social skills, not only specialist AI knowledge. The most durable roles may therefore be hybrid roles in which a person directs, evaluates and improves machine-assisted work.

For a remote professional, the strongest evidence is not a list of tools. It is a record of a real problem: what was owned, which constraints existed, how quality was checked and what changed as a result. As generated text becomes cheap, credible judgement becomes easier to distinguish when it is documented.

• Tool fluency: using models and workflow tools responsibly.
• Domain judgement: knowing what a correct answer should look like.
• Verification: testing outputs against data, policy and constraints.
• Communication: explaining uncertainty and trade-offs.
• Accountability: owning the decision when a model is wrong.

The entry-level problem is easy to miss

Public debate often focuses on experienced workers being replaced. A subtler risk is the loss of entry points. Junior analysts learn by cleaning data and checking routine cases. Junior developers learn through small bugs, tests and simple reviews. Junior support agents move from standard requests to complex cases.

If AI performs the routine layer, companies may retain fewer beginners and expect new hires to arrive already trained. That can improve short-term efficiency while weakening the future supply of experienced professionals. The labour market then wants senior judgement while removing the tasks through which judgement develops.

Transition policy therefore cannot consist only of online courses. It also needs paid apprenticeships, supervised project work, transparent assessment and realistic opportunities to demonstrate competence.

Remote work makes the transition wider and harsher

Remote work opens access to global employers, but it also exposes workers to global competition. When a task becomes standardised, a company can compare more candidates across borders. AI tools may help a professional participate in that market, yet the same tools can make basic deliverables easier to substitute.

The advantage will shift toward work that depends on context and trust: understanding a client's situation, protecting confidential information, coordinating across teams, validating results and making decisions when instructions are incomplete. These skills are not automatically safe from automation, but they are harder to evaluate through a generic keyword list.

A remote portfolio should show decisions, evidence and outcomes, not only polished screenshots. That makes a worker's contribution more legible when employers are comparing candidates in a wider market.

What should be measured next?

Employers should ask whether AI is removing low-value friction or simply removing people. Are productivity gains funding expansion and learning, or only reducing headcount? Are workers involved in redesigning the workflow?

Workers should ask whether a course produces a demonstrable capability. Can it lead to a supervised project, a portfolio sample or a recognised assessment? AI experience without evidence will quickly become another vague line on a résumé.

Policymakers should track wages, hours, job quality, entry-level hiring and the distribution of capital gains. A rising GDP figure can conceal a weak transition if workers lose bargaining power and households lose purchasing power.

• Wages and hours, not only employment counts.
• Entry-level hiring and apprenticeship availability.
• Job quality, autonomy and access to training.
• Where productivity gains and capital returns are going.
• Whether workers can move between declining and growing occupations.

Conclusion: the contest is over distribution and access

The evidence does not support a simple story in which AI either destroys all work or creates effortless prosperity. It points to a harder reality: AI changes tasks, bargaining power and the route into a career at the same time.

The technology can support a broad boom if its gains circulate through lower costs, new businesses, better public services, stronger wages, shorter working time and accessible pathways into new roles. It can also produce a narrow boom if productivity and ownership are concentrated while the transition cost is left to workers.

The decisive question is therefore not only what AI can do. It is who gets access to the opportunities it creates, who owns the productivity gains and who pays when the old career ladder disappears.