你公司上了AI客服，裁了两个坐席，HR走的时候没接到任何正式通知。国办这份文件第一次把这件事写成了制度：大规模推广前必须提前告知，还得给缓冲期。
国办这18条，到底在管什么
10月23日，国办印发《关于适应人工智能技术发展助力构建就业友好型发展方式的意见》（国办发〔2026〕29号），落款10月16日。6方面18项任务，核心逻辑就两件事。AI能创造的岗位必须创造，能替代的必须设缓冲。
6个方向拆开：创造、替代、培训、服务、保障、预警
排列顺序即优先级。第一块「强化就业创造效应」放最前，立场不是叫停，是先造新岗再动旧岗。第二块「构建就业友好型升级路径」才谈替代，限定词是「稳慎把握人工智能在劳动密集型行业应用的范围、力度和节奏，尽可能减少对从业人员的冲击」。后四块依次是培训、就业服务、权益保障、监测预警，从个体能力到制度兜底再到风险雷达，构成一条完整的传导链。
18条任务，每条写清了落款单位
国办发〔2026〕29号 六大方向传导链
「失业缓冲期」为什么是全文最具体的一句话
其余17条的主体动词是「鼓励」「推进」「完善」「健全」，方向性表述为主。唯独「鼓励企业探索建立失业缓冲期制度」给出了动作序列。大规模推广前告知，提供转型转岗支持。三个动作，有先后、有对象、有时点。在全国性文件中，这个颗粒度不常见。
你公司如果今年已在做大规模AI替代，对照这条看：告知做了没有，转岗方案写进合同了没有。
"缓"字背后的产业分层
国办发〔2026〕29号内部藏着两条轨道。「优先在劳动力紧缺、环境高危等岗位推广应用」是油门，「稳慎把握人工智能在劳动密集型行业应用的范围、力度和节奏」是刹车。同一家厂，质检线可能全速上，客服线得慢慢来。
劳动密集型行业为什么被单独拎出来说
一台质检机器人上一班，直接对应2到3个工位消失。500人工厂里替代弹性接近1，机位拆了就是空了。文件用「稳慎」而非「禁止」，留了三个可调变量：范围（先自动化哪段工序）、力度（替换还是辅助）、节奏（多快铺开）。「失业缓冲期」就是节奏变量的落地形态。
对比另一侧：放射科缺医生，加一套AI辅助诊断，一人管三人量，岗位没消失，密度上去了。两类场景分开写，不混在一起。
双轨产业分层逻辑
中小微企业那条"低成本、轻量化"路径长什么样
「低成本、轻量化」挂在「稳岗扩岗」下面，意思是用AI帮小厂留人、扩岗，不是替小厂裁人。落到30人电商运营团队：一个人接通义千问或文心一言的API跑商品文案和客户应答，按调用量付费，不建自训模型团队。80人制造工厂：质检上云SaaS，按工位月租，不招算法工程师。配套的「微专业」和订单式培训，让腾出的产能去接新订单，不是去抢新工位。
对从业者和HR的三张直接账单
文件第三、四、五方面落到具体人头上，是三张时间线错开的账单。一张在招聘端，一张在执法端，一张在培训端。三张不会同时到账。
新职业扩编：数字化管理师、AI训练师在开放哪些口子
文件用了「扩大需求」四个字，并列数字化管理师、人工智能工程技术人员、人工智能训练师。拆开看，三个口子性质不同。
数字化管理师管的是已有业务流的数字化重构，不是从零搭系统。一家跑了八年 ERP 的制造企业，把采购审批从五级签字压到两级、把质检数据从纸质单据迁到线上看板，背后需要一个既懂业务又懂数据管道的人。这个口子开在生产运营部门，不在 IT 部门。
AI训练师在 2026 年的实际工作，大概率不是训练基础模型。多数中型企业跑的是 RAG 管线或开源模型微调，「训练」更接近数据标注、评估集构建和 prompt 工程的组合。文件把它列为独立工种，等于承认了「会用 AI 工具」和「管得住 AI 工具输出质量」是两个岗位。
对 HR 来说，这两个角色没法拿现成 JD 套。文件提的「校企联合培养」解决的是供给侧长周期问题，培养周期至少一学期。实话说，第一波填坑的还是内部转岗。
招聘端算法歧视治理条款，落到执行层缺什么
文件要求「依法治理因滥用算法等造成的就业歧视、虚假招聘，以及侵犯隐私权、休息权」。四个违规类型，一个动词「治理」，没写谁来审、多久审一次、罚多少。
对比欧盟 AI Act 的分级机制，高风险系统做合规评估，处罚按严重度阶梯走。这边「依法」把球传给了现有劳动监察体系，而劳动监察的技术审计能力是短板。第一个被端出来的案例大概率是显性的：一个招聘聊天机器人按姓名或毕业年份筛简历。隐性歧视，比如排名算法按邮编降权某片区职位曝光，要等专门的执法细则落地才有人管。
