中文精读
高途的AI落地：先让被忽略的客户得到服务，再谈替代金牌销售
42章经单期精读 · 企业AI与教育服务
嘉宾：封诚，节目当期为高途AI提效与创新产品负责人，此前为连续创业者｜主持：曲凯
节目发布：2024年10月12日｜时长：35分31秒｜整理：2026年10月18日
原节目与官方介绍
企业要的结果，可能藏在最普通的响应里
一家公司把AI接到很多环节，覆盖率很好看，是否真的改善了经营？封诚的判断从获客、转化、交付三段业务出发，最后把精力集中到已经获得的线索：那些没人及时回复、问题没有被讲清的潜在客户。
这期分享的价值在于如何选业务问题、组织流程和看最终结果。它不是全公司规模验证的报告。封诚明确说当时只接了一部分流量，尚未明显改变整个公司的商业模式；转化与收益数字也缺少完整样本、周期和对照信息。本文按2024年访谈理解这些经验。
1. 技术可做的地方很多，经营上值得先做的地方未必一样
节目线索：02:15–08:30。 封诚说三段都试过。交付端可以帮助辅导教师批改、做课件、缩短服务时间，但这些人工成本在整体经营中占比有限，省时也未必转成更多收入。这与重新用AI提供教学服务，是两种不同变化。
获客端较容易算账：同样的线索更便宜，业务部门会承认价值。但流量在外部平台，方法可能被模仿或随平台变化失效。节目谈到追热点、批量内容等短期办法，封诚对其持续性缺乏信心。本文不把其中对抗平台、搬运内容的描述转为操作建议。
转化端的吸引力是，用户已进入自己的流程，企业较能控制响应、内容、安排和跟进。已有线索不再长期闲置，改善就有机会持续作用于业务。编辑判断：这里的优先级依赖高途已有线索和销售流程。一个连用户都没找到的新产品，不能照搬“先做转化”。应同时问：价值大不大、结果看不看得见、团队能不能控制改动？
2. 三四倍究竟指什么，比数字大不大更重要
节目线索：08:52–12:30、17:19–18:25。 官方章节写“成单率提高3–4倍”，但正文更具体的说法是“单效”：每条线索最终能产生的GMV，即成交金额。封诚把三四倍放在这个指标上，又另外谈转化率提高一倍，并强调是小范围尝试。
这两个指标不能互换。每条线索成交金额，可能同时受成交比例、购买课程组合与客单金额影响；本文不推测试验中是哪一项贡献，也不将其改成全公司成交翻四倍。
团队发现，更基础的问题就能产生改进：线索无人服务、响应慢、答得不专业。对技术人员而言，“先跟没人理的人聊两句”可能不够新奇；经营者看到的则是原来流失的一段机会。编辑判断：先记录原来哪些人没有得到响应，再比较改善后的服务与成交，才能理解提升来自覆盖、质量还是客户选择变化。AI覆盖98%或100%只是过程数据，不能自动当成营收证据。
封诚说有专人维护指标、对照组和对话抽检，每一两个月能观察改进。这说明团队重视迭代，但节目没有展示足以复算结果的数据，仍应保留当事人自述身份。
3. 多个Agent的价值，在于任务交接清楚
节目线索：11:35–17:06。 SOP是把常见情形和处理步骤整理成可执行流程。团队希望把优秀顾问的判断、课程知识、语气与对话经验变成AI可用的规则，并不断按反馈修改。
封诚举的链路很具体：先识别用户是否今天就想报课；高意向走相应报名流程，低意向先答疑和保持联系；后续邀请试听、留下联系方式、提醒上课、回答问题并收集反馈。最后汇总用户经历与意向，交给真人金牌销售处理高客单签约。Agent在这里是承担一段目标明确任务的AI执行单元，重点是它与下一个环节怎样衔接。
因此，提效不等于把所有销售变成同一个最强顾问。它也可能像给金牌销售配助理，提前处理信息与专业问题，让真人接手时知道用户经历过什么。
编辑判断：要检查报名、提醒、答疑有没有依据同一份客户状态，是否把过期意向当成当前决定，转交报告是否漏掉反对意见。节目说试验中用户未必知道对方是AI，不能由此推导隐瞒身份是服务设计标准；身份、授权与用户预期需要另行处理。
4. 话术可以看到，持续优化需要业务结果回流
节目线索：12:49–13:37、22:46–26:28。 从外部体验一次，能否抄走整套系统？封诚认为不行，因为分支很多，初次来访者与听过课的人得到的策略不同。一次体验只暴露适用于自己的路径。
更重要的是知道用户后来究竟有没有报课。只有把这类结果接回，团队才可能比较策略、修改下一轮服务。单靠观察“AI聊得像不像”还不够。这里不需要神秘化技术壁垒：封诚自己说，最初主要是把顾问经验转成流程，后续改进才依赖结果闭环。
团队在自建与采购上也作区分。通用声音模型等零件倾向用第三方；深度销售流程牵涉公司SOP、课程变化、客户历史和内部系统，持续沟通成本高，因此内部组织更合适。这不是“所有垂直AI都必须自研”的结论，而是对可复用程度和集成成本的判断。
招聘上，他们曾想找既懂行业又懂AI还能带队的人，后来转向有业务感、学习力强的技术人员，让其访谈与观察专家。编辑判断：关键能力是把业务判断翻译成可检查的流程。谁更适合担任这个角色，应由实际协作检验，不能从一家公司当期经验否定行业专家的学习能力。
