Exposed Is Not Replaced: ILO's 80-Million-Worker ASEAN Study

Nearly 80 million workers across Southeast Asia hold jobs that generative AI could technically do parts of. Employment in exactly those jobs has risen every year since 2017 — from 66 million to 80 million. The ILO published both numbers in the same brief last month, and the space between them is where the honest version of the AI-and-jobs story actually lives.

The brief is Generative AI and labour markets in ASEAN: Significant exposure, limited disruption, uneven preparedness, released 8 July 2026 by the ILO's Regional Office for Asia and the Pacific. Japan's JILPT flagged it for the Japanese labour policy community the same week. The title is doing a lot of work, and it deserves to be read literally — all three clauses, in order.

What the 22.9 percent actually counts

[Fact] 22.9% of total ASEAN employment — roughly 80 million people — works in occupations with more than a minimal degree of potential exposure to generative AI, based on ILO estimates for 2025.

Here is the part that gets lost when this number travels. Exposure measures the technical feasibility of automating or assisting tasks inside an occupation. It is not a forecast. It is not a redundancy notice. It is a statement about what the technology could touch, said before anyone knows whether employers will touch it.

The gradient underneath the headline matters more than the headline. [Fact] Low exposure accounts for 9.1% of employment, moderate 8.2%, significant 2.2%, and the highest band — what the ILO calls G4 — just 3.3%, or 11.7 million workers. Meanwhile about 67% of ASEAN employment sits in occupations with no identified exposure at all.

Run those against each other and you get a figure the brief does not print: of the 22.9% who are exposed in any degree, only about one in seven — 14.4% — falls into the top band. The other six are in categories where AI is more plausibly a tool than a substitute.

A 21-fold spread across nine countries

Country-level exposure diverges sharply. [Fact] Singapore records 42.2% of employment with more than minimal exposure, reflecting an occupational structure where professionals, managers, executives and technicians are roughly 64% of employed residents. The Philippines follows at 28.1%, then Indonesia at 21.7%, Viet Nam at 20.8% and Thailand at 20.6%.

In the highest band the spread is wider still: from 0.3% in Timor-Leste to 6.3% in Singapore. That is a 21-fold difference inside a single regional bloc.

One honest caveat, and the ILO raises it first: the methodology does not adjust for differences in digital infrastructure across countries. A task that is technically automatable in a Bangkok office and technically automatable in a rural district with unreliable connectivity are not the same fact. Read the country numbers as ceilings, not as conditions.

Eight years of data that decline to cooperate with the panic

If generative AI were already displacing exposed workers at scale, the exposed occupations should be shrinking. They are not.

[Fact] Employment in occupations with low, moderate, significant or high exposure grew from about 66 million in 2017 (20.9% of employment) to 74 million in 2022 (22.2%) to 80 million in 2025 (22.9%). Between 2017 and 2022, the fastest employment growth was in the highest exposure band. From 2022 to 2025 that growth moderated while the significant-exposure band accelerated. Occupations with no exposure grew most slowly throughout.

The reverse claim is equally unsupported, and the ILO says so plainly: none of this shows AI creating those jobs. It is consistent with service-sector expansion, structural transformation and labour force growth — forces that were running long before ChatGPT shipped.

The paradox: highest exposure, lowest actual use

This is the finding I did not expect, and it complicates the standard narrative more than anything else in the brief.

Using Anthropic's Claude.ai conversation data from February 2026, the ILO found usage concentrated in computer and mathematical occupations, followed by educational instruction. [Fact] Office and administrative support — the occupational group that dominates the high-exposure list — showed comparatively limited observed use. The share of Claude.ai use directed at work purposes ranged from 29.4% in Indonesia to 52.0% in Viet Nam.

Singapore's firm-level data tells the same story from the employer side. A Ministry of Manpower establishment survey covering 2,560 private sector establishments and 486,600 workers, fielded January to March 2026, found that [Fact] 71.5% of firms had not begun adopting AI in their operations at all, and only 3.8% had integrated it into core business processes. Adoption ran to 76.4% among firms with more than 500 employees but only 23.9% among firms with fewer than 25. By sector: information and communication 74.1%, professional services 57.5%, financial and insurance 56.4% — and administrative and support services just 21.1%.

So the sector employing the most highly exposed clerical workers is the sector adopting AI the least. Exposure and adoption are pointing in opposite directions, and adoption is the one that pays wages.

Where women land

[Fact] Across ASEAN, 4.8% of employed women work in the highest exposure band, compared with 2.3% of men — more than double. In Thailand and the Philippines the gap widens to three-to-four times.

