# The '88% of Garment Jobs at Risk' Statistic Is From 2016 — and It Says Otherwise

> The claim that 88–90% of garment jobs will be automated traces to one ILO working paper from July 2016, which explicitly states it does not predict job numbers or timing. What the paper says, and what the ten years since actually show.

**Source:** [https://scanerp.pro/blog/garment-jobs-automation-88-percent-statistic.html](https://scanerp.pro/blog/garment-jobs-automation-88-percent-statistic.html)

---

# The '88% of Garment Jobs at Risk' Statistic Is From 2016 — and It Says Otherwise

Santosh Rijal · September 16, 2026 · 9 min read · Automation

**TL;DR — Direct Answer:**
The figure traces to **one ILO working paper from July 2016** that applied 2013 Frey-Osborne probabilities to labour force survey data. The paper states in its own text that it **"does not attempt to predict the precise number of jobs that will be automated or displaced, nor does it identify an exact year."** It is a competent technical study that was turned into a headline it explicitly disclaims. Ten years on, Cambodia — the country with the highest risk score at 88 per cent — hit **record exports of $15.5bn in 2025** with over 900,000 workers in the sector. Where displacement did occur, the deepest measured cut was in **cutting rooms, not sewing lines**, and the cause was CNC machinery rather than AI.

You have seen this number. Nearly 90 per cent of garment workers in Cambodia face losing their jobs to automation. Between 64 and 88 per cent of garment jobs across South East Asia are at risk. It appears in NGO campaigns, conference keynotes, sourcing consultancy decks, vendor sales material and a great many news articles, usually with a photograph of a sewing floor and the word "robots" in the headline.

I own a garment factory, so this statistic is about my workers and, indirectly, about me. A few years of hearing it quoted without a source eventually sent me to find the source. It took some digging, because almost nobody citing the number cites the document. When I found it and read it, the most striking thing was not that the number was wrong. It is not wrong. It is that the paper containing it says, clearly and in its own words, that it is not the kind of number people are using it as.

## Where the number actually comes from

The source is *ASEAN in Transformation: The Future of Jobs at Risk of Automation*, by Jae-Hee Chang and Phu Huynh, published as Working Paper No. 9 by the International Labour Office's Bureau for Employers' Activities in **July 2016** ([ILO, PDF](https://www.ilo.org/sites/default/files/wcmsp5/groups/public/@ed_dialogue/@act_emp/documents/publication/wcms_579554.pdf)). Everything downstream — the 88 per cent, the 90 per cent, the half a million Cambodian sewing operators — comes from this one document or from something quoting it.

The method is a desk study, and the paper is open about it. Frey and Osborne's 2013 Oxford work evaluated 702 occupations from the US Department of Labor's O*NET database and calculated a probability of computerisation for each, concluding that around 47 per cent of total US employment sat in the high-risk band. Chang and Huynh write that their brief "takes an identical approach to assess the automation of occupations in ASEAN, deploying the automation probabilities presented in the original Frey and Osborne study and applying them to the labour force survey data of Cambodia, Indonesia, the Philippines, Thailand and Viet Nam."

The garment figure appears in a passage about manufacturing subsectors. The paper reports that the share of garment wage workers at high risk ranges "from 64 per cent in Indonesia to 88 per cent in Cambodia," and elsewhere notes that in Cambodia "close to half a million sewing machine operators face a high automation risk." Those are the sentences that became a decade of headlines.

Two features of the method matter and are almost always dropped in the retelling. The probabilities are American occupational probabilities, derived from US task descriptions, mapped onto Asian labour force data. And the unit being scored is an occupation's technical susceptibility to computerisation, not a prediction that any particular employer will buy any particular machine.

## What the paper actually says about prediction

This is the passage that should appear alongside every citation of the 88 per cent and almost never does. It sits in the methodology section, immediately after the description of the approach:

> "Lastly, the methodology does not attempt to predict the precise number of jobs that will be automated or displaced, nor does it identify an exact year when this will happen. Instead, it identifies the occupations and types of workers facing a high probability of automation over the next couple decades based on the nature of tasks involved. Some studies contend that the future of work will be less about technology displacing humans and more about a complementarity between humans and technology in the workplace. Even in this scenario, the workplace will undergo significant technological changes and many conventional jobs will be redefined."

That is the authors telling readers, in advance, exactly how their figure would go on to be misused. It is a careful sentence, written by people who understood the limits of what they had built.

