GenAI Can Help Companies Do More with Customer Feedback

许多公司正在尝试使用生成式AI实现员工的生产力目标以及与客户的互动，但只有少数公司进行了相关部署。提升员工技能、改变流程和整合技术的困难仍然存在，许多公司陷入了长期试验的怪圈。

Many companies are experimenting with generative artificial intelligence (GenAI) now, both for internal employee productivity objectives as well as customer interaction, but only a few have production deployments. Difficulties with upskilling workers, changing processes, and integrating technology persist, and many companies are caught in a perpetual experimentation loop.

对于仍在寻找合适的部署位置的公司，我们推荐“客户的声音”（voice of the customer，VoC）应用程序的相关案例——解析、解释和响应来自所有不同渠道的客户意见。与员工生产力用例相比，这些通常更容易实施，因为它们不需要改变太多行为。由于提高客户满意度通常会带来经济回报，因此衡量经济价值的提高也更容易。

For companies that are still struggling to find the right place to deploy this new tech, we recommend use cases involving “voice of the customer” applications — parsing, interpreting, and responding to customer input from all different channels. They are typically easier to implement than employee productivity use cases because they don’t require as much behavior change. It is also easier to measure improvements in economic value because improving customer satisfaction often has a financial payoff.

无论是客户中心的呼叫电话、电子邮件、社交媒体信息，还是对销售人员的评价，随时了解客户意见显然很有价值。然而，大多数企业都很难系统地捕捉、分析和回应这些反馈：涉及内容过于庞杂且无序，对其进行审查和分析过于耗费人力，而回复则过于分散且繁琐。

It is obviously valuable to stay attuned to what customers are telling you, whether in calls to a customer support call center, emails, social media messages, or even comments to salespeople. Yet historically most organizations have found it difficult to capture, analyze, and respond to this feedback in any systematic way. The content has been too voluminous and unstructured, the review and analysis of it too labor-intensive, and the response too disaggregated and burdensome.

生成式AI可以提供帮助。以下是公司需要了解的内容。

Generative AI can help. Here’s what companies need to know to put it to work.

生成式AI可以做什么

What GenAI Can Do for Customer Voice

回复客户需要一系列不同技能。对于完整的回复过程，首先，无论是通过电子邮件中的文字、电话中的音频还是其他形式，必须捕捉他们说的话，然后对这些内容进行分析，对他们的反馈进行分类（投诉、赞美、请求？），再回复客户，解决他们提出的问题。生成式AI提供的方法可以轻松大幅改进上述步骤，主要是因为它可以执行听取和回复客户意见所需的大部分功能。

Responding to customers isn’t one task — it’s really a series of tasks that require different skills. First, you have to capture what they’ve said, whether that came via text in an email or as audio over the phone, or some other means. That content has to be analyzed, their feedback categorized (complaint? compliment? request?). Then a response needs to reply to the customer and address the pertinent issues they’ve raised. Generative AI offers a means to greatly and easily improve each of these steps, primarily because it can perform most of the functions necessary to hear and respond to customer comments.

这里可以参考便捷的ChatGPT等工具会如何融入这一过程。

Consider how a tool such as the out-of-the box version of ChatGPT might fit into this process.

获取客户反馈的第一步，是将电话语音转录为文本。与人工转录相比，AI转录的速度更快，成本更低。下一步是总结客户意见，这也是生成式AI最容易获得的用例之一。大多数公司发现，AI总结的准确性不亚于人工，且同样速度更快，成本更低。

The first step of capturing customer feedback may involve transcription of voice calls into text; other channels like email and social media are already text-based. Transcription by AI is significantly faster and more cost efficient than asking a human to do the job. Then, the next step is summarization of customer comments, which is one of most accessible use cases for GenAI. Most companies find that the summarization is as accurate as a human would provide, with a much faster and less expensive process.

总结反馈后，生成式AI还擅长对内容进行主题分类。事先提供所需类别和子类别的情况下，ChatGPT和其他模型可以很好地细分内容，确定问题的常见程度，并给出具体措施。

Once feedback is summarized, GenAI is also good at topic classification of customer content. ChatGPT and other models can do an excellent job of categorizing and subcategorizing content if provided with the desired categories and subcategories in advance, can determine how common the issue is, and recommend specific actions to address it.

