南方周末城市（区域）研究中心发布报告，系统梳理人工智能大模型在中国地方政府政务领域的应用现状，涵盖政务服务、机关办公、社会治理、辅助决策四大场景1。报告指出政务AI呈现明显梯度分化，头部城市及主城区领跑，同时暴露区域失衡、基层运维薄弱、人才短缺、制度滞后及公众参与缺失等问题，并提出集约化部署、省级底座复用、人机协同与参与式治理等对策1。据智能超参数统计，2025年全国大模型中标项目达7539个，披露中标金额295.2亿元1。
在政务服务方面，深圳AI政务助手“深小i”截至2026年6月注册用户超150万，提供服务超600万次，单次解答精准率超过95%1。深圳南山区2025年2月接入DeepSeek，7月上线“AI智办”专区，材料智能核验使填报错误率降低80%、首次申报通过率提升至90%以上1。北京海淀区2025年11月推出政务服务智能体“淀小二”，覆盖全区1700多个事项，教师资格认证办理从8分钟压缩至2分钟1。福建漳州“候喜”AI智能助手学习超15万份政策文件，构建涵盖全市69个部门、5万余个事项的知识库，累计使用量突破61万次，群众办事必填材料减少25%、事项一次性办结率提升80%1。广州“穗@i企”政策智算器归集三千余条惠企政策，为280万家企业推送政策超5900万次1。江门蓬江区基于DeepSeek本地化部署的政务AI助手“江小i”覆盖1700多个政务服务事项，85个高频事项实现“边聊边办”，2025年12月获评南都“AI治理应用十大示范案例”1。惠州仲恺高新区2025年2月上线“双通道”政务服务智能支持系统，窗口单次业务咨询时间由平均6分钟缩短至3.5分钟1。
在社会治理与辅助决策方面，青岛依托城市云脑大模型实现数万件政务事件全自动派单，准确率达92%，单事件派发耗时压缩至1秒内1。2025年北京经开区落地全国首个全流程AI执法大模型，办案效率提升超50%；2026年推出AI+AR智能执法眼镜，在5到8米外即可自动捕捉车牌、秒级比对备案状态1。应急管理部“久安”大模型可在30秒内从数万路视频中锁定内涝风险1。上海“禹智河洛”防汛智能体落地后核心业务处置效率提升65%以上1。浙江财政“AI哨兵”预警准确率由23.3%提升至61.3%，风险疑点业务量增长6倍多1。上海浦东新区上线“12345市民服务热线大模型”，搭建19个智能体，单条工单审核从15分钟缩短至5分钟、审核准确率达90%；浦东2024年受理热线工单41万件、占全市三分之一1。杭州滨江区基层治理助手“滨小治”将文书工作时长缩短70%1。浙江宁波奉化区90天内将应用系统压减39.3%、运维成本降低30%1。广东“粤经济综合智能体”以“1个经济大脑+10大专业智能体”为架构，已服务全省近10万公职人员1。此外，2025年2月，深圳龙岗区在政务外网部署上线DeepSeek-R1全尺寸大模型1。
中国信息通信研究院副总工程师何伟指出，文本与机器人协同的政务大模型已分流近三分之一的人工服务1。国家信息中心副主任王冬欣指出，当前政务智能体存在身份不绑定、责任链条断裂、伦理风险难管控等问题1。中央网信办、国家发展改革委《政务领域人工智能大模型部署应用指引》要求“避免重复建设、无效建设，避免未审先建、建而不管，避免强制使用、无效使用”1。
The Southern Weekly Urban and Regional Research Center has released a report systematically reviewing the application of artificial intelligence large models in local government affairs across China, covering four major scenarios: government services, administrative operations, social governance, and decision-making support 1. The report highlights a clear gradient differentiation in government AI, with leading cities and their main urban districts taking the forefront, while also exposing issues such as regional imbalance, weak grassroots operation and maintenance, talent shortages, institutional lag, and a lack of public participation 1. To address these challenges, the report proposes countermeasures including intensive deployment, reuse of provincial-level foundations, human-machine collaboration, and participatory governance 1. Meanwhile, statistics from Intelligent Hyperparameters show that in 2025, there were 7,539 winning bids for large model projects nationwide, with a disclosed total bid amount of 29.52 billion yuan 1.