I want to be precise about where the fault lies, because the easy version of this article would be an attack on the ILO and it would be wrong. **The study is not bad work and the authors did not overclaim.** They stated their method, published their caveats, and put the disclaimer in the body of the paper rather than in a footnote. What happened next happened downstream of them: a probability band became a prediction, a prediction acquired a deadline it never had, and the caveat did not travel with the number. That is a failure of citation, not of research.

The paper also flags a qualifier that gets lost even more thoroughly than the disclaimer. It notes that "social factors specific to ASEAN businesses can offset the substitution effects of automation," and it raises the question of automation's arrival as a policy question — asking, for instance, what would happen to half a million Cambodian sewing machine operators "if automation is rapidly applied," a conditional that reads very differently from the way it is usually quoted.

## The ten-year scoreboard

A 2016 paper about the next couple of decades is now a decade old, which means part of the period it describes has actually happened. We can look.

Cambodia is the fair test, because it carried the highest risk score in the study and because the sector dominates its economy. According to a January 2026 report from the Ministry of Commerce, Cambodia exported garments, shoes and travel goods worth a combined **15.5 billion US dollars in 2025, a year-on-year increase of 15.7 per cent**. Garments alone accounted for 11.4 billion dollars, up 16.5 per cent. Data from the Ministry of Labour and Vocational Training cited in the same report puts the industry at **more than 1,500 factories and branch operations employing over 900,000 workers**, most of them women ([Cambodia Investment Review](https://cambodiainvestmentreview.com/2026/01/23/cambodias-garment-footwear-and-travel-goods-exports-hit-15-5-billion-in-2025-as-regional-trade-drives-growth/)). Industry reporting from a 2026 sector summit describes exports at close to 16 billion dollars in 2025 with continued diversification into footwear and travel goods ([Sourcing Journal](https://wwd.com/sourcing-journal/industry-news/cambodia-un-least-developed-country-garment-footwear-travel-goods-manufacturing-1239026815/)).

A decade of commentary said this workforce would be automated away. Exports and headcount are both at or near records instead.

One honest qualification, because the opposite overclaim is also available and I do not want to make it. Record exports and high employment say nothing about whether those jobs are good jobs. Wages, working hours, heat stress and bargaining coverage in Cambodia are all contested, and the same summit coverage above is largely about exactly those disputes. My claim here is narrow: the volume of garment work did not collapse. Whether it improved is a different argument, and not one that export figures can settle.

## What did displace work, and where

None of this means automation displaced nobody. It did, and the best field evidence available says so clearly enough that it deserves reporting in full rather than in summary.

A study titled *Assessment of Technological Transition in the Apparel Sector of Bangladesh and Its Impact on Workers*, conducted by Solidaridad Bangladesh, the Labour Foundation and BRAC University, with fieldwork between August and October 2024, found a **30.58 per cent reduction in the total workforce across production processes, with helpers the most affected group**. The research used mixed methods: a survey of 429 workers in Dhaka, Gazipur and Narayanganj, key informant interviews with 26 stakeholders, and four focus group discussions ([The Business Standard](https://www.tbsnews.net/economy/rmg/automation-rmg-sector-led-3058-decline-workforce-study-1024546); further coverage at [FashionUnited](https://fashionunited.uk/news/business/study-how-does-automation-affect-workers-in-bangladeshs-rmg-sector/2025012479773)).

**Read the denominator before you quote that number.** These are reductions measured per production process and per production line within the factories surveyed, from a sample of 429 workers. They are not a fall in national employment. Bangladesh's ready-made garment workforce has not shrunk by 30 per cent, and any sentence implying that it has is making the same mistake with this study that the industry made with the ILO paper. I am flagging this because it is an easy misreading and because I would rather this article be quoted correctly than quoted often.

With the denominator understood, the internal breakdown is the genuinely useful finding. The same study reports that **automation in cutting led to a 48.34 per cent drop in workforce, while sewing recorded a 26.57 per cent decline**. Sweater factories saw the highest decline per production line at 37.03 per cent, woven factories 27.23 per cent.

Sit with the ordering. The deepest measured cut landed on the cutting room, which nobody writes about. The shallowest landed on sewing, which is the entire subject of the automation debate.

