1. 先找引荐路径，再考虑写信。在现有客户、前同事、投资人里找一条通往对方的路。引荐信不靠文案取胜，它根本不在同一个池子里排队——噪音变多之后，关系存量的回报只会更高。
2. 让 AI 盯触发事件，别让它替你写那封信。招人、融资、换系统、合规期限、新上任的负责人——监控与摘要它很擅长。判断「哪个是真的现在就疼」，留给你自己。
3. 每封信至少一句模型写不出来的。你上周跟他同行聊到的具体做法、你在他们产品里卡住的那一步、他上个月发言里的一个数字。判据很简单：这句话换一家公司还成立，就不算。
4. 发送量往下砍，每封的准备时间往上加。冷邮件与外联那期的结构照旧有效——变的是「量」这一维不再能买到优势，因为它对谁都便宜了。
EN AI compresses "find five hundred plausible targets and write each of them something that looks thoughtful" from two weeks into an afternoon. So does it for your competitors. Inbox capacity didn't change and the number of senders did, so the equilibrium only points one way: reply rates per message fall until sending one more is barely worth it. What AI changed isn't how many you should send, it's where the two saved weeks should go. The evidence on personalisation is worth reading carefully here. Sahni, Wheeler and Chintagunta (2018, Marketing Science) ran large field experiments and found that simply adding the recipient's name to the subject line raised open rates by around 20%, lowered unsubscribes and produced more leads — and they called it noninformative content, because it adds no information at all. It worked because it was scarce. Once everyone can generate a bespoke-looking opener, that channel decays back toward baseline; the direction is certain, the magnitude isn't predictable. Platforms are also putting the cost back: since February 2024 Gmail and Yahoo require bulk senders (roughly 5,000+ messages a day) to authenticate with SPF, DKIM and DMARC, offer one-click unsubscribe, and hold spam complaints under 0.3%. Four places to put the saved time. Look for a referral path before you consider writing at all — a warm introduction doesn't compete in the same queue. Have AI watch for trigger events (hiring, funding, a system migration, a compliance deadline, a new person in the seat) and keep for yourself the judgement of which trigger actually hurts right now. Put at least one line in every message a model couldn't produce; the test is whether the sentence would still be true for a different company, and if it would, it doesn't count. And cut volume while raising the preparation time per message, because volume is the one dimension that no longer buys you an edge.
机制：拥挤外部性下的军备竞赛。边际成本趋近零、注意力总量固定时，「我再多发一点」永远是个体最优，加总起来却压低所有人的回复率，直到刚好覆盖那点残余成本。个性化文案同理——它有效是因为稀缺，不是因为携带信息；谁都能生成，租金就消失。
边界：不是外联死了，是「批量 + 泛化」这一格死了。噪音变多恰恰抬高了另一格的价值——少而准、有真实触发事件、发信人可核查。在很窄的市场里，五十封精准的信仍能撑起一条管道。
1. 交互式机器人一开口就说清楚。「我是 XX 的智能助理，能查价格和排期；要谈方案我直接转给负责你的顾问。」把披露和「我能干什么、干不了时转给谁」写在同一句里——单说「我是机器人」是扣分，加上出口才是给确定性。
2. 越主观、越有后果的环节，越要有人露面。报价、坏消息、异议、谈判、出事后的第一通电话，全部由人来。买家在这里评估的不是答案对不对，是「出问题时他会不会站我这边」——正是标签一贴上就关掉的那套推断。
3. 别把 AI 藏在人后面冒充人。被拆穿一次，之前所有真话一起被重估；欧盟 AI 法案第 50 条也明确要求让人知道自己正在与 AI 系统交互（2026 年 8 月适用）。判据还是销售的伦理边界那期那条——说破了还成立吗。
4. 客户越觉得自己情况特殊，越要人接手。人越认为自己的案例独特，越抗拒交给算法。听到「我们跟别家不一样」，正确回应是人露面 + 问具体差异，不是甩一份更详细的自动化报告。
EN One number here is uncomfortably large. Luo, Tong, Fang and Qu (2019, Marketing Science) ran a field experiment at a financial-services firm: chatbots that didn't disclose their identity sold about as well as proficient human agents and far better than inexperienced ones — but disclosing bot identity before the conversation cut purchase rates by roughly 80%. You cannot run that backwards. Hiding it is both a legal problem (the EU AI Act's Article 50 transparency duties require that people be informed they are interacting with an AI system, applying from August 2026) and a trust problem: get caught once and every true thing you said gets repriced. The usable reading is different — disclose, and move where the machine stands. There's a mechanism underneath. Believing a human is on the other side recruits the mentalizing