Many companies are experimenting with generative AI to achieve employee productivity goals and customer interactions, but only a few have deployed them. Difficulties in upskilling employees, changing processes, and integrating technology persist, and many companies are stuck in a cycle of long-term experimentation.
For companies that are still looking for the right deployment location, we recommend the case for a voice of the customer (VoC) application—parsing, interpreting, and responding to customer input from all different channels. These are often easier to implement than employee productivity use cases because they don't require much behavior change. Since improving customer satisfaction often pays off financially, it's also easier to measure the increase in economic value.
Whether it's a call to the customer center, an email, a social media message, or a testimonial of a salesperson, it's clearly valuable to stay informed about the customer's voice. However, most organizations struggle to capture, analyze, and respond to this feedback systematically: the content is too large and disorderly, it is too labor-intensive to review and analyze, and the responses are too scattered and cumbersome.
Generative AI can help. Here's what companies need to know.
What generative AI can do
Responding to customers requires a range of different skills. For the complete response process, first of all, whether it is text in email, audio on the phone, or some other form, it is essential to capture what they are saying, and then analyze these to categorize their feedback (complaints, compliments, requests?). ), and then reply to the customer and solve their problem. Generative AI provides a way that can easily improve the above steps dramatically, mainly because it can perform most of the functions required to listen to and respond to customer opinions.
Here's how handy tools like ChatGPT fit into this process.
The first step in getting customer feedback is to transcribe phone calls to text. AI transcription is faster and less costly than human transcription. The next step is to summarize customer opinions, which is one of the most accessible use cases for generative AI. Most companies find that AI summaries are as accurate as a human and are just as fast and costly.
After summarizing the feedback, generative AI is also good at categorizing content by topic. By providing the required categories and subcategories in advance, ChatGPT and other models can do a good job of segmenting the content, determining how common the problem is, and giving specific actions.
Generative AI can also enable breakthroughs in sentiment analysis, which distinguishes between positive, neutral, and negative customer sentiment. This feature has been around for years, but tools often struggle to identify sarcasm, humor, and other subtle expressions (for example, a hotel chain's sentiment analysis system can't determine whether "the pool is too cool" is a positive or negative review). However, the researchers found that the generative AI model was more than 95% accurate, able to accurately assess emotions in multiple languages, and could identify subtle emotions such as nostalgia and loyalty.
For voice conversations from call centers or doctor visits, generative AI can also improve customer satisfaction by analyzing employee voices. By transcribing text in real-time and analyzing conversations, generative AI can address this issue by prompting employees if they are showing enough empathy.
These uses are not just theoretical, they are already being put into practice.
Some companies are using generative AI to summarize conversations with customers. Nimish Panchmatia, chief data and transformation officer at DBS (DBS), a website that Davenport has been researching for a long time, said generative AI-powered virtual assistants could enable call transcription, summarization and action recommendations. The system has been online for several months and has been gradually rolling out new features, saving call center agents about 20 percent in handling time. In the next phase, generative AI will recommend answers/solutions to agents in real-time for them to adopt. The main purpose of using this system is to improve the quality of customer calls, not to reduce the employment rate.
In another example, one of Thompson's (co-authors) customer, a utility company, found that it could analyze conversations between customers and their cashiers. The generative AI model used by the company understands whether cashiers are responding to customers politely and thoughtfully, and whether they are complying with relevant regulatory restrictions. The model can even collect data that is ultimately used to calculate the probability that a customer will pay their bill.
Take action
The ultimate goal of knowing the voice of the customer is often to take action. If a company is using generative AI to synthesize and classify content, it's relatively easy to provide customers with personalized response messages and determine who should forward comments to so they can take action. If generative AI is told in advance which departments are dealing with which types of problems, it will also have a certain probability of success in determining which department to notify.
Using generative AI models and LangChain, a framework that integrates LLMs into the system's workflow, Thompson showed a regional grocery chain how to automatically analyze and respond to customer information. The message from the customer is as follows:
My rotisserie chicken has expired for 2 days and my loyalty card is no longer working. However, my shopping guide was very good, and when I called to ask questions, she responded very quickly and quickly took out the goods I wanted. On top of that, she had my order ready 3 hours in advance, which was amazing.