生成式AI还能在情感分析方面实现突破，即区分正面、中性和负面的客户情感。这一功能已经存在多年，但工具往往难以识别讽刺、幽默和其他微妙的表达（例如，一家连锁酒店的情感分析系统就无法确定“泳池太cool了”是正面还是负面评价）。不过研究人员发现，生成式AI模型的精确度超过95%，能够准确评估多种语言的情绪，并能识别怀旧和忠诚等细微情感。

GenAI systems also enable a breakthrough in sentiment analysis — or distinguishing positive, neutral, and negative customer sentiments. This capability has been available for many years, but tools have often struggled with sarcasm, humor, and other subtleties of customer expression. (A hotel chain’s sentiment analysis system, for example, couldn’t determine whether “the pool was too cool” was a positive or negative sentiment.) But researchers have found that GenAI models can have precision levels of over 95%, can accurately assess sentiment in multiple languages, and can identify nuanced feelings such as nostalgia and loyalty.

对于来自呼叫中心或医生看诊的语音对话，生成式AI还可以通过分析员工声音，来提高客户满意度。通过实时转录文本，分析对话内容，生成式AI可以提示员工是否表现出了足够的同理心，从而解决这方面的问题。

Given a transcribed voice conversation — from call center agents or physicians seeing patients — GenAI can also improve customer satisfaction by analyzing the voice of the employee. Analyzing the dialogue, it can suggest whether the employee is being sufficiently empathetic to the customer’s issues. As GenAI systems can transcribe text in real time, they can potentially point out problematic comments in time to address them.

以上用途不仅停留在理论层面，企业已经将它们付诸实践。

The above uses aren’t just theoretical. Companies are already putting them into practice.

一些公司正在使用生成式AI总结与客户的对话。星展银行（DBS，是达文波特长期研究的网站）的首席数据和转型官潘利明（Nimish Panchmatia）表示，生成式AI驱动的虚拟助理可以实现通话转录、总结和行动建议。该系统已经上线数月，并逐步推出了新功能，呼叫中心座席人员的处理时间也节省了约20%。下一阶段，生成式AI将向座席实时推荐回答/解决方案，方便座席决定是否采用。使用该系统的主要目的是提高客户呼叫质量，而不是降低就业率。

Some companies are now using GenAI to summarize conversations with customers. According to Nimish Panchmatia, Chief Data and Transformation Officer at DBS Bank (a long-term research site of Davenport’s), a GenAI-powered virtual assistant performs call transcription, summarization, and recommendations for action. It’s been live for several months with new capabilities progressively being launched.  Call center agents are seeing savings of around 20% on their call handling time. The next stage to go live is the GenAI recommending answers/solutions in real time to the agent to decide whether to use with the customer. The primary reason for the system was to increase customer call quality as opposed to reducing employment levels.

另一个例子中，汤普森（作者之一）的一位客户，一家公用事业公司发现，它可以分析客户与自家收银员间的对话。该公司使用的生成式AI模型了解收银员是否礼貌周到地回复了客户，以及他们是否遵守了相关监管限制。模型甚至可以收集数据，最终用于计算客户支付账单的概率。

In another example, a public utility company that is a client of Thompson’s firm found that it could analyze conversations between customers and its employee bill collectors. The GenAI model it was using could understand whether collectors were being polite and empathetic with the customer and whether they were complying with regulatory constraints on what can be said to customers. It could even collect data that ultimately could be used to calculate a probability that the customer would pay the bill.

采取行动

Using GenAI to Take Action

了解客户声音的最终目的通常是采取行动。如果一家公司正在使用生成式AI合成和分类内容，那么就可以相对容易为客户提供个性化的回复消息，并确定应该将评论转发给哪个人，以便采取行动。如果提前告诉生成式AI哪些部门处理哪类问题，它还会有一定的成功概率判断应该通知哪个部门。

The ultimate goal of understanding the voice of the customer is generally to take action. If a company is using GenAI to synthesize and classify content, it’s relatively easy to craft a personalized message to the customer in response. It can also determine the right person or part of the organization to which to forward the comment for action. If the GenAI system is pre-prompted on which departments handle which types of issues, it can make a judgment (with probabilities of being correct) of which department should be notified.

汤普森利用生成式AI模型和LangChain（一个将LLM集成到系统工作流程中的框架）向一家区域杂货连锁店展示了如何自动化分析和响应客户信息。客户发来的信息如下:

One of us (Michael Thompson of First Analytics) utilized a GenAI model and LangChain (a framework to integrate LLMs in system workflows) to show a regional grocery chain how to automate analysis of and responses to customer communications. An example survey response from a customer might look like this:

我的烤鸡过期2天了，我的积分卡也没法用。不过我的导购很好，打电话问问题时，她反应非常快，而且很快就把我要的商品拿出来了。最重要的是，她提前3个小时就准备好了我的订单，这真是太棒了。

My rotisserie chicken was expired by 2 days, and I could not use my rewards card. My personal shopper was great though. She was very responsive when I called to ask questions and quick to bring my groceries out. On top of that she had my order ready 3 hours early which was AMAZING.

