---
title: "Seventy AI 'clerks' joined a Shenzhen district — and redrew China's government service map"
date: 2026-10-31
category: AI in Action
site: NeuroAI
canonical: https://neuroai.site/a/na-app-12345-ai-hotline
language: en
---

# Seventy AI 'clerks' joined a Shenzhen district — and redrew China's government service map

> From Shenzhen to Beijing, Chinese cities are wiring large models into the 12345 hotline, cutting dispatch errors and turning complaints into governance signals.

A resident in Shenzhen's Futian district files a noise complaint. Within seconds a system reads it, tags it, routes it to the right office, and drafts the reply — no human retyping the same details into three boxes.

Behind that speed are seventy AI "clerks" that went to work in one district government, and they are part of a much larger rewiring of China's citizen-service hotline.

## Key takeaways

- Shenzhen's Futian district deployed 70 AI "digital-smart employees" on DeepSeek in February 2025, covering 240 government service scenarios.

- Measured gains include document-format correction above 95% accuracy, review time cut by 90%, and complaint routing accuracy raised from 70% to 95%.

- The 12345 government service hotline (政务服务热线) has become the flagship deployment, adopted from Guangzhou to Liaoning to Beijing.

- Beijing's line handled more than 25 million demands in the past year, with AI assisting across intake, dispatch, handling and review.

- Officials stress the AI assists, not decides: humans stay in the loop for any action affecting citizens.

## Why the hotline is the testbed

The 12345 line is China's universal non-emergency gateway — call once about anything from a pothole to a delayed pension. For years its weakness was volume: busy agents, slow routing, and complaints that vanished into the wrong department. That is exactly the kind of repetitive, document-heavy work large models (大模型) are good at, which is why governments reached for them early.

In February 2025, Shenzhen's Futian district became a visible milestone. It launched 70 AI "digital-smart employees" built on DeepSeek and a customized Government Affairs Large Model 2.0, spanning eleven categories and 240 business scenarios across document handling, public services, emergency management and investment promotion. The district says the goal is to free humans for judgment calls, not to replace them.

## What the numbers say

The published metrics are specific enough to be useful. Document-format correction accuracy exceeds 95%, review time dropped by 90%, and error rates stayed under 5%. Livelihood-complaint dispatch accuracy rose from 70% to 95%. A cross-department task-assignment assistant improved efficiency by 80% and on-time completion by 25%, while a safety-production script generator was reported to run 100 times faster. None of these are independent audits, but they set a clear baseline for what districts are claiming.

Futian also wrote what it calls the first local "management charter" for government auxiliary robots, spelling out technical standards, scope, security and oversight — a signal that the experiment is being formalized, not just piloted.

## From one district to a national habit

The pattern spread quickly. Guangzhou deployed DeepSeek-R1 and V3 671B models in mid-February 2025 for policy interpretation and 12345 work-order dispatch. Liaoning's 12345 reported data-processing efficiency roughly 20 times that of manual work, and its AI assistant "Xiao Ning" passed the national internet regulator's generative-AI service filing — described as the first such government-model approval in the province. In several regions the models run in a localized, on-premise deployment so citizen data never leaves the government domain.

Beijing is the largest stage. Its 12345 line received more than 25 million demands in the past year — over 70,000 a day, nearly 3,000 an hour — and has accumulated close to 200 million cases over seven-plus years, with resolution and satisfaction rates above 97%. The city applies AI across four steps — receive, dispatch, handle, evaluate — covering 16 districts and 343 subdistricts and townships, with ten-language channels, a business seat handling 2,500-plus demands daily, and video sign-language service for deaf callers.

## The bigger shift: from response to prediction

The interesting move is not faster replies but turning complaints into signals. By tagging and clustering millions of requests, cities say they can spot emerging problems — a spike in wage arrears, a flooded road after heavy rain — and act before a flood of calls arrives. Beijing reports using models to grade extreme-weather risks during the 2026 flood season, triggering dozens of alerts forwarded to emergency departments.

This is where "AI in practice" meets governance: the same model that drafts a reply also helps officials decide where to send inspectors tomorrow.

## Honest limitations

The efficiency figures come from district governments and Chinese state-media reports, not from independent audits, and "20x" or "100x" gains depend heavily on the manual baseline chosen. The AI "clerks" assist and route; they do not make binding decisions about citizens, and the human-in-the-loop claim should be verified case by case. Deployment quality varies widely by city and by department, and long-term effects on service quality, not just speed, are still undocumented. Cities also risk a trap: faster routing can mask underlying service gaps if human follow-through weakens. Data-security claims of on-premise deployment apply in some regions but not necessarily all.

## What readers can do now

- If you live in a Chinese city with a 12345 line, try it for a non-urgent issue and notice whether routing feels faster — and always keep the human follow-up contact.

- Civic-tech observers elsewhere should watch Shenzhen and Beijing as real-world cases of large models (大模型) in public administration, not just demos.

- When evaluating any "government AI" claim, ask the two questions that matter: is a human still accountable, and is the citizen data kept on-premise?

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-12345-ai-hotline
Free to quote with attribution and a link to the original.
