At present, AI has not completely replaced human customer service, but in some specific scenarios, AI has been able to effectively empower related businesses. In this article, the author discusses some application directions of AI+ customer service based on practical experience, so let's take a look.
When it comes to the practical application of large language models, the first thing that comes to mind is AI + customer service. This is a highly repetitive and standardizable area of the business that can be highly abstracted and summarized. Moreover, in the field of customer service, the demand for manpower is generally high, and the labor cost is relatively high.
Therefore, in this article, I will combine some practical experience and related articles I have read to talk about some application directions of AI + customer service.
1. Dismantling of customer service business
To tell how AI can be applied to customer service business, we first need to disassemble the customer service business.
We can roughly classify the customer service link according to its position before and after the business, which is mainly divided into the following two categories:
1. Pre-sales customer service:
Provide consultation and advice to customers before purchasing a product or service. Some businesses provide free services and only provide paid value-added services for some users, such as games, music and other applications. Customer service for these businesses can be categorized as "pre-sales customer service".
2. After-sales customer service:
Provide assistance to customers after they have purchased a product or service, such as returns, exchanges, repairs, etc.
The customer service business has the following characteristics:
1. High maintenance cost:
Customer service is "reactive" most of the time, dealing with a lot of user questions, complaints, and feedback, which requires a lot of human resources.
2. High management costs:
Customer service itself is a human being, and there will be various problems if it is a human being, such as laziness, fraud, abusive users, theft of corporate assets, etc., which will affect the efficiency of the entire team and even have a negative impact on the enterprise.
3. Difficulty in creating value:
The pre-sales customer service link gathers a large number of non-paying users, and it is relatively difficult to generate direct revenue from maintaining these users.
Therefore, for enterprises, customer service business is a high-cost, low-return business, and the customer service department is usually regarded as a cost-based department. How to reduce the investment of customer service manpower, or even achieve no manual customer service, this has become the ultimate vision of AI in customer service scenarios.
The customer service business can be broken down into the following parts:
1. Customer service training:
Different companies have different business situations and internal processes, so recruiting new agents can take time to train. At the same time, there is usually a large turnover in the customer service department, on the one hand, because this type of work has limited room for growth, and the average person will not be engaged in this kind of work for a long time, and on the other hand, there is a concept of peaks and troughs in the customer service business, that is, not all times the enterprise will face a high number of customer complaints. Therefore, qualified managers need to adjust their manpower requirements reasonably according to changes in business conditions.
How to get new employees up to speed quickly and reduce the training cost and trial-and-error cost caused by personnel changes is the key to the customer service training process.
2. User reception:
The user service process can be subdivided into several steps: "understanding the problem", "solving the problem", "following up", "feedback collection", and "recording and reporting". Customer service staff are responsible for resolving user issues, distilling valuable information, and conducting internal reports and summaries.
Although the customer service department is considered to be less valuable, the customer service department is the department that has the most direct contact with users. How to integrate and analyze information from the process of engaging with users, and feed it back to operations, sales, and even supply chain departments is one of the keys to improving the value of customer service departments.
3. Customer service management:
Since agents are human, there can be issues such as laziness, cheating, abusive users, theft of business assets, etc. How to prevent problems from occurring, identify them when they occur, and impose appropriate penalties after problems occur are the actions that customer service managers need to take.
Therefore, how AI plays a role in customer service business mainly focuses on customer service training, user reception, and customer service management.
2. AI in customer service training
1. Refinement of training materials
Customer service managers need to onboard new employees and conduct regular training for current employees. During this process, managers need to collect training materials on a regular basis, which includes maintaining a library of scripts, good practices, and error cases.
We can input the chat records in the process of user reception into the large language model, and use prompt words to refine the content of the maintenance speech library, excellent cases and error cases. However, because AI-generated content can be misleading, it will ultimately need to be reviewed and corrected by administrators to ensure that the content is accurate and appropriate before it can be officially used.
2. AI sparring
There is a fundamental difference between knowledge and skills. Knowledge is a summary of practice, but skills are the key to improving performance and performance. If you just blindly instill knowledge in customer service, it is difficult to ensure that customer service colleagues have really mastered the skills. Therefore, we need to "deliberate practice" our agents to ensure that our customer service colleagues internalize their knowledge into relevant skills, and in the process, review the problem and optimize the original solution.
If the customer service manager specially develops the practice questions, it will be a huge waste of the manager's time, and it will not be possible to interact with the customer service. So, here comes the power of AI. Using the chat records and prompt words of the user's reception, generalize and generate different customer complaint scenarios in batches, and conduct sparring with customer service. At the same time, we can use prompts to score the results of the sparring team, so as to assist managers in evaluating the ability of customer service, help customer service colleagues find problems, and optimize and improve.