network — medial prefrontal cortex and right temporo-parietal junction. Krach and colleagues (2008, PLoS ONE) found activity in those regions increased with how human-like the game partner was, from computer through robot to person; earlier imaging work (McCabe et al., 2001, PNAS; Gallagher et al., 2002, NeuroImage) found the same contrast between playing a person and playing a computer. Small samples, treat as direction rather than a settled quantity. But it explains the pattern precisely: label the counterpart a machine and the buyer stops inferring why you said something and starts merely checking whether it's correct. Where correctness is the whole job — scheduling, FAQs, document retrieval, meeting notes — that costs nothing. Where the buyer is really asking "will this person be on my side when it breaks" — pricing, objections, negotiation, bad news, the first call after an incident — it costs everything. And when someone insists their situation is unusual, put a human on it: Longoni, Bonezzi and Morewedge (2019, Journal of Consumer Research) showed resistance to algorithms rises with the belief that one's own case is unique.
机制：销售里的信任大半建立在一套推断上——「他为什么这么说、是不是在替我着想」。这套推断在大脑里有明确落点（内侧前额叶与右侧颞顶联合区的心智化网络），激活程度取决于对方被认为有多像人。贴上「机器」的标签，买家不再推断意图，只剩对输出的验收：只需要答案对的环节不掉分，需要「他站不站我这边」的环节，掉的是唯一那样货币。
失效场景：别读成「能瞒就瞒」。除了法律还有纯经济的理由——销售是重复博弈，被拆穿的代价按整条关系计价。另外披露的负面效应不是常数：对习惯用 AI 助手的买家小得多。
1. 把上周的时间按三栏摊开。多数人会发现第一栏占了三到四成——名单整理、纪要、跟进提醒、重复回答同样的问题。这一栏本周交出去，它已经不构成任何人的优势。
2. 第二栏定一条硬规则：验证成本必须低于省下的时间。邮件初稿、竞品对比、商业案例，AI 起草能省一半；但要是得花四十分钟核对它编的三个数字，这一项就退回第三栏自己写。省时间的假象是这栏最贵的坑。
3. 把第三栏写成具体动作，不是形容词。不是「建立信任」，是：这个月拿到三家客户内部的真实审批流程；说出一条「这条你们别买」；给一个出事时你到场的承诺并写进条款。顾问式销售那期的处方逻辑在这里就是护城河——模型能生成建议，不能承担建议错了的后果。
4. 如果你是新手：这是追赶的窗口，也是关得最快的门。AI 能把你的话术、结构、跟进纪律拉到平均线，但平均线本身正在贬值。省下的时间全用在拿真实客户对话上——那是模型给不了你的输入。
EN The strongest field evidence so far all points at one shape: AI turns the best performers' tacit know-how into something anyone can call, so novices gain most and veterans gain almost nothing. Brynjolfsson, Li and Raymond studied 5,179 customer-support agents at a software firm as a generative assistant rolled out (NBER working paper 2023; Quarterly Journal of Economics 2025) and found issues resolved per hour rose about 14% on average — with roughly a third more for the newest and least-skilled agents, and little to no gain for the most experienced. Dell'Acqua and colleagues (2023, Harvard Business School working paper) found a matching pattern and a warning in a randomised study of BCG consultants: inside the model's competence, AI users completed 12.2% more tasks 25.1% faster at higher quality; on a task deliberately placed outside it, consultants using AI were about 19 percentage points less likely to reach the right answer. The practical reading isn't "learn AI faster." It's that whatever skill you live on, if a model can copy it, its rent is going to zero — and the rent left sits in three places: inputs AI can't get, judgement someone has to sign, and relationships. So take last week's calendar and split it three ways. Column one — list building, notes, follow-up reminders, answering the same question again — is a third of most people's week and should be handed over now; it is nobody's advantage any more. Column two, where AI drafts and you verify, needs one hard rule: verification must cost less than the time saved, or the item goes back to column three. And write column three as actions, not adjectives: get three buyers' real approval chains this month, say out loud which option they shouldn't buy, commit in writing to showing up when it breaks. A model can generate advice; it cannot carry the consequences of the advice being wrong.