That matches what I see. I run a cutting room, and I can name the roles an automatic spreader and a computer-controlled cutter remove: the people who walked the fabric back and forth laying plies, the hands that held the marker, the operator pushing a straight knife through a lay, and some of the helpers who tied and moved bundles afterwards. That work is now a machine and one or two trained people. It is real displacement and it fell mostly on helpers — lower paid, often women, often older or with less schooling, exactly the profile the ILO paper predicted would be most exposed.

But the technology that did it is a CNC cutter and an automatic spreader. That is mechanisation. It is computer-controlled machinery that has been commercially available for decades, and it contains no machine learning whatsoever. Filing it under "AI is taking garment jobs" gets both the cause and the location wrong, which is why the debate keeps aiming at the wrong process. I have written separately on [what AI genuinely runs on a sewing floor in 2026](https://scanerp.pro/blog/ai-garment-manufacturing-sewing-floor-reality.html), and on why [the flagship robotic sewing factory now runs with human sewers](https://scanerp.pro/blog/sewbot-factory-what-happened.html). The practical cutting-room side is covered in our guide to [cutting room management and fabric waste](https://scanerp.pro/blog/cutting-room-management-reduce-fabric-waste.html).

## What a factory owner actually feels today

Here is the part that makes the whole debate feel strange from inside a factory. The pressure I feel is not a surplus of operators about to be displaced by machines. It is the opposite: finding and keeping skilled operators is hard, and it has been getting harder.

Nepal has had a generation of working-age people leave for jobs abroad, and I compete for staff against that option as much as against the factory down the road. A trained machinist who can hold quality across a range of operations is the scarcest input I buy, ahead of machines and well ahead of software. When I plan capacity, the binding constraint is people, not capital. I do not think my position is unusual; the same summit coverage from Cambodia is full of skills-framework and workforce-progression initiatives, which is not what an industry about to shed 88 per cent of its workforce spends its time on.

So for planning purposes, over a realistic horizon of the next several years, I would treat it this way. Plan on people sewing your garments, and plan on it being hard to hire them. Spend on cutting-room automation, where the payback is demonstrable and the technology is mature. Spend on retaining operators, because replacing them costs more than the arithmetic usually admits. And keep a record of what your floor actually does, operation by operation, because whatever automation eventually arrives will need to be justified against a baseline — and a factory that cannot say what its current cost per operation is cannot evaluate any machine that claims to beat it. There is more on that planning question in our guide to [production planning and capacity](https://scanerp.pro/blog/garment-production-planning-capacity-guide.html).

## What would change the picture

I am not arguing that garment work is permanently safe, and the 2016 paper does not argue that it is doomed. The defensible claim from the last decade is about timing and mechanism, not permanence. Three developments would genuinely move the assessment.

A robotic system that holds a general sewing line — mixed styles, mixed sizes, changing fabrics, real shade-lot variation — rather than flat high-volume goods. That has not happened, despite serious money and serious engineering aimed at it.

A sustained rise in wages relative to capital cost. This is what actually drives automation decisions, and it is why nearshoring markets adopt before low-wage ones. Anyone modelling automation risk should be modelling wage trajectories, not occupational probabilities.

And structural demand shifts — tariffs, trade preferences, nearshoring — which change where garments are made without necessarily changing how. Those have moved more garment jobs in the last decade than any robot has, and they get a fraction of the attention.

None of those is the story the 88 per cent tells. That number was a careful answer to a narrow technical question, asked in 2016 and answered honestly, including about its own limits. It deserves to be cited the way it was written.

## Frequently Asked Questions

### Where does the 88 per cent garment jobs automation statistic come from?

From ASEAN in Transformation: The Future of Jobs at Risk of Automation, by Jae-Hee Chang and Phu Huynh, published as Bureau for Employers' Activities Working Paper No. 9 by the International Labour Office in July 2016. It is a desk study that takes the occupational computerisation probabilities from Frey and Osborne's 2013 Oxford study and applies them to labour force survey data for Cambodia, Indonesia, the Philippines, Thailand and Viet Nam. The 88 per cent is the share of garment wage workers in Cambodia in the high-risk band; the paper gives a range across countries of 64 per cent in Indonesia to 88 per cent in Cambodia.

### Does the ILO paper predict that 88 per cent of garment jobs will be lost?

No, and it says so directly. The paper's own text reads: "Lastly, the methodology does not attempt to predict the precise number of jobs that will be automated or displaced, nor does it identify an exact year when this will happen. Instead, it identifies the occupations and types of workers facing a high probability of automation over the next couple decades based on the nature of tasks involved." A high technical probability of automatability is not a forecast of job losses, and the authors were careful about the difference. The misuse happened downstream of them.