5. 教育里的“讲懂”，必须同时守住“解对”
节目线索：26:31–31:16。 封诚看好讲题：答案和步骤早就能在教辅里找到，学生真正愿意得到的服务，是老师根据自己的背景找出没懂的位置，再换一种讲法。封闭题库可配置较确定的解释，开放题则先要保证解答正确；他对当时八九成正确率仍不满意。
这里有两道验收。错误答案讲得生动仍然错误；答案正确、学生听不懂，也没有完成教学服务。学生已有知识、当下追问与实际理解，需要成为后续解释的输入。封诚认为不能只靠系统自我对练判断“是否讲懂”，而要有真人反馈。
他还把教师任务拆开：备课、批改等可以变化，激发兴趣、帮助寻找目标和动力仍很重要。“AI取代不了老师”是嘉宾的判断，不是本篇证明的终局结论。编辑判断：更可执行的做法，是观察某项任务移交后，教师和学生分别发生什么变化，避免用省下多少批改时间直接代表教育效果。
节目尾声还设想成人沟通辅导等软硬件结合，并提出产品三问：不推广时有人愿试吗，能持续进入生活或在需求出现时被想起吗，商业化能闭环吗？这些能帮助定义验证顺序；有关教育消费增长和融资环境的乐观说法，仍属于2024年个人市场判断。
练习：画出一条能追到结果的服务链
以下是编辑练习，适合已有客户流程的组织。选一类课程、一批来源一致的线索，标清从触达到试听、正式购买及后续服务的状态。先找到一个原来经常无人处理的环节，再试有限改动。
| 观察层次 | 保留证据 | 不能直接推出 |
|---|---|---|
| 服务覆盖 | 首次响应等待、未响应人数 | 成交一定提高 |
| 服务质量 | 抽样对话、事实错误、转交完整性 | 每个人都像金牌销售 |
| 经营结果 | 同口径转化、每线索成交金额、退款 | 小范围提升能线性放大 |
| 教学结果 | 学生能否解释与完成相似任务 | 回答正确就已经讲懂 |
设置对照、观察期和人工接手条件，把人员省时与最终增收分开记录。若团队无法获得成交或学习反馈，应先补结果通道，再扩大AI使用范围；模型数量和覆盖比例不能补上这个缺口。
回听导航与关联阅读
- 01:56：获客、转化、交付的价值选择。
- 09:43：章节大数字与正文单效定义。
- 14:23：意图、试听、提醒和人工交接。
- 17:11：先服务原来没人理的用户。
- 22:46：采购通用能力与内部业务流程。
- 26:31：答案正确和真正讲懂。
- 32:45：产品兴趣、持续使用与商业化三问。
相关主题：企业Agent、经营指标、教育服务。对照Answer.AI的学习与增长闭环和Dify的持续运营，可比较业务人员怎样参与AI产品的维护。
来源边界：依据完整提供的转写、官方介绍与既有精读重组，未逐句复听。开头字幕污染未当人物信息；高途与封诚按官方校正。效益数字、教学能力、市场与融资判断均为当期口述，未独立核验；章节与正文的指标差异已明确，不外推全公司或当前能力。
English edition
高途’s AI Implementation: Serve Neglected Customers Before Discussing the Replacement of Top Salespeople
42 Zhangjing episode analysis · Enterprise AI and education services
Guest: Feng Cheng, introduced as head of AI efficiency and innovative products at 高途, previously a serial founder | Host: Qu Kai
Published: October 12, 2024 | Duration: 35 minutes 31 seconds | Prepared: October 18, 2026
Original episode and official introduction
Business improvement can begin with an ordinary response
If a company connects AI throughout its processes and reports impressive coverage, has its business improved? Feng Cheng examines acquisition, conversion and delivery, then concentrates on leads already obtained: potential customers who receive slow responses or inadequate explanations.
