AI and the Future of Work

What the best evidence says — and does not say — about AI and jobs: why “exposed” is not the same as “replaced”, what controlled experiments found, the argument over entry-level work, where Malaysia stands, and what you and your employer can do.

A customer-service agent in Petaling Jaya starts her shift. A year ago she searched three knowledge bases to answer a billing question; now a suggested reply appears beside each chat, and her job is to judge whether it is right, fix it, and handle the angry customer the software cannot calm. Is her job safer or more at risk than it was? The honest answer is that nobody yet knows for certain — but the evidence gathered since 2023 lets us say much more than the headlines do. This page separates what has been measured from what is being predicted.

Exposure is not replacement

Most big numbers about AI and jobs are measures of exposure: researchers break each occupation into its tasks and ask how many of those tasks overlap with what AI can currently do. A highly exposed job is one that is likely to change. Whether it shrinks, grows or simply shifts depends on whether AI does the tasks instead of the worker or helps the worker do them.

The best-known estimate comes from the International Monetary Fund. In January 2024 an IMF staff discussion note found that almost 40% of global employment is exposed to AI. Exposure rises with income, because office work is more common in richer economies:

Share of jobs exposed to AI. Source: IMF, Gen-AI: Artificial Intelligence and the Future of Work (January 2024).

The International Labour Organization (ILO) uses a stricter, task-by-task method. Its 2023 global study concluded that generative AI is more likely to augment jobs than to automate them, with clerical work the clear exception. Its refined index, published with Poland’s research institute NASK in May 2025, put about one job in four worldwide in occupations with some exposure to generative AI, but only 3.3% in the highest-exposure category. That top category was larger for women (4.7%) than for men (2.4%), mainly because women hold more clerical and administrative jobs.

Consensus Among labour economists the common ground is this: AI changes tasks within jobs far more often than it removes whole occupations, and exposure figures are a map of where change is likely, not a count of jobs that will be lost.

What the experiments show

Exposure studies are estimates. Since 2023, a handful of controlled studies have measured what actually happens when real workers get generative AI.

Brynjolfsson, Li and Raymond studied more than 5,000 customer-support agents at a software firm as an AI assistant was rolled out in stages. In the published version (Quarterly Journal of Economics, 2025), agents with the assistant resolved about 15% more issues per hour on average. The gains were concentrated among newer and less-skilled agents — over 30% in the 2023 working paper — while the most experienced agents gained little. Customer sentiment in the chats improved, and fewer agents quit.

This guide’s framing Taken together, the experiments point to three patterns. AI helps most on tasks with clear, checkable outputs; it tends to lift less experienced workers most, at least at first; and it harms performance when people use it for tasks it does badly without noticing. That last pattern is why the habits in Checking AI’s Work are a job skill, not a nicety.

Two caveats matter. These are short-run studies of particular firms and tasks; productivity in one call centre does not tell us how many call-centre jobs there will be in ten years. And productivity gains can cut either way for workers: a firm can use them to serve more customers with the same staff, or the same customers with fewer.

Which tasks change first

Across the studies, the tasks most affected early share a profile: they are done on a screen, mostly with language or code, follow recognisable patterns, and have outputs someone can check. Drafting routine letters, summarising documents, first-line customer queries, translation, simple coding, data entry and scheduling all fit. Tasks that change more slowly involve physical presence and dexterity (nursing, plumbing, construction), building trust and relationships (sales, counselling, teaching young children), judgement under real accountability (signing off an audit, a diagnosis, a legal opinion), and coordinating people.

This is a modern version of Moravec’s paradox, described on Lessons from History: what is hard for people (writing a polished report) can be easy for machines, while what is easy for people (folding laundry, reading a room) remains hard. It also means the effect lands unevenly within a single job. A lawyer may lose most of her document-review hours while her court and client work grows more central.

The argument about entry-level work

If AI is best at the routine tasks that used to be given to beginners, what happens to beginners? This is now one of the most closely watched questions in labour economics.

In August 2025 Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford, using payroll records from the largest US payroll provider, reported that employment of 22- to 25-year-olds in the most AI-exposed occupations, such as software development and customer service, had fallen about 13% relative to less-exposed jobs since late 2022, while older workers in the same occupations were unaffected (a November 2025 revision put the figure at 16%). The decline was concentrated in jobs where AI tends to automate rather than augment. In contrast, an October 2025 analysis by the Yale Budget Lab and Brookings found no discernible economy-wide disruption to the US labour market in the 33 months after ChatGPT’s release, noting that shifts in the mix of occupations had begun before generative AI.

Debated Whether AI is already shrinking the first rung of the career ladder is genuinely unsettled, and the evidence so far is mostly from the United States.