高校"微专业"和专项培训，谁先受益
文件要求普通高校和职业学校「增加人工智能微专业布点」，同时实施专项培训计划，对象列了三类人：高校毕业生、农民工、失业人员。
受益顺序是清晰的。高校毕业生已经在体系里，微专业挂在现有学位旁边，按学期节奏走，成本由学校和企业分摊。农民工和失业人员走「项目制、订单式」通道：时长更短、企业补贴，但地理上绑死在 AI 落地的那座城市。一线城市服务业从业者先拿到课时，县域劳动力市场要等政策扩散。
「订单式」的意思是用人单位出题目。一家正在部署 RPA 的工厂订了六周「数字化工具操作」课，教的是按按钮，不是判断什么时候该停机。这是缓冲，不是转型。
三张账单的到账时间线
落点：对照18条检查你自己的节奏
管团队的人：部署AI前该补哪几步流程
如果你三个月前签了AI客服上线的审批单，当时没跑影响评估，18条现在给了你一份可以倒填的清单。操作条款集中在第8条「建立人工智能应用就业风险评估机制」和第11条「鼓励企业探索建立失业缓冲期制度」。拆开来看，是三步：
第一步，写一页影响备忘。在你把任何工具从试点扩到全组之前，列出哪些岗位会被吸收、哪些会缩减、时间线多长。两页以内。塞不进去说明你还没想清。
第二步，提前告知。失业缓冲期条款的原文是「在大规模推广应用人工智能过程中提前告知受影响劳动者」。注意是「过程中」，不是「裁完之后发个内部邮件」。如果下一次扩围涉及三人以上，在工具切生产之前安排一对一谈话，把日期、受影响岗位、可转方向三样说清楚。
第三步，把转岗路径落到具体岗位。「加大转型转岗就业支持力度」这句话要能回答一个追问：被替代的那个人，内部哪个角色能接住，中间缺的技能靠什么补。答不上来，这步就是空转。
边界要讲清。第18条的「完善就业监测预警和风险应对体系」意味着数据层是持续积累的。200人的公司今天不会因为跳过第一步被罚款，但当你的行业出现一批标注为AI替代的裁员时，预警阈值会触发，你的公司进入监测数据集。有没有跑过清单，事后都会被看到。
AI规模化部署前的三步门禁
面临转岗的人：培训入口和失业缓冲怎么接
培训侧有三个入口，时效各不一样。
人工智能专项培训计划由地方人社局组织，形式是项目制、订单式的数字化技术技能培训，紧贴市场需求。你不需要处于失业状态才能报名。查你所在城市的人社局官网，找2026年度AI培训招生日历，通常每季度开放一期。这是离你最近的一条路。
高校「微专业」和交叉学科方向是第二条。2026年有多所高校新增了AI相关微专业布点，学制短于完整学位，可以叠在现有学历上。适合入职五年内、想跨到AI训练师或数字化管理师方向的人。公共平台上的公益通识课程是第三条，门槛最低，但定向性也最弱，适合补词汇和基础概念，不适合直接对接岗位。
失业缓冲期目前仍处于「鼓励企业探索」阶段。截至2026年10月下旬，没有省份公布实施细则。你现在能靠住的底线是「按规定对参保失业人员及时足额发放失业保险待遇」。缓冲期是天花板，还没封顶。
文件里有一处判断值得单独拎出来：它把「新职业需求扩编」和「传统岗位替代」放在不同段落处理。如果你的岗位在替代那一栏，最经济的动作不是等缓冲期落地，而是在替代落到你工位之前，把技能往数字化管理师、AI训练师、AI工程技术人员这几个正在扩编的方向上挪。流程还没跑完，工具已经上线了，但你的转岗窗口不会等你把文件读完。
参考文献
延伸入口
- 原文归档：tobemagic.github.io/ai-magician…
- 公众号：计算机魔术师
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Your company got into the AI client's uniform, cut two seats and HR left without any formal notice. For the first time, the Office of the Secretary-General has written this into a system: mass outreach must be informed in advance and a buffer period must be granted.