The useful subject is how to select a problem, organize work and measure outcomes. This is not a company-wide validation report. Feng explicitly says only part of the traffic had entered the system and it had not visibly changed the entire business model. Conversion and revenue figures lack full samples, periods and comparison data. This analysis retains their 2024 interview context.
1. Technically possible tasks have different business priorities
Episode reference: 02:15–08:30. Feng says the team tried all three stages. In delivery, AI can help teaching assistants mark work, make materials and reduce service time. Those labor costs represent a limited share of the overall business, and saved time does not necessarily create revenue. Directly redesigning teaching around AI would be a different change.
Acquisition is easier to account for: cheaper leads create recognizable value. But traffic belongs to outside platforms; methods may be copied or undermined by platform changes. Feng lacks confidence in the durability of short-term approaches such as trend monitoring and bulk content. Descriptions of platform evasion and copied material are not converted here into operating recommendations.
Conversion offers greater control over responses, information, scheduling and follow-up after users enter the company’s process. Previously neglected leads provide an opportunity for continuing improvement. Editorial interpretation: This priority depends on existing leads and sales processes. A new product still searching for users cannot simply copy it. Ask whether value is substantial, results are visible and the team controls the change.
2. What “three or four times” measures matters more than its size
Episode reference: 08:52–12:30 and 17:19–18:25. An official chapter heading says order conversion rose three to four times. The fuller conversation assigns that figure to 单效, transaction value generated per lead, expressed as GMV. Feng separately discusses doubling conversion and emphasizes a small-scale trial.
These measures are not interchangeable. Transaction value per lead can reflect the purchase rate, course combinations and order values. This article does not infer which contributed or turn the statement into a company-wide fourfold increase in orders.
The team identified basic failures: unanswered leads, slow replies and weak expertise. Speaking to previously neglected people may seem unremarkable to technologists, while an operator sees a lost opportunity. Editorial interpretation: Establish who previously received no service, then compare service and purchases to understand coverage, quality and selection effects. AI coverage of 98% or 100% is a process measure, not revenue evidence.
Feng reports dedicated staff, comparison groups and sampled conversation reviews, with improvements observed every month or two. This shows attention to iteration, but the episode does not disclose enough data to reproduce the findings.
3. Multiple agents matter when handoffs are clear
Episode reference: 11:35–17:06. An SOP organizes common situations and responses into an executable process. The team wants to turn effective advisers’ judgments, course knowledge, tone and conversational experience into rules AI can use and update.
The described chain is concrete: identify whether someone intends to enroll today; direct high-intent customers toward enrollment while answering lower-intent questions and maintaining contact; invite trial attendance, obtain contact details, remind students, answer questions and collect feedback. Finally, summarize the user’s history and intent for a human salesperson handling a high-value contract. An agent here is an AI unit responsible for a bounded task; its connection to the next stage matters.
Efficiency can therefore resemble equipping a top salesperson with assistants who resolve information and professional questions in advance. The person taking over receives an account of what the customer experienced.
Editorial interpretation: Check whether enrollment, reminders and answers use consistent customer state, whether outdated intent becomes a current assumption and whether reports omit objections. Feng says trial users did not necessarily know they were speaking with AI. That does not establish concealed identity as a design standard. Identity, authorization and expectations need separate treatment.