Other views
  • An early warning: the Stanford authors and others argue young workers are the “canaries in the coal mine”: firms can stop hiring juniors long before they lay anyone off, so this is where effects would show first.
  • Other causes: sceptics point to higher interest rates, the post-pandemic correction in tech hiring and offshoring, which hit the same jobs at the same time and are hard to separate from AI.
  • A changed ladder, not a missing one: others argue that if juniors with AI can do more advanced work sooner — as the call-centre study suggests — entry-level roles may be redesigned rather than removed, though that depends on employers choosing to train.

Speculative All long-range numbers — how many jobs AI will create or destroy by 2030 or 2040 — are forecasts. Past automation forecasts have been wrong in both directions (see why predictions fail), and the popular claim that AI will take every job is examined on Myths and Misconceptions.

Malaysia’s labour market

The most detailed Malaysian evidence so far comes from TalentCorp, the talent agency under the Ministry of Human Resources. Its impact study of AI, digitalisation and the green economy, released in November 2024, covered about 3.5 million employees in ten sectors that together account for roughly 60% of GDP, including electrical and electronics, global business services, ICT and wholesale and retail. It found:

A second phase covering further sectors, including healthcare, tourism, construction and the public sector, was announced for 2025. Note what these figures are: employer- and expert-informed assessments of impact on jobs, which include change as well as loss — not a count of layoffs.

Malaysia also has particular features that shape how AI arrives. A large share of employment is in small and medium-sized enterprises, which adopt new technology more slowly than big firms; a substantial informal and gig workforce has fewer employer-provided training routes; and a sizeable global business services and shared-services industry does exactly the kind of screen-based, language-heavy work that is most exposed. On the other side, government programmes have made basic AI literacy free: the AI untuk Rakyat self-learning course, launched in January 2024 with Intel and offered in Bahasa Malaysia, English, Tamil and Mandarin, reported over a million completions within six months As of 2026. The wider national picture is on AI in Malaysia.

What you can do

No one can promise a particular job is safe. But the experiments point clearly to the kinds of worker who benefit: those who know their field well enough to judge AI’s output and who use it on the right tasks. A practical ladder:

  1. Map your own tasksList what you actually do in a typical week. Mark which tasks are screen-based, patterned and checkable — those are where change will come first.
  2. Try AI on those tasks, safelyUse the tools your employer allows, with no confidential or personal data, following your organisation’s acceptable-use policy (see AI in the Workplace).
  3. Get good at checkingThe consultants study shows the cost of trusting AI outside its frontier. Being the person who catches its mistakes is valuable.
  4. Deepen what AI cannot do for youDomain judgement, relationships with clients and colleagues, physical skills and accountability become more valuable as routine tasks shrink.
  5. Keep learning on purposeUpskilling works best when tied to a real task. In Malaysia, check HRD Corp-claimable courses through your employer and free government programmes.

What employers and governments owe

Individual effort only goes so far; the shape of change depends heavily on choices made by firms and states. Here the disagreement is about values as much as evidence.

Other views
  • Employers should share the gains: trade unions and many labour researchers argue that if AI raises productivity, workers should share it through training, redeployment before redundancy, a say in how tools are introduced and, where hours fall, protected pay.
  • Markets adjust best: many economists and business groups argue that firms must be free to adopt productivity-raising tools, that new jobs historically appear in ways no one predicts, and that heavy rules would push investment elsewhere.
  • Strengthen the safety net: others focus on government: portable training accounts, unemployment insurance (Malaysia’s Employment Insurance System under PERKESO is one example), support for older workers, and, at the radical end, proposals such as universal basic income, which remain highly contested.

Whatever one’s view, two points draw wide support: training needs to be easier to reach for people outside big companies, and workers who will use AI tools should be involved in deciding how they are introduced.

Quick check

The IMF says about 40% of jobs worldwide are “exposed” to AI. What does that mean?

Exposure measures overlap with AI capabilities. The IMF itself suggests roughly half of exposed jobs in advanced economies may benefit.

In the call-centre study, who gained most from the AI assistant?

The assistant spread the know-how of top agents to newcomers; experienced agents gained little.

What happened when consultants used AI on a task outside its “frontier”?

They were 19 percentage points less likely to be correct — a reminder to check AI’s work.

// Pause and reflect

  1. Which three tasks in your work (or the work you hope to do) are most exposed? Which are least?
  2. If AI made your team 20% more productive, what would you want your employer to do with that gain?
  3. Where will the next person starting in your field learn the routine skills that AI now handles?

Your notes are saved privately in this browser.

Key terms on this page

Sources & further reading