ILO on AI in Chinese firms: 13 of 21 companies can't measure the gain

Thirteen of the 21 Chinese companies interviewed for a new ILO brief could not put a number on what AI has done to their productivity. The other eight could, and their numbers are large: recruitment cycles cut from 30 days to 13, a customer-service floor of 300 people going from 6,000 to 15,000 resolved queries a day, a document-reading team of more than 10 shrunk to two or three. Read together, the two facts say more about where China's enterprise AI wave actually is than either does alone.

The brief, "Artificial intelligence adoption in Chinese enterprises: Productivity effects, workforce implications, and policy challenges," was published by the ILO Research Department on September 15, 2026. Ekkehard Ernst of the ILO wrote it with six co-authors from Renmin University of China (Zhong Zhao, Huilin Zhu, Yuhui Li, Xin Wei, Hao Zhang, Zeyang Chen), who ran the interviews and the survey during the first half of 2026. I read the 12-page PDF in full; every figure below comes from it.

Two datasets, one small and one wide

[Fact] The first source is 21 semi-structured interviews with executives, department heads and core staff at firms spanning manufacturing (6), finance, insurance and real estate (4), business services (4), and one each in communication, construction and education technology; four firms are listed as unclassified in the brief's own Table 1. The size range is extreme: an 8-person media startup and a 21-person venture capital firm at one end, an opto-electronics manufacturer with more than 20,000 workers and an insurance conglomerate with 270,000 employees and annual revenue above 1.14 trillion yuan at the other.

[Fact] The second source is a survey of 1,591 Chinese professionals across industries, ownership types (state-owned, private, foreign-invested) and career levels from entry to senior management. The brief reports four headline shares from it: 56% see AI adoption as an inevitable trend, 47% believe AI creates more jobs than it displaces, roughly a third express concern about social stability or "human uniqueness," and 39% expect AI to reduce their income.

The two sources do different jobs. The interviews explain how firms are adopting; the survey says how workers feel about it. Neither measures what happened to employment.

What the eight firms with numbers reported

[Fact] The opto-electronics manufacturer's "lighthouse factory" reported a 30% production-efficiency gain after deploying small AI models in 2021, with other client factories at roughly 20%. The same firm cut a customer-specification reading team from more than 10 people to two or three, halved a parameter-changeover time from 120 to 60 minutes, and says its AI products can reduce labour cost by about 80% in specific production scenarios.

[Fact] In services, an insurance company said 300 customer-service employees raised their daily throughput from 6,000 to 15,000 issues. The largest insurance group's HR department cut recruitment cycle time from 30 to 13 days. An AI-native talent management firm said project managers went from running two or three projects at once to five or six, client response speed improved 30%, and core efficiency metrics improved by 30% to 100%. An HR services firm measured a 30% drop in R&D working hours in Q1 and a 28% rise in per-capita customer-service output in Q3. A travel services firm estimated its AI data handling equals five or six full-time employees. An ed-tech company said it halved its service pricing, with AI study rooms running 80% below the cost of traditional tutoring classes.

[Estimate] Some arithmetic the brief does not do. The insurance floor's numbers work out to 20 resolved issues per agent per day before AI and 50 after, on the same headcount. The specification team's cut is a 70% to 80% reduction in labour on that task. And the productivity table (Table 3) lists 13 metrics from eight firms, which means 13 of the 21 interviewed companies, 62%, offered no quantified gain at all. The brief's own wording for the majority is that "most firms lack systematic evaluation frameworks": the biggest insurer has "no formal metrics on the business side," the VC firm describes better research reports "without any quantitative measurement," and the media firm reports "no quantitative assessment at all."

That 62% is the number I would lead with if I were a policymaker. Firms are spending on AI without a way to tell whether it worked.

Three ways in, three depths

[Fact] The brief sorts adoption into three organisational models. A centralised specialist team builds proprietary tools (the opto-electronics manufacturer, the talent firm). Business-embedded integration weaves AI into existing units, which is the pattern in insurance and banking. Spontaneous diffusion spreads bottom-up in small, flexible firms; at the VC firm, employees started using AI tools before the company formalised support.

[Fact] It also describes three depths of impact: tool-assisted efficiency (speed up existing tasks), process-embedded optimisation (rebuild workflows around AI), and business-model innovation. Most sampled firms are at the first or second stage. Only three, the talent firm, the opto-electronics manufacturer and the ed-tech company, show signs of the third.

One detail that cuts against the "China is racing ahead" framing: several firms described deliberately waiting. A state-linked digital services firm cited "sunk cost" risk from premature investment and entered only once open-source foundation models had matured and compute costs had fallen. The tools these firms name are mostly general-purpose (Doubao, Kimi, Yuanbao, Copilot, the Feishu collaboration suite), often trialled side by side before one is picked.

The barrier list looks like everyone else's

[Fact] Output quality is the most-cited problem, at eight or more firms. The HR services firm says AI customer-service accuracy falls below the 95% threshold it needs before letting AI answer clients directly. The ed-tech firm abandoned AI-generated educational animation because correction costs were prohibitive. The media founder rates AI output as equivalent to "an ordinary clerical worker." The opto-electronics manufacturer draws a line between small models that reliably solve specific production problems and generative AI, which it describes as in a phase of "technological disenchantment."