4. Scripts can be observed; optimization needs business outcomes
Episode reference: 12:49–13:37 and 22:46–26:28. Could a competitor copy the system by trying it? Feng argues that many branches and different histories prevent one encounter from revealing everything. A newcomer and a former trial student receive different paths.
More importantly, the system needs to know whether someone eventually enrolled. Returning that result allows comparison and revision. Judging whether AI conversation sounds convincing is insufficient. Feng does not mystify the technology: the initial system largely translates advisers’ experience; further improvement depends on outcomes.
The team also separates buying components from building processes. General-purpose voice models can come from suppliers. Deep sales workflows involve company-specific procedures, changing courses, customer histories and internal systems, making continual coordination costly and internal implementation more suitable. This is a judgment about reuse and integration, not a rule that every vertical AI product must be built internally.
The initial recruiting ideal—industry expertise, AI knowledge and leadership in one person—proved scarce. The team shifted toward technically trained people with business awareness and strong learning ability who could interview and observe experts. Editorial interpretation: The needed capability translates business judgment into an inspectable process. Actual collaboration should determine who does that best; one team’s experience cannot dismiss industry experts’ ability to learn.
5. Explaining successfully still requires solving correctly
Episode reference: 26:31–31:16. Feng sees value in explaining problems. Answers and steps already appear in study guides; students need someone to identify their particular gap and explain differently. A closed question bank can support more controlled explanations. Open problems first require correct solutions, and he remains dissatisfied with an estimated 80–90% accuracy at the time.
There are two acceptance questions. An engaging explanation of a wrong answer is still wrong. A correct answer that the student cannot understand has not completed teaching. Prior knowledge, current questions and actual understanding should influence the next explanation. Feng argues that self-play alone cannot establish successful understanding; human feedback is necessary.
He also decomposes teachers’ work. Preparation and marking can change while cultivating interest, goals and motivation remains important. His confidence that AI cannot replace teachers is a judgment, not an established final outcome. Editorial interpretation: Observe what happens to teachers and students after a particular task moves. Marking time saved cannot directly represent educational impact.
The closing discussion proposes adult communication support combining hardware and software, then offers three product questions: will anyone try it without paid promotion, will it remain part of daily life or come to mind when needed, and can it support a business? These help define validation order. Optimistic statements about education spending and financing remain personal market views from 2024.
Exercise: Draw a service chain whose outcome can be traced
This editorial exercise suits organizations with existing customers. Select one course category and a comparable source of leads, marking states from contact to trial, purchase and service. Find one regularly neglected stage and test a bounded change.
| Layer | Evidence | What it cannot directly establish |
|---|---|---|
| Coverage | First-response wait and unanswered users | Purchases necessarily improve |
| Quality | Sampled dialogue, factual errors and complete handoffs | Everyone performs like a top salesperson |
| Business outcome | Consistently defined conversion, transaction value per lead and refunds | Small-trial gains scale linearly |
| Learning outcome | Students explain and solve a similar task | A correct answer has been understood |
Set a comparison, observation period and human-takeover conditions. Record labor savings separately from increased revenue. If purchase or learning feedback is unavailable, improve the result channel before expanding AI coverage. More models and higher coverage cannot supply the missing evidence.
Listening guide and related reading
- 01:56: Selecting acquisition, conversion and delivery opportunities.
- 09:43: The headline figure and the definition of value per lead.
- 14:23: Intent, trials, reminders and human handoff.
- 17:11: Serving previously neglected users.
- 22:46: Purchased capabilities and internal business processes.
- 26:31: Correct solutions and successful explanation.
- 32:45: Interest, continuing use and commercialization.
Related themes: enterprise agents, business metrics and educational service. Compare Answer.AI’s learning and growth loop and Dify’s continuing operations to examine business colleagues’ role in maintaining AI products.
Source boundary: Reorganized from the complete supplied transcript, official introduction and earlier close reading; audio was not replayed sentence by sentence. Opening subtitle contamination was excluded from biographical information. 高途 and Feng Cheng follow official metadata. Benefit figures, educational capabilities and market and financing views remain contemporary statements without independent verification. The mismatch between chapter and dialogue metrics is explicit; findings are not extrapolated to the entire company or current capabilities.