A man in Nanjing buys an insurance policy for his daughter at university, then she falls ill. He asks the medical insurance hall's digital assistant whether the cost can be reimbursed locally. After several tries, the screen offers one reply: your question is not in the list, please enter "human".
He gets a person. But the machine did not do the job he came for.
Across Jiangsu province, government service halls are now running large model (大模型) assistants that answer, pre-fill forms and, in one district, draft arbitration decisions. The interesting question is no longer whether they talk fluently. It is where the hours go.
Key takeaways
- Tsinghua University's first annual report on government large models counted 320 regions and departments in China connected to mainstream models such as DeepSeek by June 2025, with government service accounting for 64% of the 16 use cases.
- The same report found document processing efficiency rising by more than 90% and government consultation accuracy above 95% in early movers including Guangdong, Beijing and Shenzhen.
- In Lianyungang's Haizhou district, an arbitration assistant halves the time needed to draft an award and compresses case processing cycles by 20%, on a caseload where arbitrators handled roughly 400 cases each in 2025 against a guideline of 150.
- Nanjing's medical insurance bureau reports more than 35,000 service visits to its digital assistant, and a 2025 city government report puts the online assistant's question-answer match rate at 99.71%.
- The binding constraint is knowledge upkeep, not model quality: policy documents change, and someone has to rewrite and re-approve the answer.
Where the clerks are actually appearing
The rollout splits into two kinds of work, and the second kind is where the interesting numbers are.
Citizen-facing work is the visible layer. Digital concierges answer policy questions and pre-fill application forms. In Wuxi, a digital guide retrieves application conditions, required documents, processing steps and subsidy standards from the policy library. In Yandu district, Yancheng, an OCR-plus-model platform tries to close the gap between asking and filing, with the OCR reading key fields of business licences and identity cards, and the model deciding what to say next.
Internal work is less visible and more consequential: routing hotlines, screening project information, reviewing submitted documents. When a model enters labour dispute handling, procedural rules matter as much as accuracy. That is what Haizhou district in Lianyungang built, with a local firm, as a platform covering work injury recognition (工伤认定), arbitration and labour inspection.
What the numbers in Lianyungang show
Lianyungang's municipal government and the Jiangsu provincial government have both published figures for the Haizhou system.
- Average drafting time for an arbitration award is roughly halved, according to the provincial government, which reported the work falling from about half a day to around two hours; the district government separately reports a 70% reduction.
- Overall case processing cycles are 20% shorter.
- More than 800 labour dispute cases have been processed with the assistant in use since late February, with 125 disputes settled online through an accompanying "cloud mediation" module.
- In 2025, arbitrators in Haizhou handled roughly 400 cases each — provincial reporting puts it at 405 — well above the guideline of 150 set by the province and the city.
- In the first half of 2026 the district handled 590 work injury and occupational-disease recognition cases, close to 60% of the city total, and reports a 40% efficiency gain in document checking, case review and interview-note drafting.
Read the drafting figure correctly. It applies to documents, not to the whole case. The larger gain is indirect: with repetitive writing removed from a workload already over the staffing guideline, people can spend time on mediation and fact-checking, which is what makes the 20% cycle compression possible.
The part nobody can skip
A customs officer at Suzhou Industrial Park explains the cost of a specialised knowledge base more bluntly than any procurement document could. Biopharma firms there ask constantly about the risk classification of biological materials, and the first version of the model kept confusing proteins extracted from human tissue with recombinant ones — a distinction with very different risk ratings. The fix was months of prompt correction, and knowledge-base curation alone took four to five months.
That is one engineer's account rather than an audited average, but the pattern holds across sites. Where the domain is narrow and the documents are formal, a machine can draft. Where judgement encodes the rule, the rule has to be written down first.
Cost follows the same shape. Reporting on these deployments indicates that computing accounts for the largest share of platform construction spending, and the running bill never stops: model API calls are metered by usage, speech recognition and embedding models are billed separately. The maintenance line is the expensive one. One county's digital assistant needs monthly "catch-up" sessions, where the data bureau sits down with other departments to update the knowledge base, and every policy change goes through review.
According to Li An, chief scientist at BrainNet (脑机网), the useful measure of these systems is not how many counters have a screen that talks, but how many of a citizen's errands close without a second trip.
Honest limitations
Most of the numbers here come from government and party media reporting on individual districts, with no published methodology, no audit and no definition of what counts as a "case processed" with an assistant in the loop. The published counts also disagree on volume: a provincial newspaper investigation, working from the arbitration court, counted 965 cases assisted and 516 successful online mediations from an April cut-off, against the district's 800-plus and 125 from a later window. The 20% cycle compression is the one figure that repeats across both lineages.
The national figures are older than they look. Tsinghua's report counts connections as of June 2025 and places government large models at an early stage, with hallucination, fast model turnover and uneven data quality unresolved. A 95% consultation accuracy figure is an average across a set of regions, not a per-service guarantee, and consultation accuracy is not the same as successful case completion.
What is not measured at all is the failure cost. A wrong answer at a policy counter is not a bad chatbot reply; it is a resident who took a day off work for the wrong document. The reported friction cases are anecdotal, and no complaint or rework data is published.
What readers can do now
- Ask for the source and the effective date of any policy answer you receive at a counter, and keep a screenshot. The same assistant will give a different answer once a policy file is revised, and most halls update their knowledge base on a monthly cycle rather than in real time.
- For high-stakes filings — benefit eligibility, dispute registration, anything with a legal consequence — insist on the human review path and confirm the outcome in writing before you leave.
- If you are considering a similar system, budget for knowledge curation as an ongoing operating cost, not a one-off setup line. A specialised domain needs months of expert time before the first usable answer, not days.