What a Person-Time Cannot Tell a Courier

Item six of State Council General Office Document No. 29, signed on 16 October 2026 and carried by Xinhua a week later, turns on one verb: 鼓励, to encourage. Employers are encouraged to "探索建立失业缓冲期制度", to explore an unemployment buffer period, warning affected workers in advance and helping them into other work as AI rolls out. Item seven goes further. AI projects in market-access sectors that could cause a large employment shock are to receive an employment risk assessment during review.

Zhang Ying, writing in Huxiu, calls it the first State Council text to treat AI's job effects as a matter for advance public governance, and she is careful about its limits: "large" has no definition, the method is unwritten, there are no penalties, and an assessment resting on self-reporting will miss a good deal. All true, and all fixable in the follow-up rules she expects. The defect I worry about more sits one level down, in how China counts the retraining that is meant to catch these workers. I think the country has made a reskilling promise on a scale few governments would attempt, and that it is measuring that promise in a unit that cannot show whether it was kept.

The unit is 人次, the person-time. On 10 October, Vice Minister Li Zhong announced the subsidised training numbers for 2026 to 2030, and the Beijing News account two days later spelled out what they mean with unusual care: more than 50 million is the target, "统计单位是人次，不是去重后的劳动者人数". One worker on three courses is three. A second target of 40 million new skills certificates includes certificates earned through schools, employers and the market, so it is no success rate for the subsidised programme. The last plan aimed for 75 million person-times and reported 95.73 million. That beat tells you seats were filled. Renmin University's Zhao Liqiu told the paper what it leaves out: "部分培训项目仍存在重参训取证、轻就业增收的问题." Some programmes care about attendance and certificates and care less about jobs and income.

Set that beside the most visible private retraining pledge in the country. At the APEC China business forum on 21 June, JD.com's founder Liu Qiangdong said robots would make couriers unnecessary and that he did not want his 700,000 frontline workers left without income. In an internal video reported on 27 May, he had been blunter: "对于被机器取代的一线员工，京东一个都不会开除." Not one displaced frontline worker fired. The Nirvana Plan sends them in batches to more than 120 contracted vocational colleges to learn robot repair, maintenance and fault diagnosis.

Now the classroom arithmetic. Workers' Daily, the union federation's paper, reported on 28 July that JD has eight robot repair centres, a few thousand repair engineers, and plans to hire more than 10,000 repair specialists within three years. Ten thousand is about 1.4% of 700,000, by my division. The paper asked the obvious question, whether end-of-line repair can absorb 700,000 people, and noted that JD had not disclosed course length, certification standards, assessment methods or pay during training. A vocational teacher reading that list would stop at the second item. You cannot plan a term without knowing what the student must be able to do at the end of it.

The fair reply on JD's behalf is that repair is only one door. The same report says JD is pursuing certification in infant care, nutrition and auto repair, and JD Services will spend 200 million yuan this year training more than 10,000 engineers across trades, which works out to at most 20,000 yuan a head. On 15 October, at its Double 11 launch, the company claimed the world's largest blue-collar training system, more than a million trained, with a million more over three years. "World's largest" is JD's phrase and I have seen no audit of it. Read the Chinese closely and the running total is again in 人次. The direction is sensible: human services absorb more people than robot pits do. Then the plan has a different destination from the one its founder announced, and the people boarding it should be told which.

Here is how the main counters compare.

That third row is the state of evidence inside firms. The ILO's research brief of 15 September, drawing on 21 firm interviews and a Renmin University survey of 1,591 professionals, reports gains such as a 57% cut in recruitment time and 150% more customer queries handled. Then it says what they are worth: all figures are self-reported, and "no firm described a controlled experimental evaluation design." One digital manufacturer called AI's effect on overall productivity "difficult to evaluate and quantify." An ed-tech firm in the sample found employees over 40 "demonstrate markedly weaker AI application capabilities." If firms cannot yet measure the gain from the tool, the 45-year-old sorter sent on a course has no reason to trust a certificate as proof of anything.

The fourth row is the most interesting line of Chinese labour policy I have read this year, and it comes from a province. Under Jiangsu's rule, as the Beijing News describes it, if at least 50% of a project-based class is in stable employment within six months of finishing, a locality may pay the training provider up to 20% above the standard subsidy. Stable means an employer has paid at least three consecutive months of basic pension contributions. A pension ledger is a dull instrument, and that is its merit. A certificate is issued by someone paid to issue it. A contribution record is created by an employer paying real money for a real worker, and the state already holds the file. Shandong, by comparison, has since autumn 2025 required at least 32 hours of AI general education for technical-school students, which is a sound syllabus decision and a measure of input.

The serious objection to scaling Jiangsu's test comes from the economists who studied the American version. Heckman, Heinrich and Smith, examining the US Job Training Partnership Act in 2002, found that the short-run measures used to monitor performance were "weakly, and sometimes perversely, related to long run impacts." Their finding should chill anyone tempted by a six-month bonus. Card, Kluve and Weber's meta-analysis of more than 200 studies adds that average programme effects are close to zero in the short run and "become more positive 2-3 years after completion." A six-month test could therefore reward the provider who picks the easiest students and punish the slow, deep course that pays off later.

I accept both findings, and I think they argue for a longer ledger, with no retreat to counting seats. China's social-insurance file keeps running after the course ends. The same record can be read at 36 months at little extra cost, and because contribution bases follow declared wages, imperfectly and within a floor and a ceiling, it can say something about income too. As for selection, the Heckman paper's second finding was that the efficiency gains or losses from cream skimming were small. The fairness problem remains, so publish results by age band and prior schooling, and the sorter over 40 stays visible.

Inside JD the test needs changing, because a no-firing pledge means the pension record keeps ticking whatever the course achieved. There the outcomes are role, pay and retention against a comparison. And JD has been handed the comparison for free. Batches mean someone goes first and someone goes later. Assign the order by lot, depot by depot, and the later batches become the control group the ILO could not find anywhere in its sample. It costs almost nothing beyond a random number and the discipline to report it. I would not accept any figure from this programme, or from the national one, that cannot say how many distinct workers went in and what they earned afterwards.

Document No. 29 asks for shared cross-agency employment data and an intelligent monitoring model in item sixteen, so the plumbing is at least on paper. Zhang Ying's warning stands beside it: firms that report AI-driven cuts candidly may take on more legal risk than firms that call them restructuring. A monitor fed by that incentive will see what it is shown.

I can find no date from JD for when the first batch finishes, and no name for the job on the other side. When a sorter comes back from a term of fault diagnosis to a company that has promised not to fire her, and the robot pit has no opening, who will be allowed to count what happened to her?

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.