What do you care about these 18 articles?
On 23 October, the Office issued its Opinions on Adapting to Artificial Intelligence Development Helping to Build Employment-Friendly Development. No. 29), set aside for October 16. The core logic of the 18 mandates in 6 is two things. The jobs that AI can create must be created and the alternatives must be buffered.
Open in 6 directions: creation, substitution, training, services, safeguards, early warning
Order is the priority. Before the first "enhanced job creation effect", the position was not to stop, it was to create a new job and move on. The second item, “Constructing an employment-friendly path to promotion”, was replaced by the qualifier, “Surely grasping the scope, strength and tempo of artificial intelligence applications in labour-intensive industries and minimizing the impact on practitioners”. The latter four components, in turn, are training, employment services, security of interests, monitoring of early warning, from individual capabilities to systems to risk radar, which forms a complete chain of transmission.
18 missions, each of which clears the down payment unit.
National office (2026) Number 29, six-way chain.
Why is "unemployment buffer" the most specific sentence in the whole text?
The main verb for the remaining 17 articles is "encourage" "promoting" "improvement" and "sound" in the main direction. The only action sequence is given by "Encourage enterprises to explore the establishment of a buffer period for unemployment". Provide transition support prior to large scale roll-out. Three moves, sequenced, targeted, sometimes. This particle size is not common in national documents.
If your company is already working on a large-scale AI replacement this year, then let's see if you've done it.
The industrial layer behind the word "slow"
National office (2026) Two tracks are hidden inside 29. "Priority in the use of jobs with a labour force shortage and high environmental risks" is the oil door, and "sure of the scope, strength and rhythm of artificial intelligence applications in labour-intensive industries" is the brake. At the same plant, the test line may be full speed and the passenger line may have to be slow.
Why are labor-intensive industries being carried alone? Come out.
The first class of a mass-checking robot directly corresponds to the disappearance of two to three jobs. The alternative elasticity in the 500-person factory is close to 1 and the space is empty. The document leaves three variable adjustable variables, using the word "conserve" rather than the word "prohibited": scope (which process is automated first), strength (replacement or support), rhythm (how fast to spread). The unemployment buffer period is the landing pattern of the rhythm variable.
On the other side of the equation, there is a shortage of radiologists, plus a set of AI-assisted diagnostics, one person, three persons, no job, density. The two scenarios are separated and not mixed.
Double-track industry hierarchy logic
What's the "low-cost, light-quantization" path for MSMEs?
The term "low cost, light quantization" is attached to the term "stable job expansion" and means "sheltering and expanding small factories with AI, not laying off small factories." (c) 30-person electrician operating team: one person answers Qwen or Ernie API running merchandise files and client responses, no self-training model team based on paid call. 80-man manufacturing plant: Mass inspection on cloud SaaS, monthly rent, no algorithm engineer. A combination of “micro-professional” and order-type training has allowed the released capacity to take new orders, not to rob new jobs.
Three direct bills for practitioners and HR.
The third, fourth and fifth aspects of the document fell on specific individuals, with three staggered bills. One at the recruitment end, one at the law enforcement end and one at the training end. Three won't arrive at the same time.
New career expansion: Digital managers, AI trainers who are opening up
The document uses the word "extension of demand" and lists digital managers, artificial intelligence engineering technicians, and artificial intelligence trainers. Open it, three mouths are different in nature.
The digital manager manages the digital re-engineering of existing business flows rather than a piecemeal system. An ERP manufacturing company that has been running for eight years has pressed procurement approvals from level 5 signatures to level 2, moving quality data from paper documents to online boards, requiring a person who knows both business and data conduits. It's in the business sector, not in the IT sector.
The actual work of the AI trainers in 2026 was probably not a basic training model. Most medium-sized enterprises run RAG pipelines or open source models fine-tuned, "training" closer to the combination of data labels, assessment set construction and prompt projects. The document lists it as a stand-alone job, which is tantamount to recognizing two jobs as "Ai tools" and "Ai tools output quality".
For HR, these two characters can't make a JD. The paper refers to “co-education” that addresses the supply side of the cycle, which is at least one semester. To be honest, the first wave of landfills is still internalized.
Recruitment end algorithms discriminate against governance provisions, to what is missing at the executive level
The document calls for the "legal regulation of employment discrimination, false recruitment, violation of privacy, right to rest, etc. due to the misuse of algorithms". Four types of violation, one verb, "governance", does not write who comes to court, how often they are tried, how many penalties.
