Artificial intelligence is already shifting economic growth - but the benefits aren't landing everywhere, or evenly, for every worker. At the IMF and World Bank annual meetings in Bangkok, IMF Managing Director Kristalina Georgieva described AI as a "slow-moving tsunami" that could eventually affect about 40% of jobs worldwide. In advanced economies, the figure could be closer to 60%.
That does not mean 40% of jobs will vanish. Georgieva's point is about exposure: AI can change what people do, reduce demand for some roles, and raise productivity for workers who stay employed. Which outcome shows up in practice depends on the job itself, the technology, and what employers and governments do next.
The early concern is for younger workers. Entry-level roles often bundle routine research, drafting, analysis, and administrative work - tasks that AI tools can now assist with. At the same time, AI-related investment is supporting growth and boosting hardware-producing economies. Below, I unpack what the IMF's figures really imply, why early-career jobs are often the first to feel it, and what businesses can do without treating automation like a workforce plan.
What the IMF's 40% figure actually means
The headline number is easy to misread. Georgieva said the IMF expects AI to eventually touch about 40% of jobs worldwide. The estimate rises to around 60% in advanced economies and is roughly 23% in low-income countries.
But "touch" (or "exposure") doesn't automatically mean "eliminate." A job can be exposed because AI can take on part of the task list, because workers use AI to complete work faster, or because the role changes shape. A bookkeeper using software to produce routine reports is exposed to automation, for example - yet that alone doesn't tell you whether the employer cuts staff, expands the service, or shifts the bookkeeper toward exception handling and client support.
That distinction matters for business decisions. A share of exposed jobs is not a forecast of layoffs at the same rate. It's a signal about the scale of possible change, not a precise prediction of when it happens or who is hit first. Real results will vary by occupation, the firm, the regulatory environment, and how reliable the tools are in day-to-day use.
Exposure is about tasks, not job titles
Most jobs mix different types of work. A single role might include repetitive data entry, judgment calls, customer communication, and responsibility for the final outcome. AI may handle some of those parts well while being poorly suited for others.
That's why statements like "AI will replace accountants" or "AI can't replace teachers" don't help much when you're trying to plan hiring. The better question is simpler: which tasks in this role can the system do reliably, which ones still require human judgment, and who is accountable when something goes wrong?
In real projects, a task-by-task view is also where implementation costs show up. A tool may draft a report quickly, but someone still needs to check sources, handle sensitive information, fix errors, and fit the output into existing workflows. If review and exception handling take longer than the original task, the "productivity gain" may be more illusion than improvement.
Why advanced economies show greater exposure
The IMF's estimates point to higher exposure in advanced economies partly because AI tools are more available and many jobs can be supported by digital work. The lower number for low-income countries shouldn't be read as immunity. Georgieva said poorer countries score lower mainly because AI hasn't reached them to the same extent.
Of course, adoption depends on more than whether a model exists. It depends on internet access, electricity reliability, relevant data, language support, skills, and whether employers can pay for and integrate new systems. Countries with weaker infrastructure might see less immediate task automation - but they may also miss out on productivity gains and new business opportunities.
So the IMF's figures describe both technological exposure and unequal access. They're a reason to plan for different timelines - not a reason to assume workers and economies with lower exposure have nothing at stake.
Why young workers may feel the change first
Georgieva warned that young people entering the workforce are likely to be hit first. AI can absorb routine tasks that often sit inside entry-level jobs. At the same panel, Singapore President Tharman Shanmugaratnam said demand for graduates' first jobs is already falling.
The issue isn't that every graduate role is made of tasks AI can do. It's that early-career work often includes structured, repeatable tasks that are easier to delegate: preparing a first draft, assembling background research, summarizing documents, checking data, or answering standard questions. Those are also the first tasks companies tend to hand off to AI.