传统的自然语言处理（NLP）很难对这条消息进行分类，因为它包含了多个主题，混合了积极和消极情绪。而基于生成式AI的系统能够解析整个消息，识别不同问题，并起草一份可供人工审查的回复——目前这点是必需的。(以下斜体内容为生成式AI的回复。）

A traditional natural language processing (NLP) system would have difficulty categorizing this message as it contains more than one topic and is a mix of positive and negative sentiment. The GenAI-based system created is able to parse the entire message, identify the different issues, and draft a response for human review — currently required, but eventually a second GenAI model could prioritize the responses most likely to need review. (All GenAI model comments are in italics below.)

亲爱的（客户姓名），

Dear [Name],

感谢您抽出时间与我们分享这段经历。对于过期烤鸡和积分卡问题给您带来的不便，我们深表歉意。您的反馈是无价的，我们向您保证，我们正在尽快解决这些问题。

Thank you for taking the time to share your experience with us. We’re truly sorry to hear about the inconvenience you faced with the expired rotisserie chicken and the issue with your rewards card. Your feedback is invaluable, and we want to assure you that we are taking immediate steps to address these concerns.

我们也很高兴听到您的导购能够提供积极高效的服务。很高兴知道您的订单提前准备好了，她对您的需求做出了回应。我们一定会向她转达您的肯定，因为我们理解提供卓越的服务是多么重要。

We’re delighted to hear that your personal shopper was able to provide a positive and efficient service. It’s great to know that your order was ready ahead of schedule and that she was responsive to your needs. We’ll certainly pass along your praise to her, as we know how important it is to offer exemplary customer service.

感谢您的惠顾，我们会努力让您之后的购物过程更加便捷愉快。

We greatly appreciate your patronage and will strive to make your future visits nothing short of excellent.

祝一切顺利。

Warm regards,

客户服务中心

[Customer Service]

除了起草高度个性化的回复，该模型还能确定需要通知哪些部门，这些通知可以根据模型总结和分配责任的方式自动完成:

In addition to drafting highly personalized responses to each customer comment, the model identifies which departments to notify as part of the workflow. These notifications, in turn, can be automated based on how the model summarizes the situation and assigns responsibility:

这条评论是关于杂货店的混合体验：顾客遇到了与产品新鲜度（烤鸡过期2天）和无法使用积分卡的问题。然而，导购反应迅速、动作敏捷，甚至提前准备好了订单，给顾客提供了积极体验。

The comment is about a mixed experience at the grocery store. The customer experienced problems related to the freshness of a product (rotisserie chicken expired by 2 days) and the inability to use their rewards card. However, the customer had a positive experience with their personal shopper who was responsive, quick, and even had the order ready early.

1. 运营部门：该部门需要调查过期食品，这是一个需要立即关注的重大食品安全问题

• 事实：负责商店的运营管理。可能性：100%
• 假设：他们可以立即改正食品安全和产品质量问题。可能性：90%

1. Store Operations: This department needs to know about the expired Rotisserie chicken, as it’s a significant food safety issue that requires immediate attention.

• Fact: Store Operations is responsible for managing the stores. Likelihood: 100%
• Assumption: They can take immediate corrective action on food safety and product quality. Likelihood: 90%

    2. IT部门：无法使用积分卡可能是系统问题

• 事实：把控着公司的技术水平。可能性：100%
• 假设：可以解决积分卡问题。可能性：85%

2. IT: The inability to use a rewards card could be a system issue that needs to be resolved.

• Fact: IT manages the company’s IT capabilities. Likelihood: 100%
• Assumption: IT is the likely department to fix rewards card issues. Likelihood: 85%

    3. 人力资源：顾客的积极反馈可用于员工奖励。虽然与问题没有直接关系，但认可良好的工作有助于保持高标准

• 事实：负责管理人事需求。可能性：100%
• 假设：HR会对员工的积极反馈感兴趣，以便进行表彰。可能性：80%

3. Human Resources: The positive feedback about the personal shopper could be used for employee recognition. While not directly related to the problems, acknowledging good work helps in maintaining high standards.

• Fact: Human Resources is responsible for managing HR needs. Likelihood: 100%
• Assumption: HR will be interested in positive employee feedback for recognition. Likelihood: 80%

    4. 供应链：虽然没有直接提到，但如果货架上有过期产品，供应部门可能需要调查库存周转问题

• 事实：负责将物品送到商店。可能性：100%
• 假设：可能知道货架上为什么有过期产品。可能性：60%

4. Supply Chain: While not directly mentioned, if expired products are on the shelf, there may be an issue with inventory turnover that the Supply Chain department could investigate.