Compared to the EU AI Act classification mechanism, the high-risk systems are evaluated for compliance and the penalties follow a serious step. The ball is passed to the existing labour inspection system "in accordance with the law" and the technical audit capacity of labour inspection is short-lived. The first case that was taken out was presumably obvious: a recruitment chat robot sifted by name or year of graduation. Inhibited discrimination, such as ranking algorithms that are exposed to a post in a district subject to postage, is subject to a specific rule of law enforcement.
"Micro" and specialized training in higher education, who benefits first?
The document calls for general higher education and vocational schools to “addition of artificially intelligent micro-profiling points” and special training programmes for three categories of people: graduates of higher education institutions, migrant workers and unemployed persons.
The order of benefits is clear. Graduates of higher education institutions are already in the system, and microprofessionals are attached to existing degrees, at the pace of the semester, at the cost of schools and enterprises. Farmers and the unemployed take the "project, order" route: longer and shorter, enterprise-subsidized, but geographically tied to the city where AI landed. First-line urban service providers are given lessons first, and the district labour market is spread.
The word "order" means the name of the employer. A factory that is deploying RPA has scheduled six weeks of "Digital Tool Operations" to teach buttons, not when to stop. It's buffer, not transition.
Timeline for three bills.
Droppoint: Check your own rhythm against 18
Team people: Which steps should be completed before AI is deployed
If you signed an AI passenger service on-line three months ago without an impact assessment, 18 articles now give you a list that can be filled out backwards. The operative provisions focus on article 8, on the establishment of an employment risk assessment mechanism for the application of artificial intelligence, and article 11, on encouraging enterprises to explore the establishment of a buffer period for unemployment. Open it. It's three steps:
The first step is to write a page to influence the memo. Before you expand any of the tools from the pilot to the group, list which posts will be absorbed, which will be scaled down, and how long the timeline will be. Within two pages. It doesn't fit. It means you haven't figured it out yet.
Step two, advance notice. The provision for a buffer period for unemployment reads “to inform the affected workers in advance of the mass introduction of artificial intelligence”. Attention is given to "the process", not "the internal mail after the cut." If the next extension involves more than three persons, a one-to-one interview is arranged before the tool is produced, with the date, the position affected and the direction to be taken.
Step three, move the transfer path to specific positions. The phrase "enhanced support for the transition to employment" would answer the question: who is replaced, what role inside, and what skills are missing. I can't answer that. This step is empty.
The boundaries are clear. Article 18, "Improving early warning and risk management systems for employment monitoring", means that data layers are constantly accumulating. The 200-person company will not be fined today for skipping the first step, but the warning threshold will be triggered when your industry shows up as an AI-substitute layoffs, and your company enters the monitoring data set. If you run through the list, you'll be seen after.
Three-step ban on the AI scale prior to deployment
People facing a transfer: training entrances and unemployment buffers Answer.
There are three entry points on the side of the training and the time limits vary.
Specific training programmes for artificial intelligence are organized by the Local People ' s Social Services Bureau in the form of project-based, order-based training in digital skills, which is tailored to market demand. You don't have to be unemployed to register. Check the network of social welfare officers in your city and find an AI admission calendar for 2026, usually open quarterly. This is the closest road to you.
“Microprofessionals” and cross-cutting disciplines in higher education are the second. In 2026, a number of institutions of higher learning added a micro-professional site for A related subject, with a shorter degree than a full degree, which could be folded into existing academic qualifications. It is suitable for a person who, within five years of entry on duty, wishes to move into the direction of an AI trainer or digital manager. The public interest course on public platforms is the third and lowest threshold, but also the least targeted, suitable for supplementing terminology and basic concepts and not for direct job matching.
The unemployment buffer period is still at an “encouraging enterprise to explore”. As of late October 2026, no province had issued implementing regulations. The bottom line on which you can now rely is "the full and timely payment of unemployment insurance to the unemployed who are insured". The buffer period is the ceiling, not the top.
One of the findings in the document is worth showing: it treats "new job needs expansion" and "traditional job substitution" in different paragraphs. If your position is in the substitute column, the most economical move is not to wait for the buffer period to fall, but to move the skills in the direction of digital managers, AI trainers, AI engineers and technicians before replacing you. The process is not over, the tools are online, but your shift window will not wait for you to finish reading the files.
References
Extension entrance
- Original archive: tobemagic.github.io/ai-magician...
- WeChat Official Account: Computer magician
Auto-fetched and translated by AI Atlas. All rights belong to the original author.