AI 银行信贷全景图 · 一次信用的定价 · 2026-07 · 钱的机器本身Banking & Credit × AI · Pricing One Act of Trust · Jul 2026 · The Machinery of Money Itself
对一次信任报价
Quoting a price on trust
把银行拆到最小单位,它只做一件事:对一次信任报价,然后承担报错价的代价。AI 已经吃透了这条链上标准化、高频、可量化的环节——开户核身、支付反欺诈、信贷初筛、语音催收;却啃不动「承担后果」的部分:穿越周期的风控验证、最后贷款人、挤兑时的信心
Strip a bank to its atom and it does one thing: quote a price on an act of trust, then bear the cost of quoting wrong. AI has eaten through the chain’s standardised, high-frequency, quantifiable links — onboarding, payment anti-fraud, credit screening, voice collections — and cannot chew the consequence-bearing part: risk models proven across a full cycle, the lender of last resort, confidence in a bank run
本图相信:「完成时」比「进行时」更有解释力。柜员消亡是一部放完的电影——离柜率 2008 年 37.93% 爬到 2020 年首破 90%用了十余年;网点与 ATM 双双 2018 年前后见顶;美国分支 2009 年达峰。替代的节奏(三十年渐进)、抵抗(报批/县域刚需/ATM 悖论)、去向(转岗而非清零)——全有现成答案可考(A/B)。This map believes: the perfect tense explains more than the present tense. The teller’s demise is a film already screened — the off-counter rate climbed from 37.93% (2008) to past 90% in 2020 over a decade; branches and ATMs both peaked around 2018; US branches peaked in 2009. The tempo (thirty gradual years), the resistance (approval regimes, county-level need, the ATM paradox) and the destination (redeployment, not erasure) — all have answers on the record (A/B).
本图不相信:「普惠」与「掠夺」是两套系统。同一套机器学习风控,既造出「3 分钟申贷、1 秒放款、0 人工」的普惠效率(每笔小微贷运营成本约 2.3 元 vs 传统约 2000 元),也在 24%–36% 区间用「日息/双融担/会员费」话术,把最脆弱的多头共债人群深度杠杆化——效率与掠夺是同一枚硬币,用的是同一套模型(B)。It does not believe «inclusion» and «predation» are two systems. One and the same ML risk stack built the «3-minute application, 1-second disbursal, zero humans» efficiency (≈¥2.3 operating cost per micro-loan vs ≈¥2,000 traditionally) — and, in the 24–36% band, leverages the most fragile multi-debt borrowers through daily-rate, dual-guarantee and membership-fee scripts. Efficiency and predation are one coin, minted by one model (B).
本图管的是「钱的机器本身」(银行/消金/助贷/支付/征信反欺诈/催收/不良处置);理财投顾归wealth。主脊:一笔钱的旅程七环节(开户 KYC→存款负债→支付清算→信贷审批→贷后管理→催收→不良处置),④⑥标红;三条结构带:普惠 vs 利率现实的诚实层、柜员消亡的完成时样本、支付跳代与货币地缘;判断层:消费者端真实年化 IRR ＞ 一切营销话术,机构端负债成本与风控能力 ＞ 规模冲动。姊妹:财税→ledger,稳定币→crypto,柜员转岗→work,责任基建→liability。
This map covers the machinery of money itself (banks, consumer finance, loan facilitation, payments, credit bureaus and anti-fraud, collections, distressed assets); wealth management belongs to wealth. The spine: one sum’s journey in seven links (KYC → deposits → payments and clearing → credit underwriting → post-loan management → collections → distressed disposal), ④ and ⑥ flagged; three bands: the inclusion-vs-rates honesty layer, the teller’s perfect-tense sample, the payments leapfrog and monetary geopolitics; the judgment: for consumers, the true annualised IRR ＞ every marketing script; for institutions, funding cost and risk capability ＞ the urge to scale. Siblings: accounting → ledger, stablecoins → crypto, teller redeployment → work, liability infrastructure → liability.