If an employer automates those activities without redesigning the role, it may reduce the number of junior openings. That creates a second problem over time: fewer people get the supervised experience that turns newcomers into experienced analysts, engineers, marketers, lawyers, or managers.
The entry-level work pipeline can break quietly
A company might get a short-term boost by reducing repetitive work. The less visible downside is a weaker training pipeline. Senior staff still need people who understand the details, can investigate anomalies, and can take responsibility for decisions. Those skills don't appear just because a team added an AI assistant.
This trade-off is familiar. Removing routine work can improve efficiency, but routine work is often where new employees learn the business, its customers, and how far the company's processes actually go. If that work disappears, employers have to replace the learning path - through supervised projects, simulations, rotations, or more structured mentoring.
There's another risk for younger workers too: expectations can rise faster than capability. If a graduate is given an AI tool and expected to produce expert-level work without review, the company hasn't removed the need for experience. It's just pushed the risk onto someone less experienced - and it can make mistakes harder to catch.
Employers should redesign jobs, not just reduce headcount
A sensible workforce plan starts with the tasks in a role. One practical way to map work is to sort tasks into three buckets:
- Automate: repetitive work where the process is stable and mistakes are easy to detect.
- Augment: work where AI can produce a draft or recommendation, but a person should review it.
- Keep human-led: decisions involving real accountability, sensitive relationships, or context the system can't reliably assess.
This isn't meant to be permanent. Tasks can move between groups as tools improve and business needs change. The goal is to make the decision explicit, test it in an actual workflow, and track not only speed but quality.
For early-career roles, the same exercise should include a tough question: what replaces the learning value of any task you automate? If a junior employee no longer prepares the first analysis, they may need a different way to learn how to question assumptions and spot a weak result. Productivity gains tend to last longer when the next generation is still building the skills the business will need later.
AI's economic gains are unevenly distributed
The IMF's warning sits next to a more optimistic growth picture. Georgieva said AI investment is helping keep global growth near 3% a year despite an energy shock. The IMF's July outlook projected global growth of 3% in 2026, down from 3.5% in 2025, with higher energy prices - linked to the war in the Middle East - weighing on the forecast.
The benefits aren't spread evenly. Economies that build the hardware behind AI are among the early winners. The IMF pointed to Taiwan, South Korea, Thailand, and Malaysia as major net exporters of AI hardware. In its July update, those countries exceeded first-quarter growth forecasts by an average of 4.4% on an annualized basis. South Korea, specifically, grew at a 7.5% pace compared with the 1.8% the Fund projected in April.
Georgieva also said AI adoption could add 0.5 to 1 percentage point a year to growth in advanced economies and 0.2 to 0.8 points in emerging economies. For context, a percentage point is the difference between a 2% growth rate and a 3% growth rate. She said the high end of the range could more than offset the drag from aging populations.
Infrastructure investment is not the same as broad prosperity
AI-related hardware and technology now make up more than a tenth of global goods trade, according to Georgieva's remarks in Singapore. She also said AI investment as a share of global GDP could surpass the investment that went into railroads or the electricity grid.
That scale helps explain the momentum. It doesn't automatically tell you who captures the returns. Hardware exporters can benefit from demand for chips and related products, and firms that own valuable technology may benefit from productivity improvements. But workers whose tasks get automated may see weaker bargaining power or fewer opportunities - especially if income shifts from labor toward owners of capital.
There's also a public-finance angle. Georgieva warned that governments rely heavily on taxes tied to workers' incomes. If more value flows to capital owners instead, the tax base can weaken even while the economy keeps growing. In that sense, how gains are distributed isn't a side issue - it affects how governments fund education, infrastructure, and support for people changing jobs.
Lower exposure can mean lower access, too
The IMF's estimates also suggest a gap between countries that can deploy AI and those that can't. A country with limited connectivity or computing infrastructure may see less immediate job exposure. But those same limits can also make it harder for local businesses, schools, clinics, and public services to use AI in practical ways.