The mechanism is occupational segregation, not anything about the technology. Women are concentrated in clerical, administrative and selected professional roles; men are more concentrated in manual and operational work with lower exposure. If you want to see the specific job families driving this, the ILO's high-exposure clerical list maps onto roles like data entry keyers, bookkeeping clerks, payroll clerks, word processors and human resources assistants.

Exposure pays better, which cuts against the usual story

Here is the reversal. The ILO looked at wages by exposure level and found that occupations in the significant-exposure band were consistently the highest paid in the sample. Average hourly earnings rose with exposure in Cambodia, Indonesia, the Philippines and Viet Nam; the pattern held in Thailand but less cleanly.

The occupations sitting there are accountants, personal financial advisors, software developers and economists. Exposure in ASEAN is not a marker of precarity. It is closer to a marker of the knowledge economy — which is why the ILO's policy worry is distributional: that the gains land on people who are already well paid.

The one signal worth watching

Youth and adult exposure are broadly comparable region-wide, but not everywhere. [Fact] In Indonesia 4.0% of youth employment is in the highest band versus 3.2% of adults; in the Philippines it is 4.6% versus 3.9%.

More interesting is what happened next. In the Philippines and Thailand, youth employment in highly exposed occupations fell faster than overall employment between 2022 and 2024, with sharp declines in highly exposed clerical roles specifically. Tertiary-educated young people in exposed sectors also fared worse than adults — a degree did not neutralize it.

And yet the aggregate youth indicators held. [Fact] Philippine youth unemployment stayed flat at 6.9% from 2022 to 2023 while the NEET rate fell from 13% to 12.4%; Thai youth unemployment fell from 5.1% in 2022 to 4.6% in 2024, with NEET down from 13.3% to 12.8%.

[Claim] The ILO reads this as reallocation — shifting hiring patterns, different education choices, changed entry pathways — rather than a broad deterioration in young people's prospects. That reading is plausible and it is also unproven. Entry-level change showing up before aggregate change is exactly what the "canaries in the coal mine" hypothesis predicts, and two countries over two years cannot settle it. This is the number to check again next year.

Preparedness, not exposure, is the variable

[Fact] The IMF's AI Preparedness Index scores the ASEAN-5 at an average of 0.60 against 0.68 for advanced economies. The ILO sorts the region into three tiers: Singapore alone at the top; Malaysia and Thailand, then Brunei Darussalam, Indonesia, the Philippines and Viet Nam holding the foundations with real gaps in advanced skills, research capacity and compute; and Cambodia, Lao PDR, Myanmar and Timor-Leste further back.

Exposure and preparedness do not move together, and the mismatches are where the risk concentrates. The Philippines ranks second on exposure but fourth on preparedness. Myanmar carries one of the lowest preparedness scores in the sample alongside higher exposure than Cambodia, Viet Nam or Lao PDR. Singapore is the only economy that is highly exposed and highly prepared — which is a much better place to be exposed from.

What this means if you are the one doing the job

The most useful thing in this brief is a distinction, not a number. Exposure tells you where a technology could plausibly rearrange tasks. Adoption tells you whether anyone is actually doing it. Preparedness tells you what happens to you when they do. Only the first is currently large in ASEAN; the second is small and lopsided toward big firms in knowledge sectors; the third is a policy choice countries are making right now.

For workers in administrative assistant and clerical roles, the practical read is that you are in the most exposed category and the least-adopting sector at the same time. That is not safety, but it is time — and the Singapore data suggests the firms most likely to move first are the large ones in information, professional and financial services. Learn the tools before the tools arrive with a reorganization attached. The ILO's own framing is that GenAI is "a tool to be mastered rather than a solution to all," and the workers who master it early are the ones the productivity gains get shared with.

Sources

• ILO, Generative AI and labour markets in ASEAN: Significant exposure, limited disruption, uneven preparedness, ILO Brief, Geneva, 8 July 2026. DOI 10.54394/00034834. Prepared by Christian Viegelahn, Phu Huynh, Makiko Matsumoto, Jong Eun Oh, Diego Rei and Felix Weidenkaff. Licensed CC BY 4.0.
• JILPT (Japan Institute for Labour Policy and Training), ASEAN地域の労働者8,000万人が生成AI技術の影響を受ける可能性（ILO：2026年7月）, 海外労働情報, 8 July 2026.

Figures in this article were read directly from the ILO brief PDF, including its annexes and boxes; the JILPT summary was used as a cross-check. Exposure estimates draw on the methodology of Gmyrek et al., ILO Working Paper No. 140 (2025). Singapore firm-level figures originate with Singapore's Ministry of Manpower, "Adoption of Artificial Intelligence Among Firms," April 2026, as reported in the ILO brief.

AI-assisted analysis: this article was drafted with AI assistance from primary source documents and reviewed before publication. The ILO brief itself carries a comparable disclosure — its authors used ChatGPT for editorial assistance and validated the final content.