• Fact: Supply Chain is responsible for getting merchandise to stores. Likelihood: 100%
• Assumption: They may have insights into why an expired product was on the shelf. Likelihood: 60%

    5.采购部门：该部门与提出的具体问题最不相关，但应注意采购的产品已经过期

• 事实：负责采购产品。可能性：100%
• 假设：对店内问题（如产品过期）的直接控制较少，但也要了解情况。可能性：50%

5. Purchasing: This department is least relevant to the specific issues raised but should be aware that a product they sourced was found to be expired.

• Fact: Purchasing is responsible for procuring products. Likelihood: 100%
• Assumption: They have less direct control over in-store issues like product expiration but should still be informed. Likelihood: 50%

很少有人能对客户的问题进行更好的分类并确定可能的责任方。生成式AI模型甚至可以确定造成客户问题的根本原因（如供应链问题），而投诉中并没有提到这些问题。

Few humans could better classify customer issues and identify possible responsible parties. The GenAI model was even able to identify possible root causes (supply chain issues, for example) for the customer’s problem that were not mentioned in the complaint.

嵌入流程

Integrating GenAI with Systems and Processes

对于大多数组织来说，将模型的功能与管理客户输入和反馈的系统与流程整合在一起已然都是一项挑战。如果生成式AI模型与API、定制GPTs（定制模型）、客户关系管理系统（CRM）、VoC数据源，以及用于接收和发送客户通信的自动机器人配对，则大部分流程都可以实现自动化。对于一些使用LangChain等开源框架进行开发的企业而言，这种整合在技术上比较复杂，也比较耗时。

For most organizations, it will be challenging to integrate a model’s capabilities with its systems and processes for managing customer input and feedback. If GenAI models are paired with API connections, custom GPTs (tailored models), customer relationship management (CRM) systems, voice of the customer data sources, and automated bots for receiving and sending customer communications, much of the process can be automated. This integration has been technically complex and time-consuming for some organizations to develop using open source frameworks like LangChain.

不过，新推出的无代码工具可以提供帮助。包括Zapier和UiPath在内的几家供应商已经开发出自动工具，可以从生成式AI系统的输出中提取相关信息，用于工作流程。工作流程可以包括由生成式AI系统创建消息，通过自动机器人将信息发送给客户或员工，并将采取的行动更新到CRM系统。

However, new no-code tools are being introduced to help out. Several vendors, including Zapier and UiPath, have developed automation tools that can pull out relevant information from a GenAI system output for a workflow. The workflow could include creation of messages by a GenAI system, sending them to customers or employees with an automation bot, and updating a CRM system with the actions that have been taken.

当然，人工仍需参与其中，以解决客户问题，并找出解决这些问题的根本方法。人工必须审查客户“痛点”出现的频率，决定哪些是最关键的问题，并着手加以解决。也许从例行分析和回复客户中省出来的时间，可以用来关注根本问题。毕竟，我们真正想听到的客户反馈，是对我们的赞不绝口。

Of course, humans still need to be involved to resolve customer issues and figure out how to address their underlying causes. They’ll have to review the frequency of customer “pain points,” decide which ones are most critical, and move to address them. Perhaps the time freed up from routine analysis and response activities can be used to focus on the underlying issues. After all, the only customer voices we really want to hear are those singing our praises.

托马斯·达文波特是巴布森学院信息技术管理学的总统特聘教授，麻省理工学院数字经济倡议访问学者，德勤人工智能实践高级顾问。他是《AI行动方案：传统企业如何决胜人工智能转型》（All-in on AI: How Smart Companies Win Big with Artificial Intelligence，哈佛商业评论出版社，2023；中文版由中信出版集团出版，2024）的合著者。吉姆·斯特恩于2002年创办了“营销分析峰会”（Marketing Analytics Summit），并于 2004 年参与创办了“数字分析协会”（Digital Analytics Association）。他的第十二本书《人工智能营销:实际应用》（Artificial Intelligence for Marketing: Practical Applications）于 2017 年出版。近期他最受欢迎的研讨会是“创建生成式AI应用路线”（Creating a Generative AI Adoption Roadmap）和“生成式AI：创造力工具”（Generative AI: Creativity Powertool）。迈克尔·汤普森博士是First Analytics的CEO，这是一家专业服务公司，专门为多个行业设计和实施高级分析解决方案。

托马斯·达文波特（Thomas H. Davenport）吉姆·斯特恩（Jim Sterne）迈克尔·汤普森（Michael E. Thompson）| 文  

飞书、DeepL、Kimi | 译   孙燕 | 编辑