AI and work in Asia

In January 2026 a government wrote this into a national AI framework: as agents take over entry-level tasks, which typically serve as the training ground for new staff, organisations risk losing the basic operational knowledge those tasks used to build. That is Singapore's Infocomm Media Development Authority, not a think tank, and the closest thing to official corroboration of the apprenticeship argument found anywhere in this research.

Singapore names the problem, in the terms this research uses

IMDA's Model AI Governance Framework for Agentic AI, version 1.0 of 22 January 2026, was read in full at source. Section 2.4.3 states:

"As agents take over entry level tasks, which typically serve as the training ground for new staff, this could lead to loss of basic operational knowledge for the users. Organisations should identify core capabilities of each job and provide sufficient training and work exposure so that users retain foundational skills."

A few pages earlier the framework warns of "the potential loss of trade craft", and requires that "sufficient training, especially in areas where agents are prevalent, must be provided to ensure that humans retain core skills".

It goes further than most frameworks on oversight too. It names automation bias directly, describing it as "the tendency to over-trust a system that has performed reliably in the past". It requires that overseers be trained to identify failure modes such as inconsistent agent reasoning. And it requires that the effectiveness of the oversight itself be audited, which is a step almost nobody takes.

It is also honest about the limit. The executive summary concedes that "continuous human oversight over all agent workflows becomes impractical at scale", which is the concession most governance documents avoid making. The argument about what oversight can actually bear is at human in the loop is not a safeguard.

Two qualifications. It is guidance rather than statute. And it states a risk and a duty to train, with no threshold for what counts as retaining core skills and no test of whether the training works.

The sharper finding is the twenty months before it

Singapore published a Model AI Governance Framework for Generative AI on 30 May 2024. Both published versions of that document were read and searched at source. They contain zero occurrences of "human oversight", "human-in-the-loop", "over-reliance", "automation bias" or "deskill". Human oversight is not among its nine dimensions. Its only use of the word competency concerns third-party auditors rather than the person doing the overseeing.

Twenty months later, the same issuing body built a pillar around oversight and added deskilling to it. That shift is documented, dated and quotable. It is better evidence of how fast this problem became visible to policymakers than any comparison between countries would be.

What Singapore measures, and what the number actually says

Singapore's Ministry of Manpower reported in June 2026 that 28.5 per cent of firms had adopted AI, rising to 74.1 per cent in information and communications, 57.5 per cent in professional services and 56.4 per cent in financial and insurance services.

The more useful pair of numbers sits underneath. Only 6.2 per cent of firms reported AI-related reductions in headcount or hiring, against 18.9 per cent reporting redesign of job functions. The ministry's own conclusion: AI is "having a greater impact on job redesign and work processes than on broad-based job displacement".

Redesign running roughly three times ahead of displacement is a useful corrective to forecasting that treats job losses as the headline. It is also not reassuring on its own, because redesign is not neutral. The question this research asks is which tasks the redesign removes, and a survey counting headcount cannot answer it. That is the gap IMDA's own framework points at.

Hong Kong: the contrast is sharp

Hong Kong has published no consolidated territory-wide AI strategy. That is not an inference. The government was asked in the Legislative Council in October 2025 to map out strategies and set phased targets, and again in April 2026 to formulate a comprehensive AI development blueprint. Neither reply announced a document. The April 2026 answer was that initiatives "are ongoing and being consolidated". The operative overarching document remains a general innovation and technology blueprint from December 2022, in which AI is one of three strategic industries.

On measurement the position is cleaner still. The Census and Statistics Department's business survey of IT usage, released 27 February 2026, was read in full including all thirty tables. Artificial intelligence appears nowhere in it. The survey measures cloud computing at 98.1 per cent, QR codes at 37.6 per cent, RFID at 20.3 per cent, internet of things at 7.4 per cent, and augmented or virtual reality at 1.5 per cent.