[Fact] Skills and resistance come next (six or more firms), then regulation and data security (six or more), system integration (five or more), and cost versus return (five or more). The ed-tech company reports that employees over 40 show "markedly weaker" AI capabilities. The largest insurer reports explicit resistance from HR staff who see AI as a threat to their own roles, and a gap between headquarters staff, who adapt, and branch staff, who adapt less. Multinationals' Chinese subsidiaries face a further layer: the HR services firm says some overseas headquarters permit only private model deployment and prohibit open models.

Only two firms cite data quality or annotation as a barrier. That is lower than I expected from a manufacturing-heavy sample and may reflect who was interviewed rather than what is hard.

The optimism and the cases sit uneasily together

Here is the tension the brief flags but does not resolve. 47% of surveyed professionals believe AI creates more jobs than it displaces, which the authors note is more optimistic than surveys in many OECD countries. Yet the interviews contain a specification team cut by 70% to 80%, a travel firm avoiding five or six hires, a talent firm whose junior consultants now deliver "work previously requiring mid-to-senior expertise," and a VC partner who says that in future "people in this industry will become 'cheaper.'"

[Claim] The brief itself points out that the junior-consultant finding is "in notable difference to current discussions in the United States about the risk of entry-level jobs disappearing," citing the Stanford Digital Economy Lab's "canaries" paper. That is a fair contrast, but it is a contrast between one firm's account and a payroll-data study, not between two measurements. The Chinese evidence here is that juniors are being made more productive; whether fewer of them are hired as a result is a question this brief cannot answer, because, in its own words, it "does not directly include worker voices" and the survey is "limited to attitudinal measures rather than objective employment outcomes."

The 39% who expect income declines is the figure that connects to the cases. The brief reports a gender pattern: male employees show both more optimism and more pessimism about AI's income effects, and high-skilled men are over-represented among those expecting gains. It gives no cross-tabulation, so we cannot tell how much the 47% and the 39% overlap. They are not mutually exclusive positions, and a worker can believe both that AI creates net jobs and that it will cut their own pay.

What the brief tells you to make of it

[Fact] The authors' policy section names five areas: skills and lifelong learning (with mid-career and older workers singled out), managing displacement (they say the gains concentrate in "routine clerical, administrative, and customer-service roles"), AI governance across national, sectoral and multinational layers, standardised productivity measurement that includes job quality and not only output per worker, and SME access, since many small firms "don't dare to invest and don't know how to invest."

[Claim] One line worth keeping: the opto-electronics manufacturer says it has not put AI into daily production "as autonomous systems," because "black box" operation in manufacturing carries too much risk, and in some processes the system is designed to withhold results if it has not consulted the required documents first. That is a description of human-in-the-loop as an engineering control, not as a slogan, and it matches the direction of the ILO's manufacturing-meeting conclusions we covered last week.

Limits, in the authors' words and mine

[Fact] The brief's own limitations section is unusually direct. The 21 firms are "purposively selected rather than randomly drawn" and "cannot be statistically generalised." All productivity data are self-reported by managers, "have not been independently verified," and "may reflect optimism bias, measurement inconsistencies, or selective reporting." No firm ran a controlled evaluation, so the gains "cannot be rigorously attributed to AI adoption." The metrics are heterogeneous and "cannot be directly aggregated or compared." Some firms are represented only by "summary bullet points."

Two things I would add. First, the survey's sampling method, response rate and field dates are not reported in the brief, so the 1,591 figure should be read as a large convenience sample, not a representative one. Second, this is the third China data point we have covered this year and they do not measure the same thing: the 752 million job-ad analysis tracked what employers post, this brief tracks what managers say, and neither observes what workers actually do all day. The nearest thing to that in our recent coverage is the ECB's worker survey, which is European.

If this is your job

The tasks the brief says AI is eating are specific: résumé screening, routine query handling, document reading, data collection. If you work in customer service or human resources, the Chinese cases are a preview of the throughput expectations that follow AI deployment, and the 300-agent floor that now handles 15,000 issues a day is the template. If you are in production management, the manufacturer's small-model-first, generative-AI-later sequence is a reasonable one to copy. And if you write code, the HR services firm's 30% cut in R&D hours came with "persistent bottlenecks" in generation quality, which is consistent with what software developers elsewhere report.

The one thing every firm in the sample agreed on is that they would hire for "AI literacy" and, more specifically, an "AI mindset": knowing where a tool fits, writing the prompt, and checking the output. Nobody in the brief has yet measured whether that mindset protects a job. They have measured that they are hiring for it.

Sources

• ILO, "Artificial intelligence adoption in Chinese enterprises: Productivity effects, workforce implications, and policy challenges," Research Brief, Ernst, Zhao, Zhu, Li, Wei, Zhang, Chen, September 15, 2026. Licensed CC BY 4.0. https://www.ilo.org/publications/artificial-intelligence-adoption-chinese-enterprises-productivity-effects (PDF: https://doi.org/10.54394/00035342)

The per-agent throughput (20 to 50 a day), the 70% to 80% team reduction, and the "13 of 21 firms without a quantified gain" count are this site's arithmetic on the brief's tables. Occupation-page links are this site's mapping; the brief does not use O*NET categories.