机制：这类工具把「最好的那批人怎么做」从隐性知识变成可复制的公共品。复制成本一低，这项技能的租金就消失——受益的是原来离最佳实践最远的人，靠它赚溢价的人什么也没多拿。剩下的溢价只能来自模型拿不到的东西：客户内部的私有信息、要担责的判断、不能被生成的人际存量。
边界：地板被抬高对新手是好事，但也意味着「比平均强一点」不再值钱。另外，第一栏交出去之后若没把时间投到第三栏，净结果只是更快地做同样多不产生差异的事——用数据做销售那期的口径在这里很有用：比上线前后的单位时间产出，别比感觉。
1. 数字回源系统，客户事实回记录。发给客户的任何百分比、金额、周期，必须能指回 CRM、财务系统或一份真实通话记录；模型生成的数字一律不出门。不是洁癖：一个编出来的客户案例被戳穿，你损失的是这条关系里剩下的全部单子。
2. 每条对外内容有一个具名的人签发。不是流程审批，是「出问题客户打给谁」。署名会让人真的看一眼——最便宜的一道闸。
3. 越省时间的地方，验证越要写成动作。「仔细检查一下」等于没检查。写成：核对三个数字、点开一次来源、确认客户名与职位。两分钟内做得完的检查才会被真的做。
4. 给自己留可修改的接口，别做全自动。这也是你能不能长期用下去的问题：人看到算法犯一次错就倾向彻底弃用，而只要允许自己稍微改一改输出，使用率就明显回升。AI 出稿、你终审，比全自动更可能活过第三个月。
EN There's only one way to use this badly: let it generate and have nobody actually check. The danger isn't that models err, it's that they err exactly where you're least able to notice — confident tone, clean formatting, plausible-looking citations. Skitka, Mosier and Burdick (1999, International Journal of Human-Computer Studies) documented automation bias in a flight-simulation task: with an automated aid present, people made both omission errors (missing events the aid didn't flag) and commission errors (following the aid into a mistake). Dell'Acqua and colleagues (2023) supply the sales-relevant version: outside the model's competence, AI users were about 19 percentage points less likely to be right. So design the loop around verification cost, not prompt craft. Three rules worth putting on the wall. Every number you send a customer — a percentage, an amount, a timeline — must trace to something that isn't the model: the CRM, the finance system, a real call recording. This isn't fastidiousness; one invented case study, caught once, costs you every remaining deal in that relationship. Every outbound piece gets a named person who signs it, meaning the person the customer calls when it's wrong — a signature is the cheapest gate there is, because it makes someone actually read the thing. And write verification as an action: "check it carefully" is not a check; "verify three numbers, open one cited source, confirm the name and title" is, and it fits in two minutes, which is why it gets done. Finally, keep an editing seat for yourself rather than going fully automatic. Dietvorst, Simmons and Massey (2015, JEP: General) showed people abandon an algorithm quickly after seeing it err even when it beats them on average; their 2018 Management Science follow-up showed that letting people modify the output, even slightly, restores use substantially. Draft-then-human-approve isn't just safer output — it's the version you'll still be running in three months.
机制：人把「流畅、自信、格式规整」当成正确性的线索，所以生成式输出天然骗得过一次快速检查；而只要有自动化提示在场，人的独立核查就会减少（自动化偏差）。叠加的结果是错误率不高、漏检率很高，漏掉的恰好是最贵的那类——落在模型能力边界之外、你又没有独立信息可对照的那部分。写成动作的、便宜的验证，是唯一能同时压住这两条的东西。
失效场景：别把验证做成第二遍全文重写——那等于没省时间，团队三周内就会偷偷跳过。验证要短、要具体、两分钟内做得完。另一个常见的坑：把客户敏感数据贴进不受控的工具里，先问清合规，别等法务来问你。
一小时，把「要用 AI」换成四件能查、能改的事。