World Bank President Ajay Banga said the bank is backing "small AI" tools that don't depend on large language models, with potential uses in farming, clinics, and classrooms in emerging markets. The point here is straightforward: "useful AI" doesn't always require the biggest model. A smaller tool designed around a specific task may be cheaper to run, easier to deploy, and better suited to local constraints.
So the real choice isn't just whether a country adopts AI. It's whether adoption supports locally useful services and skills - or concentrates value among a small set of companies and hardware-producing regions.
What businesses and policymakers can do now
There's no one-size-fits-all response. A bank, a small retailer, a hospital, and a software company have different risks, data needs, and tolerance for errors. Still, a similar sequence helps keep decisions tied to day-to-day reality.
For businesses: measure the workflow before buying the tool
Start with the work, not the vendor's demo. Pick a process with a clear baseline: how long it takes, what quality looks like, how often it fails, and how much review it requires. Then test whether AI improves the whole process - not just the time needed to produce a first draft.
A useful pilot should track at least four things:
- Time: Does the workflow take less total staff time, including review and corrections?
- Quality: Are outputs accurate, complete, and consistent with the required standard?
- Risk: Does the process expose personal, confidential, or regulated information?
- Human ownership: Is it clear who checks the result and who is accountable for acting on it?
A tool that saves a few minutes but creates a bigger review burden - or a fuzzy failure path - can be technically impressive and still be a bad operational decision. Expand only if the pilot shows value under normal conditions, including awkward inputs and edge cases.
Companies should also think about how benefits are shared. If productivity rises, does it translate into better service, more output, shorter backlogs, or improved working conditions? If the only measurable outcome is fewer entry-level roles, the company may be weakening its own pipeline of future talent.
For workers: build skills around judgment and verification
Workers don't have to become AI researchers to stay valuable in an AI-enabled workplace. They do need to understand where the tools help, where they fall short, and how to verify what the tools produce. Domain knowledge matters even more when systems generate answers that sound plausible but are incomplete or wrong.
Useful skills include asking precise questions, checking claims against trusted sources, handling exceptions, explaining decisions, and communicating with people who have different needs. None of this guarantees that jobs won't change. But it makes it easier to move from routine production toward work that involves context and responsibility.
For graduates and early-career workers, experience still matters. A portfolio of real work - internships, supervised projects, and the ability to explain how a result was checked - can show competence that goes beyond familiarity with a tool. Employers, for their part, shouldn't treat "knowing the tool" as a stand-in for training people in the actual job.
For policymakers: make access and transition part of the plan
Governments have multiple linked problems to address: access to infrastructure, education and retraining, worker support during transitions, and the tax implications of shifting income from labor to capital. Treating these as separate debates often produces policies that fix one part of the system while another part fails.
Training should connect to real demand, not just be offered as a generic response to technological change. That means working with employers and education providers to identify which tasks are changing, which skills people need, and how workers can demonstrate they've learned them. It also means investing in basic access so smaller businesses and less-connected communities aren't locked out of useful applications.
The IMF's full World Economic Outlook is scheduled for release on October 13, its first update to the 3% global growth forecast since July. The forecast will give a broader economic picture, but it won't settle the central labor-market question: how much of AI's productivity gain turns into better work - and how much turns into fewer opportunities for people entering the labor market.
Conclusion
The IMF's warning is best read as a planning signal - not a countdown to mass unemployment. AI is already influencing investment and growth, but what it means for work depends on how companies redesign tasks, how governments support access and transitions, and whether workers share in productivity gains.
For businesses, the next step is practical: map one real workflow, test AI against a measurable baseline, and decide what happens to the tasks - and learning opportunities - that the tool changes. For workers and policymakers, the priority is building skills and access without assuming every job will shift in the same way.
The technology will keep moving. The quality of the transition depends on human decisions. Read the original report from Decrypt and use the IMF's upcoming outlook to reassess the economic context, not to replace job-level analysis.