AI 保险全景图 · 一张保单的生命周期 · 2026-07 · 七段流水线 + 制度暗线Insurance × AI · The Life of One Policy · Jul 2026 · Seven stations + the hidden line

感知归 AI,裁量归人 Perception to AI, judgment to people

保险的原子动作只有一个:把一份不确定性从个体转移到池子里,并给它标一个价。这个动作拆开是七段流水线——精算、渠道、核保、承保、保全、理赔、再保。理解 AI 对保险的冲击,关键不是问「AI 能不能做保险」,而是问哪一段的本质是「从数据到判断」(AI 的主场),哪一段的本质是「从判断到承诺」(AI 的禁区)。流水线之下还压着一条制度暗线:负债成本——它不出现在任何一段流程里,却决定了整条线上所有人的命运:一家被利差损压垮的公司,无论它的 AI 有多先进,都会死。 Insurance has one atomic act: move a unit of uncertainty from a person into a pool, and price it. Unfolded, that act is a seven-station line — actuarial design, distribution, underwriting, issuance, servicing, claims, reinsurance. To understand AI here, don’t ask whether AI can do insurance; ask which stations are «data to judgment» (AI’s home ground) and which are «judgment to promise» (AI’s forbidden zone). Beneath the line runs a hidden institutional wire: the cost of liabilities — absent from every process step, yet deciding everyone’s fate: a company crushed by negative spread dies no matter how good its AI is.

本图相信:车险外观定损是保险业唯一可以说「AI 已经赢了」的战场(分钟级已成常态,一家供应商部署在 125 家险企);秒级核保与保全客服的替代真实发生——安联已据此裁员 1500–1800 人;专业化顾问>人海战术已被实证:五大上市寿险人数砍掉七成,人均新业务价值反涨 43.3%(A/B)。This map believes: exterior auto-damage appraisal is insurance’s one battlefield where AI has plainly won (minutes are routine; one vendor runs in 125 insurers); second-level underwriting and servicing substitution is real — Allianz cut 1,500–1,800 jobs on it; and specialist-beats-headcount is proven: the top five life insurers cut 70% of agents while NBV per agent rose 43.3% (A/B).
本图不相信:「AI 颠覆保险」的整体叙事——78% 的险企在用生成式 AI,仅 4% 规模化;某巨头 218 个 AI 原型只有 5 个投产;不到 5% 的险企能披露 AI 的具体财务影响;过去十年保费翻倍,费用率只降了 1 个百分点。也不相信把代理人塌方全记在 AI 头上——912 万→212 万的三击顺序是清虚→模式失效→报行合一+AI,AI 是补刀,不是主犯(A/B)。It does not believe the wholesale «AI disrupts insurance» story — 78% of insurers use generative AI and only 4% run it at scale; one giant turned 218 prototypes into 5 production systems; under 5% of insurers can state AI’s financial impact; a decade of doubled premiums cut the expense ratio by one point. Nor that AI felled the agent army — the 9.12M→2.12M collapse came in three blows: registry purges, model failure, then fee reform plus AI. AI was the finisher, not the killer (A/B).

主脊:一张保单的七站(精算→渠道→核保→承保→保全→理赔→再保),②⑤标红;三条结构带:利差损暗线、投保易理赔难的机制解剖、试点丰收损益表歉收;判断层:AI 能做感知/分类/检索/批准,不能做裁量/承诺/共情——这条能力边界与保险的责任边界高度重合,而且不是巧合。姊妹:需求侧互为因果→wealth,AI 责任险新物种→liability,智驾重构→car,销售职业的一般性重构→sell The spine: one policy, seven stations (actuarial → distribution → underwriting → issuance → servicing → claims → reinsurance), ② and ⑤ in red; three bands: the negative-spread wire, the buy-easy-claim-hard mechanism, pilot harvest vs P&L famine; the judgment: AI does perception, classification, retrieval and approval — not discretion, promises or empathy; that capability line coincides with insurance’s liability line, and not by accident. Sisters: wealth, liability, car, sell.