Traditional natural language processing (NLP) makes it difficult to classify this message because it contains multiple topics with a mix of positive and negative emotions. A generative AI-based system can parse the entire message, identify different issues, and draft a response that can be reviewed by a human — which is currently required. (The following is a reply from generative AI.) ）
Dear (Client's Name),
Thank you for taking the time to share this experience with us. We apologize for the inconvenience caused by the expired roast chicken and loyalty card issues. Your feedback is invaluable, and we assure you that we are working on these issues as quickly as possible.
We're also glad to hear that your shopping guide can provide a positive and efficient service. It's good to know that your order is ready in advance and she responds to your needs. We will be sure to pass on your affirmation to her, as we understand how important it is to provide exceptional service.
Thank you for your patronage, and we will strive to make your future shopping experience as convenient and enjoyable as possible.
All the best.
Customer Service Center
In addition to drafting highly personalized responses, the model identifies which departments need to be notified, which can be automated based on how the model summarizes and assigns responsibilities:
This review is about a mixed experience at the grocery store: the customer encountered issues with product freshness (rotisserie chicken expires 2 days) and the inability to use loyalty cards. However, the shopping guide was responsive, agile and even prepared the order in advance, providing a positive experience for the customer.
1. Operation Department:
Fact: Responsible for the operation and management of the store. Likelihood: 100%
Hypothesis: They can correct food safety and product quality issues immediately. Likelihood: 90%
2. IT Department:
The inability to use the loyalty card may be a system issue
Fact: control the company's technical level. Likelihood: 100%
Hypothesis: The scorecard problem can be solved. Likelihood: 85%
3. Human Resources:
Positive feedback from customers can be used for employee rewards. While not directly related to the issue, recognizing good work helps maintain high standards
Fact: Responsible for managing personnel needs. Likelihood: 100%
Hypothesis: HR will be interested in positive feedback from employees in order to give recognition. Likelihood: 80%
4. Supply chain:
Although not mentioned directly, if there is expired product on the shelf, the supply department may need to investigate the inventory turnover issue
Fact: Responsible for delivering items to the store. Likelihood: 100%
Hypothesis: It is possible to know why there are expired products on the shelves. Likelihood: 60%
5. Purchasing department:
This sector is least relevant to the specific issues raised, but attention should be taken that the products purchased have expired
Fact: Responsible for sourcing products. Likelihood: 100%
Hypothesis: There is less direct control over in-store issues such as expired products, but it is also important to be informed. Likelihood: 50%
Few people can better classify a customer's problem and identify the likely responsible parties. Generative AI models can even identify the root cause of customer issues, such as supply chain issues, that aren't mentioned in the complaint.
Embed the process
For most organizations, integrating the functionality of the model with the systems and processes that manage customer input and feedback is already a challenge. Most of the processes can be automated if generative AI models are paired with APIs, custom GPTs (custom models), customer relationship management systems (CRMs), VoC data sources, and automated bots for receiving and sending customer communications. For some companies that use open-source frameworks such as LangChain for development, this integration is technically complex and time-consuming.
However, new no-code tools can help. Several vendors, including Zapier and UiPath, have developed automated tools that can extract relevant information from the output of generative AI systems for use in workflows. Workflows can include creating messages by generative AI systems, sending information to customers or employees via automated bots, and updating the actions taken to the CRM system.
Of course, humans still need to be involved to solve customer problems and figure out the root cause of solving them. Humans must review how often customer "pain points" occur, decide which are the most critical issues, and set out to address them. Perhaps the time saved from routine analysis and response to customers can be used to focus on the underlying issues. After all, what we really want to hear from our customers is praise for us.
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
Thomas Davenport is the Presidential Distinguished Professor of Information Technology Management at Babson College, a visiting scholar at the MIT Digital Economy Initiative, and a senior advisor to Deloitte's AI practice. He is the co-author of All-in on AI: How Smart Companies Win Big with Artificial Intelligence (Harvard Business Review Press, 2023; Chinese edition by CITIC Publishing Group, 2024). Jim Stern founded the Marketing Analytics Summit in 2002 and co-founded the Digital Analytics Association in 2004. His twelfth book, Artificial Intelligence for Marketing: Practical Applications, was published in 2017. His most popular recent workshops are "Creating a Generative AI Adoption Roadmap" and "Generative AI: Creativity Powertool." Dr. Michael Thompson is the CEO of First Analytics, a professional services firm that specializes in designing and implementing advanced analytics solutions for multiple industries.
Feishu, DeepL, Kimi | Translated by Sun Yan | edit
Smart people in the workplace never work hard
The reason why you have trouble getting rid of anxiety is because you have fallen into these thinking traps