In the realm of government services and administrative operations, various cities have launched innovative AI applications to enhance efficiency. In February 2025, Longgang District in Shenzhen deployed the full-scale DeepSeek-R1 large model on its government extranet 1. Nanshan District in Shenzhen also integrated DeepSeek in February 2025 and launched the "AI Smart Service" zone in July, reducing form-filling error rates by 80% and boosting the first-time application approval rate to over 90% 1. Shenzhen's AI government assistant "Shen Xiao i" has registered over 1.5 million users and provided more than 6 million services by June 2026, with a single-answer accuracy rate exceeding 95% 1. In Beijing, Haidian District introduced the "Dian Xiao Er" government service intelligent agent in November 2025, covering over 1,700 matters and compressing the processing time for teacher qualification certification from 8 minutes to 2 minutes 1. Shanghai's Pudong New Area launched a "12345 Citizen Service Hotline Large Model" with 19 intelligent agents, shortening the review time for a single work order from 15 minutes to 5 minutes and achieving a 90% review accuracy rate 1. Notably, Pudong handled 410,000 hotline work orders in 2024, accounting for one-third of the city's total 1. In Fujian, Zhangzhou's "Hou Xi" AI assistant studied over 150,000 policy documents to build a knowledge base covering more than 50,000 matters across 69 departments, with cumulative usage exceeding 610,000 times, reducing required materials by 25% and increasing the one-time completion rate by 80% 1. Guangzhou's "Sui @ i Qi" policy intelligent calculator aggregated over 3,000 business-friendly policies, pushing them to 2.8 million enterprises more than 59 million times 1. Jiangmen's Pengjiang District deployed the "Jiang Xiao i" assistant based on the localized deployment of DeepSeek, covering over 1,700 matters, enabling "chat-and-process" for 85 high-frequency items, and it was awarded as one of the "Top Ten Demonstration Cases of AI Governance Applications" by Nandu in December 2025 1. Huizhou's Zhongkai High-tech Zone launched a "dual-channel" intelligent support system in February 2025, cutting the average window consultation time from 6 minutes to 3.5 minutes 1.
AI is also transforming social governance, law enforcement, and decision-making support. Qingdao relies on its city cloud brain large model to achieve fully automatic dispatching of tens of thousands of government events with a 92% accuracy rate, compressing the dispatch time for a single event to under one second 1. In 2025, Beijing's Economic-Technological Development Area implemented the country's first full-process AI law enforcement large model, boosting case-handling efficiency by over 50%, and in 2026, it introduced AI+AR smart law enforcement glasses that can automatically capture license plates from 5 to 8 meters away and compare them with filing statuses in seconds 1. The Ministry of Emergency Management's "Jiu An" large model can identify waterlogging risks from tens of thousands of video feeds within 30 seconds 1, while Shanghai's "Yu Zhi He Luo" flood control intelligent agent has improved core business processing efficiency by more than 65% 1. Zhejiang's financial "AI Sentinel" increased its early warning accuracy from 23.3% to 61.3%, with the volume of risk suspicion business growing by more than six times 1. Guangdong's "Yue Economy Comprehensive Intelligent Agent," structured with one economic brain and ten professional agents, has served nearly 100,000 public servants across the province 1. In Hangzhou, Binjiang District's grassroots governance assistant "Bin Xiao Zhi" reduced document processing time by 70% 1. Furthermore, Fenghua District in Ningbo, Zhejiang, reduced its application systems by 39.3% and lowered operation and maintenance costs by 30% within 90 days 1.
Experts and policymakers are closely monitoring the development and risks of these technologies. He Wei, Deputy Chief Engineer at the China Academy of Information and Communications Technology, pointed out that government large models combining text and robotics have already diverted nearly one-third of manual services 1. However, Wang Dongxin, Deputy Director of the State Information Center, noted that current government intelligent agents face issues such as unbound identities, broken responsibility chains, and difficult-to-control ethical risks 1. In response to these challenges, the Cyberspace Administration of China and the National Development and Reform Commission issued the "Guidelines for the Deployment and Application of Artificial Intelligence Large Models in the Government Sector," which explicitly require "avoiding redundant and ineffective construction, preventing unapproved construction and unmanaged operations, and prohibiting forced or ineffective use" 1.
评论
还没有评论，欢迎留下第一条。