912 → 212.6 万
中国保险代理人 2019 峰值→2026Q1,减少近 700 万人、降幅近八成——中国单一行业最大规模的销售劳动力收缩;塌方发生在 AI 之前,AI 是第三击(A/B)China’s insurance agents, 2019 peak to Q1-2026: down nearly 7 million, almost 80% — the largest sales-labour contraction in any single Chinese industry; the collapse predates AI, which struck third (A/B)
只批,不拒
宏利核保 AI 的设计:只能批准、不能拒保——批准的后果由保司承担,拒保的后果由消费者承担。这条线画在哪里,是理解 AI 在保险业边界的钥匙(B)Manulife’s underwriting AI can approve but never decline — approval’s consequences fall on the insurer, denial’s on the consumer. Where that line sits is the key to AI’s boundary in insurance (B)
78% vs 4%
用生成式 AI 的险企占比 vs 实现规模化的占比;十年保费翻倍、费用率仅降 1 个百分点——「人人都在用,没人算得清收益」是当下最准确的行业切片(B)Insurers using generative AI vs those running it at scale; a decade of doubled premiums cut expenses one point — «everyone uses it, nobody can count the gains» is the industry’s most accurate cross-section (B)
90%
日本破产寿险的责任准备金削减比例,叠加预定利率被强制降到 1%——1990 年按 6.25% 买的三十年保单,2000 年只剩名义躯壳,而这是「合法的」(A)The reserve haircut at failed Japanese life insurers, stacked with guaranteed rates forced down to 1% — a thirty-year policy bought at 6.25% in 1990 was a nominal shell by 2000, all of it «legal» (A)
口径裁判:中国险企的 AI 效果数字(93% 秒级核保、44 亿次 AI 服务、响应提速 260 倍)几乎全部为公司自述,无独立第三方审计,分母与计数口径大多未公开;代理人 912 万(代理制)与 973 万(全部销售人员)是包含关系,212.6 万系 67 家偿付能力报告合计口径、与往年不完全可比;穿戴设备「死亡率降 76%」未校正自选择偏差,不能当因果证据;车险欺诈「千亿规模」为媒体测算;道交法修订草案「智驾专项责任险」细节仅见行业媒体解读、未见官方文本。Refereeing the numbers: Chinese insurers’ AI metrics (93% instant underwriting, 4.4B AI service calls, 260× faster response) are almost entirely self-reported and unaudited, with denominators undisclosed; the 9.12M (agency) and 9.73M (all sales staff) figures nest rather than conflict, and the 2.126M sums 67 solvency reports — not fully comparable across years; the wearables «76% lower mortality» is uncorrected for self-selection and proves no causation; the «hundred-billion» fraud figure is a media estimate; the draft road-law’s «autonomy liability policy» exists only in trade-press readings, not official text.
中心装置 · 912 万→212 万:一场发生在 AI 之前的塌方The central device · 9.12M → 2.12M: a collapse that predates AI
任何把销售岗位萎缩全部归因于 AI 的叙事,在保险业这里会被证伪——这是全景图系列里最重要的「反 AI 决定论」对照组。Any story blaming AI alone for shrinking sales jobs is falsified here — the atlas’s most important control group against AI determinism.
塌方:三击的真实顺序The collapse: three blows, in order
2014–2019 年代理人从 325 万膨胀到 912 万,五年翻近三倍。然后:第一击「清虚」(2019–21),监管清核执业登记、剔除影子人力;第二击「模式失效」(2020–23),自保件熟人市场耗尽、新人留存率不足 30%、金字塔增员塌方、疫情切断面访;第三击「报行合一+AI」(2023–),佣金费率平均降约 30%,AI 承接线索筛选、需求分析、方案比价——代理人垄断的信息中介职能被接管。2026 年一季度末:212.6 万人,平均月收入约 6000 元(2019 年为 8000–10000 元)。Agents ballooned from 3.25M to 9.12M in five years to 2019. Then: blow one, the purge (2019–21) — registry audits strip phantom headcount; blow two, model failure (2020–23) — the self-purchase and acquaintance market exhausts, rookie retention drops under 30%, the pyramid recruiting engine collapses, the pandemic cuts face time; blow three, fee reform plus AI (2023–) — commissions fall ~30% while AI takes over lead scoring, needs analysis and product comparison, the information brokerage agents once monopolised. Q1-2026: 2.126M agents averaging ¥6,000 a month (¥8–10k in 2019).
重建:人数砍七成,产能翻倍以上The rebuild: 70% fewer people, twice the output
五大上市寿险代理人 2023 年 142.38 万→2025 年 133.34 万,企稳但回不去了。产能却在涨:平安人均新业务价值 +43.3%、月均收入 10395 元;太保月人均首年保费 +30.4%;中国大陆 MDRT 会员 2.1 万人、全球第一。被砍掉的 700 万人做的恰是不需要温度的工作(找线索、讲产品、递单子)——那正是 AI 与线上化的正面靶心;留下的人做的是另一种工作The top five listed insurers’ agents went 1.42M (2023) → 1.33M (2025): stabilised, never returning. Output climbs instead: Ping An’s NBV per agent +43.3% at ¥10,395 monthly income; CPIC’s first-year premium per agent +30.4%; mainland China leads the world with 21,000 MDRT members. The 7 million cut did the warmth-free work — lead-finding, product-pitching, form-ferrying — precisely AI’s and the internet’s bullseye; those who remain do a different job.
判词VerdictAI 是补刀,不是主犯——但 AI 决定了塌方之后重建成什么样:不是人海,而是「AI 做信息中介 + 人做信任中介」的双层结构。死亡话题的销售永远是人对人的工作,但「人对人」的人数会持续下降、单价会持续上升——这不是岗位存亡问题,是岗位性质的置换。AI finished the job, it didn’t start it — but AI dictates what gets rebuilt: not a human wave, a two-layer structure where AI brokers information and humans broker trust. Selling around death stays person-to-person forever; the number of those persons keeps falling as their price keeps rising — not a survival question but a substitution of the job’s very nature.
Reading the Map

从这张图带走的五条规律Five patterns to take away

立场声明:本图为保险产业结构分析,不构成任何保险配置建议、投资建议或法律意见;产品举例全部停留在结构层,不做具体产品推荐或优劣评价;险企 AI 效果数字凡为公司自述均已逐处标注;日本破产处置数据来自金融厅官方表格;消费者「该怎么配保险」归 wealth 图,本图不越界;市场地图 ≠ 推荐背书;仅供研究与批判性参考。Stance: a structural read of the insurance industry — not allocation advice, investment advice or legal opinion; product examples stay structural, with no recommendations or rankings; every self-reported AI metric is flagged as such; Japanese failure data comes from the FSA’s official tables; how consumers should buy insurance belongs to the wealth map, and this map does not trespass; a market map is not an endorsement; for research and critical reference only.