Hong Kong's statistical office measures AR and VR adoption at one and a half per cent of firms and does not ask about AI at all. Three companion publications, including the household IT survey and the flagship information society compendium, contain no AI content either. The consequence for anyone quoting a Hong Kong AI adoption figure is that it came from a private survey with a self-selected sample.

Hong Kong's privacy regulator has done better work than the government has. Its 2024 model framework requires that personnel exercising oversight "remain aware of the tendency to over-rely on the output produced by AI", and states that human oversight should not be "merely a gesture". But deskilling appears in none of the five Hong Kong instruments examined. The territory addresses whether the human starts competent and whether they over-trust the machine in the moment. It does not address whether capability degrades through sustained use.

Japan and Korea: the adoption gap, measured twice

Japan's Institute for Labour Policy and Training surveyed 22,000 employees, stratified on the 2020 Census with the OECD involved in the design. It found 12.9 per cent reporting any AI use by their employer and 8.4 per cent using it themselves. Adoption is a fraction of what the discourse implies, in the country most often described as behind.

The Korea Development Institute put the two halves of the question side by side, and the gap is the finding. It scored 38.8 per cent of Korean jobs as technically automatable across more than 70 per cent of their tasks. Its survey of 800 firms with ten or more staff found actual adoption at 2.7 per cent. Almost every widely circulated number in this field measures the first quantity and gets discussed as though it described the second.

Korea's realised effects are the ones worth carrying into a boardroom: no aggregate employment change, lower earnings, and the impact concentrated on younger, tertiary-educated workers and women. France's national statistics office found the same shape independently. Two countries, two methods, one silhouette. It runs opposite to what twenty years of automation commentary trained everyone to expect.

India, and what this page does not claim about it

No verified Indian official statistic on AI adoption or AI-related employment has been checked at source for this research, so none is quoted here. That is a gap rather than a finding, and it will be closed rather than filled with numbers that circulate without provenance.

What can be said is structural. India's exposure runs through IT services and business process work, which is the sector where France measured the sharpest fall in employment of under-thirties and where Korea's effects concentrated on the young and educated. If the pattern found in three countries holds, India is among the most exposed labour markets in the world to precisely the mechanism this research describes, and the least measured.

What a leadership team in the region can take from this

• Singapore's framework is a usable internal test wherever you operate. Ask its question of your own deployment: which entry-level tasks were the training ground, and what replaced them.
• Adoption numbers are not capability numbers. Singapore's 28.5 per cent, Japan's 12.9 per cent and Korea's 2.7 per cent all measure use. None measures whether anyone could still do the work unaided, which is the test at dependency.
• Redesign is where the risk lives. Nearly three times as many Singaporean firms redesigned jobs as cut them. Redesign decides which repetitions survive, at missed reps.
• Check the provenance of any regional AI statistic. One of the four markets here publishes a government AI adoption figure. The others do not, and the numbers circulating for them are private surveys.
• Audit the oversight, not just the output. IMDA requires it. Almost no organisation does it.

Key research and primary sources

• Infocomm Media Development Authority (2026). Model AI Governance Framework for Agentic AI, Version 1.0. Singapore, 22 January 2026.
• AI Verify Foundation and IMDA (2024). Model AI Governance Framework for Generative AI. Singapore, 30 May 2024.
• Ministry of Manpower, Singapore (2026). Labour Market Report, First Quarter 2026. 15 June 2026.
• Census and Statistics Department, Hong Kong (2026). Survey on Information Technology Usage and Penetration in the Business Sector, 2025 Edition. 27 February 2026.

Related SuperSkills research

For the wider international position, AI and work by country, Japan and the Gulf. For the mechanism Singapore names, the missing rungs and capability debt. For the category, human capability in the age of AI. For measurement, measuring adoption properly.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. He has presented across Asia, including in Singapore, Hong Kong, China and India, and built EtonX operations in China and in India. Every document cited here was read at the issuing body's own page on 28 August 2026. Claims that could not be verified at source, including the reported version 1.5 of the IMDA agentic framework, are excluded rather than softened.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.