### What happened to garment employment in the ten years since?

In the country with the highest risk score, it grew. Cambodia's garment, footwear and travel goods exports reached a record 15.5 billion US dollars in 2025, up 15.7 per cent year on year, with garments alone accounting for 11.4 billion. Ministry of Labour data cited in the same report puts the sector at more than 1,500 factories and branch operations employing over 900,000 workers. Exports and employment both stand at or near records, not in the collapse the headline implied. That is a statement about export volume and headcount only; wages and working conditions are a separate and contested question.

### Has automation displaced any garment workers?

Yes, and it should be reported straight. A study by Solidaridad Bangladesh, the Labour Foundation and BRAC University, with fieldwork between August and October 2024 and a survey of 429 workers in Dhaka, Gazipur and Narayanganj, found a 30.58 per cent reduction in workforce across production processes, with helpers the most affected group. Read the denominator carefully: these are reductions per production process and per line among the factories surveyed, not a fall in national RMG employment. Bangladesh's garment workforce as a whole has not dropped by 30 per cent.

### Which garment jobs were actually displaced, and by what?

Cutting, and by CNC machinery rather than AI. In the same Bangladesh study, automation in cutting led to a 48.34 per cent drop in workforce while sewing recorded a 26.57 per cent decline — the deepest cut landed on the process everyone ignores, and the shallowest on the one everyone fears. Automatic spreaders and computer-controlled cutters remove spreading, marking and manual cutting labour, most of it helper roles. That is mechanisation, not machine learning, and it has been commercially available for decades.

### What would change the picture?

Three things worth watching. First, a robotic system that holds a general sewing line across mixed styles, sizes and fabrics rather than flat high-volume goods — that has not happened, and the flagship attempt now runs with sewers. Second, a sustained rise in wages relative to capital cost, which is what actually changes an automation payback calculation. Third, structural demand shifts such as nearshoring, which move where garments are made without necessarily changing how they are made. The defensible claim from the last decade is about timing and mechanism, not about garment work being permanently safe.

---

### Scan ERP by Country

[🇮🇳 India](https://scanerp.pro/india/)
[🇧🇩 Bangladesh](https://scanerp.pro/bangladesh/)
[🇻🇳 Vietnam](https://scanerp.pro/vietnam/)
[🇳🇵 Nepal](https://scanerp.pro/nepal/)
[🇰🇭 Cambodia](https://scanerp.pro/cambodia/)
[🇱🇰 Sri Lanka](https://scanerp.pro/srilanka/)
[🇪🇹 Ethiopia](https://scanerp.pro/ethiopia/)

### Know What Your Floor Costs Before You Automate

Any machine that claims to beat your current cost per operation has to be measured against a baseline you can produce. Scan ERP records production as it happens: bundle QR codes carrying lot, article, colour, size, component and quantity, scanned at each operation to capture operator, machine, operation and timestamp. Priced at $5 per machine or $0.001 per piece, whichever bills lower, from $39 per month.

[Request a Free Demo](https://scanerp.pro/#contact)

If you take one thing from this, make it a habit rather than a figure. When a statistic about your industry has been repeated to you for ten years, go and find the document. In this case the document is a free PDF, the relevant sentence is in the methodology section, and it says the opposite of the headline. That took me an afternoon, and it changed how I plan.

*Santosh Rijal is the founder of [Scan ERP](https://scanerp.pro/), a garment manufacturing ERP system designed for factory floor operations. He works directly with sewing lines, cutting rooms, and production supervisors across Nepal's garment manufacturing sector.*

---

### Related guides from the field

- [What Happened to the Sewbot Factory?](https://scanerp.pro/blog/sewbot-factory-what-happened.html) — A ten-year follow-up on the plant that was going to make 800,000 shirts a day.
- [AI in Garment Manufacturing: The Sewing Floor Reality](https://scanerp.pro/blog/ai-garment-manufacturing-sewing-floor-reality.html) — What actually runs on a sewing floor in 2026, and what doesn't.
- [Cutting Room Management and Fabric Waste](https://scanerp.pro/blog/cutting-room-management-reduce-fabric-waste.html) — The room where measurable automation actually landed.
- [Is Piece-Rate Pay Bad for Garment Workers?](https://scanerp.pro/blog/piece-rate-pay-garment-workers-evidence.html) — What the evidence actually says, read from the source studies.