ILO's First AI-in-Manufacturing Conclusions: What 54 Countries Cut Before Agreeing

Zero times. That is how often the words "shall" or "must" appear in the International Labour Organization's first-ever tripartite conclusions on artificial intelligence in manufacturing, a five-page text that government, employer and worker delegates from 54 countries adopted unanimously in Geneva on 17 April 2026. "Should" appears 13 times. The record of proceedings, published on 7 September 2026, shows sentence by sentence which words were softened, which were deleted, and who objected. If you work in a factory, the deletions tell you more than the text that survived.

What the meeting was

[Fact] The ILO Governing Body decided in March 2025 to convene a technical meeting on "challenges and opportunities for promoting decent work, productivity and a just transition arising from artificial intelligence in the manufacturing industry". It met from 13 to 17 April 2026 with 115 participants: 36 government representatives from 24 Member States, 46 government observers from 19 more, 8 employer and 8 worker representatives. The chair was Magnus Norddahl of the Icelandic Confederation of Labour; the three vice-chairs were Jason Dale (United States government), Delphine Rudelli (CEEMET, the European tech and industry employers) and David White (International Association of Machinists and Aerospace Workers).

[Fact] The adopted text (reference TMDWAI/2026/8) has 17 paragraphs plus an appendix listing 40-odd ILO instruments. It goes to the 358th Session of the Governing Body in November 2026 "for its consideration". The record (TMDWAI/2026/9) is the ILO secretariat's summary of who said what, including a paragraph-by-paragraph account of the drafting fight in the Working Party on 15 and 16 April.

One number first, because three ILO documents disagree on it. [Fact] The conclusions say manufacturing "provided direct employment to more than 489 million people" in 2024, citing ILOSTAT. The background report prepared for the same meeting says 494 million in 2023, citing ILOSTAT. The ILO press release of 21 April says "almost 500 million". Same source, different reference years, a one-percent wobble. It matters only because the rest of this article is about precision of wording.

The scorecard: 11 of 17 paragraphs were rewritten

[Estimate] My own count from the record: six paragraphs (1 to 3, 6, 8 and 15) were "approved as proposed" by the ILO Office. The other 11 were "approved as amended". The amendments run in one direction far more often than the other.

The regulation fight. [Fact] Paragraph 4 explains why small manufacturers lag on AI. The Office draft blamed "regulatory environments". The Workers' group proposed "the absence of an adequate regulatory environment". Employers, backed by the Government group, won "unsupportive and complex regulatory environments". The Worker vice-chair agreed "in the interest of consensus, but noted that it represented a major concession for his group".

Here is the odd part. [Fact] The meeting's own background report cites a survey in which 19 percent of manufacturing employers name government regulation as a barrier to AI adoption, against 58 percent naming technology cost and 43 percent naming lack of skills. [Claim] The word the room fought hardest over describes the barrier the employers' own survey ranks last.

The intensity fight. [Fact] In paragraph 5 the Workers' group asked to add "deskilling", "pace of work", "human dignity and integrity, discriminations, surveillance". Employers replied that "skills" covered deskilling. After a round trip through "pace of work" the text reverted to the Office's original "intensity of work". In the final five pages, "surveillance" appears zero times, "algorithmic" zero times and "monitoring" once.

The consultation fight. [Fact] Paragraph 13 is the one that reaches the factory floor. The Office draft said workers "should be informed, involved and consulted, in a timely manner, on the introduction of AI systems likely to have major effects on them". The adopted version reads: "while recognizing that employers are the drivers of investment and innovation, and have the prerogative of business operations and systems, workers should be informed, involved and consulted, when applicable, in a timely manner on the introduction of AI systems likely to have an effect on them". The phrase "when applicable" now appears three times in the text. The Employer vice-chair had said on day three that decisions on deploying AI "fell within managerial prerogative rather than the scope of social dialogue".

The guidelines fight. [Fact] Workers, "supported by several members of the Government group", wanted a recommendation that the ILO develop guidelines on AI in manufacturing. Employers and the United States said there was "insufficient data available to formulate robust guidelines". The Office offered a middle sentence about convening a future meeting. That too was cut. The adopted paragraph 17(a) ends: "Such research is to inform possible future action of the ILO." The Workers' and Government groups placed on record that all their members "with the exception of the United States, had wished to include a recommendation on developing guidelines". The word "guideline" does not occur in the body of the conclusions. The Workers' group had opened the week asking for a path toward an ILO convention.

The risk-assessment fight. [Fact] The draft "recommended" that enterprises assess AI risks and referred to the UN Guiding Principles on Business and Human Rights. Employers argued companies already carry heavy due-diligence obligations. The final paragraph 16 says enterprises "are encouraged to carry out risk assessments", and the UN Guiding Principles moved to the appendix.

Deleted outright. [Fact] References to data privacy and intellectual property in the skills clause of paragraph 14, the words "job insecurity" in paragraph 12, and the whole of clause 17(e). "Wages" survive three times; the one operative mention, "improve working conditions, including wages" in 14(c), was accepted by employers "on the basis that references to wages or sharing the gains of progress did not appear elsewhere in the text". The Workers' opening demand for "reduced working time without wage cuts" left no trace: "working hours" appears zero times.