90.88%
2020 年中国银行业离柜率——首破 90%(2008 年仅 37.93%);农行柜员 2016–2020 从 14.8 万降至 8.3 万;ATM 2018 年峰值 111.08 万台→2023 年 84.54 万——前 AI 时代自动化的完成时样本(A/B)China’s banking off-counter rate in 2020 — first past 90% (37.93% in 2008); one major bank’s tellers fell from 148k to 83k over 2016–20; ATMs peaked at 1.11M in 2018, down to 845k by 2023 — the perfect-tense sample of pre-AI automation (A/B)
2.3 元/笔
网商银行每笔小微贷的平均运营成本(vs 传统银行人力约 2000 元/笔;「3 分钟申贷、1 秒放款、0 人工」为该行自身口径 D);模式已复制到 50 家机构、服务 6800 万小微商家——信贷审批是全图 AI 密度最高的核心工厂(B/D)MYbank’s average operating cost per micro-loan (vs ≈¥2,000 of human cost traditionally; the «3-1-0» claim is the bank’s own basis, D); the model replicated to 50 institutions serving 68M small merchants — underwriting is the map’s densest AI factory (B/D)
24–36%
助贷灰色定价区间:「日息万五」=名义年化 18%+;等额本息真实 IRR ≈表面利率近 2 倍;「双融担」把 36% 包装成 24%——2025 助贷新规(10-1 施行)以白名单+增信费计入综合成本直指这套话术(A/B)The grey pricing band of loan facilitation: «0.05% daily» reads 18%+ annualised; equal-instalment repayment runs a true IRR near 2× the sticker; the dual-guarantee structure dresses 36% as 24% — the 2025 facilitation rules (in force Oct 1) target exactly this playbook with whitelists and all-in cost accounting (A/B)
0.48 折
个贷不良批量转让的平均折扣率(本金回收率 7.3%,个别资产包低至 0.17 折);2024 年成交 1583.5 亿元、同比 +64%——未穿越完整周期的模型,正在真实周期里交学费(A/B)The average discount on bulk transfers of non-performing personal loans (7.3% principal recovery; some pools at 1.7% of face); 2024 volume ¥158.35B, +64% — models never proven across a full cycle are now paying tuition inside one (A/B)
⚠️ 口径裁判(先读):① 「离柜率」(2014–2021)与「电子渠道分流率」(2022 起)是两套口径不可连线,「突破 90%」应锚定 2020 年;② 美国柜员峰值有全口径(~60 万,2007–2010)与抽样口径(39.8 万,2014)之分,勿混用;③ 居民杠杆率 62% 与 70%+ 存测算分歧;④ 数字人民币采用 2025 年 9 月末官方口径(14.2 万亿),「11 月末 16.7 万亿」未获一手源;⑤ 「近 60% 催收由 AI 负责」为媒体财报解读待 IR 复核;「0 人工干预」为机构自身宣传口径;⑥ 对流传表述的更正:所谓「贷后催收 JR/T 行业标准」并不存在,实为三份递进文件(2024 团标→2025 国标 GB/T 45251→2026 中银协试行指引);⑦ AI 外呼市场「80 亿→268 亿」等来自选型软文已降级不采;⑧ 四路研究交叉核验(关键数字 ≥2 来源);渗透%为编辑估值。
⚠️ Basis rulings (read first): ① the «off-counter rate» (2014–21) and the «e-channel diversion rate» (from 2022) are two bases that cannot be spliced — anchor «past 90%» on 2020; ② US teller peaks split between the full basis (~600k, 2007–10) and the survey basis (398k, 2014) — never mix; ③ household leverage splits 62% vs 70%+ by estimate; ④ e-CNY uses the official end-September 2025 basis (¥14.2T); the «¥16.7T by November» claim lacks a primary source; ⑤ «~60% of collections by AI» is a press reading of one filing, pending IR verification; «zero human intervention» is the institution’s own marketing basis; ⑥ a circulating correction: the alleged «JR/T collections standard» does not exist — the reality is three escalating documents (the 2024 association standard → the 2025 national standard → the 2026 banking-association trial guideline); ⑦ AI-calling market sizes from vendor-selection advertorials are demoted and unused; ⑧ four research tracks cross-verified (key figures ≥2 sources); penetration %s are editorial.