What the United States put on the record

[Fact] The record lists the same government objecting, losing, and asking that its objection be noted, at least six times: to "just transition" (proposed replacement: "successful worker transitions"), to "inclusive" in skills and social protection, to "gender equality", to the "transformative gender agenda" wording of the Centenary Declaration, to "participation", and to the just-transition guidelines in the appendix. On "gender" versus "sex" in data collection the chair brokered "sex/gender". At paragraph 153 the US representative "questioned why nearly every drafted document required repeating the same discussions".

[Fact] The US case rested on numbers: early findings from the 2022 OECD AI surveys in which roughly 80 percent of AI users said the technology improved their work performance and 8 percent said it worsened it. The German worker representative answered with a survey of more than 10,000 employees in which 58 percent saw no or only minor productivity benefit from AI. [Claim] Neither figure made it into the conclusions, and neither was collected on a factory floor specifically for this question. That absence is the honest finding: the meeting agreed the evidence base is thin and then argued about wording anyway.

The counter-narrative: this meeting was not about robots

[Fact] On day one the Employer vice-chair insisted that "AI should not be equated with robots" and that "using examples of robotization to draw conclusions about job losses risked obscuring the discussion". The Indian worker representative had just described an industrial corridor where new robots "displaced many manual jobs". In effect, the most vivid job-loss story in the room was ruled outside the mandate.

The background report backs the narrower framing. [Fact] It reports that manufacturing workers are generally less exposed to AI than service workers; that only 0.4 percent of workers in low-income countries have high occupational exposure to AI, against 5.5 percent in high-income countries; and that in high-income settings women's average exposure (8.5 percent) is more than twice men's (3.9 percent). It says cost savings and revenue gains from AI adoption have so far been "commonly below 10 and 5 per cent, respectively". [Estimate] The AI-in-manufacturing market it cites, US$23.4 billion in 2024 to US$155 billion by 2040, is a 6.6-fold rise over 16 years, or roughly 12.5 percent a year; that is a supplier forecast, not an employment forecast.

[Claim] So the honest reading for a machinist or a line supervisor is not "AI will take your job by 2030". It is that the people who set global norms just agreed, in writing, that they do not yet know what AI does to a factory job, that employers keep the decision, and that you should be consulted "when applicable".

Two survivors are worth knowing. [Fact] Paragraph 12 recognises "new risks, including those resulting from increased monitoring practices". Clause 14(o) asks governments to "strengthen labour administration, including labour inspection". The Swedish worker representative described a plant that moved from manual to AI-driven automated production under a union agreement: a skills assessment first, then upskilling, and "no worker had been laid off for lack of skills" by the time the new line ran. Singapore's delegate described company training committees led by unions, with government transformation grants that require union sign-off on wage and career outcomes.

We track AI exposure for the roles this text talks around. See our pages for industrial engineers, industrial production managers, industrial machinery mechanics, welders and robotics engineers. For how an ILO text reads exposure differently from adoption, see our piece on the ILO's ASEAN generative-AI study; for the monitoring risk the conclusions only name once, see ILO on algorithmic management of Cambodian drivers.

What this record cannot tell you

A tripartite consensus text is a lowest common denominator by design. Unanimity means every group held a veto, so a sentence that all three could sign cannot carry any one group's strongest claim. That is why the score is 13 "should", zero "shall". It is also why reading the deletions is legitimate: they are the only place the disagreements are visible.

The record is a secretariat summary, not a transcript. Positions "noted for the record" are not votes, and the Working Party met in closed session. The conclusions are, in the Office's own words at paragraph 130, "aspirational and non-binding"; the Governing Body in November decides only on their dissemination. My word counts are of the English text and could shift in the French and Spanish versions. And our own exposure scores are occupation-level, which bundles exactly the AI-versus-automation distinction the employers insisted on separating.

What a manufacturing worker can do with this

1. If your plant introduces AI-driven monitoring or scheduling, paragraph 12 and clause 14(i) on future-proof privacy rules are the ILO wording to cite. Ask whether the "encouraged" risk assessment of paragraph 16 was done.
2. Paragraph 9 makes skills "a shared and continuous responsibility of governments, employers and workers". The Swedish and Singapore examples show what that looks like when it is written into an agreement before the line changes.
3. Watch the 358th Session of the Governing Body in November 2026. If the guidelines idea returns there, it will be because the data the employers asked for started to arrive.

Sources

• ILO, Record of proceedings: Technical meeting on challenges and opportunities for promoting decent work, productivity and a just transition arising from artificial intelligence in the manufacturing industry (Geneva, 13–17 April 2026), TMDWAI/2026/9, published 7 September 2026. https://www.ilo.org/resource/record-proceedings/record-proceedings-technical-meeting-challenges-and-opportunities-promoting
• ILO, Conclusions and recommendations, TMDWAI/2026/8, adopted 17 April 2026. https://www.ilo.org/sites/default/files/2026-04/TMDWAI-2026-8-EN.pdf
• ILO Sectoral Policies Department, AI in manufacturing: Challenges and opportunities for promoting decent work, productivity and a just transition (background report), TMDWAI/2026, Geneva, 2026, CC BY 4.0. https://www.ilo.org/sites/default/files/2026-03/TMDWAI-2026-EN%20-.pdf
• ILO news, "ILO adopts first-ever conclusions on AI in manufacturing work", 21 April 2026. https://www.ilo.org/resource/news/ilo-adopts-first-ever-conclusions-ai-manufacturing-work

Quotations are short excerpts from ILO documents, attributed above. Paragraph and word counts are the author's own from the English texts.

AI-assisted analysis: this article was drafted with AI assistance from the primary ILO documents and reviewed before publication.