3. AI in customer reception
1. Intelligent routing
Intelligent routing is a system that leverages artificial intelligence technology to intelligently assign calls to the most appropriate agent or service channel based on a variety of factors, such as customer needs, customer sentiment, service agents' areas of expertise, and real-time business conditions.
When the company's business is large, requires a large number of customer services, and the user scenarios are complex enough, the customer service team is divided into different skill groups, and each group is responsible for different scenarios. In this case, it is necessary to rely on the "intelligent routing" system to allocate users with different demands to maximize the "efficiency" and "effect" of reception.
The construction of intelligent routing can be based on user attribute custom rules for allocation, or a small model can be trained for conditional attribute allocation. However, none of these methods can be assigned based on user speech. In this case, we can use large language models (LLMs) to classify users' demands as the basis for intelligent route allocation.
In general, there are three main distribution methods for intelligent routing:
- Rule assignment based on user criteria.
- Algorithmic assignment based on small models.
- Allocation based on LLM refinement results.
2. Intelligent Q&A
Intelligent customer service Q&A refers to the method of using artificial intelligence technology to help customer service answer questions raised by customers or solve customer queries.
Currently, there are several solutions that implement intelligent question answering:
1. Knowledge Base Matching Q&A:
This scenario requires a pre-built knowledge base to answer the user's questions using the answers of the knowledge base through text matching or semantic matching. This approach is able to quickly and accurately answer questions that have clear answers in the knowledge base, and is suitable for Q&A scenarios in fixed domains or specific topics, without the AI hallucination problem.
The picture comes from Baidu, which is the customer service system of Wisdom Tooth Technology
2. Self-trained customer service AI:
For specific business scenarios or needs, machine learning and natural language processing technologies are used to train enterprise-specific customer service AI based on existing corpora. This approach can generalize the corpus, respond to a wider range of customer complaint scenarios, and provide personalized Q&A solutions that meet specific business needs.
However, AI is a summary of repetitive content, and it is prone to hallucinations in generalized scenarios, which may give users inaccurate answers.
3.RAG+LLM:
LLMs lack vertical knowledge and will answer customers' questions beyond what they ask. Therefore, RAG technology can be used to meet the needs of customer service scenarios. RAG conducts information retrieval through a large-scale corpus, obtains possible answer fragments, and then feeds these fragments into the LLM for further processing and answer generation. This approach takes advantage of the characteristics of LLMs and allows them to generate appropriate answers based on the customer's context and knowledge base answers. For small and medium-sized businesses that don't have the ability to train customer service AI, they can also take advantage of this approach to use the capabilities of AI.
However, the RAG + LLM approach also has the problem of hallucination, which may give users the ability to generate AI-made answers.
These methods mainly solve the problem of answer matching and distribution, and solve the problem of "what to answer and when". At the same time, the intelligent Q&A system can also introduce multi-round Q&A, multi-modal answers, and diversified knowledge bases to expand the problem-solving capabilities of intelligent Q&A.
1. Multiple rounds of Q&A:
When customers find customer service, they often have a purpose, some of which can be answered in one sentence, but some of which are not, such as returns, refunds, complaints, etc. At the same time, most customers cannot accurately describe the problem in one sentence, and there will be problems such as "synonyms" and "missing subject and guest". In this case, it is generally necessary to conduct multiple rounds of Q&A to solve it.
Therefore, the corpus of the intelligent Q&A system can not only be a single-sentence corpus, but also a multi-sentence corpus, which enables the intelligent Q&A system to cover more scenarios.
2. Multimodal Answers:
Multimodal answers not only contain common pictures, voices, videos, etc., but also include links, forms, and questionnaires to improve the information density and improve the efficiency of problem solving when interacting with users.
3. Diverse Knowledge Base:
Some problems may not be solved well, so in addition to being able to answer professional questions, intelligent customer service also needs to have the ability to greet, chat, and comfort customers to better serve customers. Therefore, we need to prepare relevant corpora for adapting to different scenarios, and here we can use keywords, intent recognition and other capabilities to control what knowledge base capabilities to use in what scenarios.
Finally, it is worth mentioning that "intelligence" is not a castle in the air, it needs to be accumulated step by step.
The three implementation schemes mentioned above are very dependent on the accumulation of corpus, and need to manually summarize the common customer complaint problems in the business first, and then give them to the intelligent question and answer system for use. The accumulation of data and the effect of answering are a spiraling process of mutual cause and effect.
Therefore, for small and medium-sized enterprises, in order to use a good intelligent customer service, a dedicated corpus operation member is indispensable, who needs to be responsible for collecting and refining high-quality corpus, and regularly updating the corpus in combination with business development.
At the same time, the illusion problem is difficult to eradicate, unless a large cost is invested in model training (or prompt word optimization) and corpus tuning, for small and medium-sized enterprises, instead of using such a high cost for maintenance, it is better to use the "knowledge base matching Q&A" scheme without illusion problem.
3. Customer Service Recommendations
Because the "customer service AI" and "RAG + LLM" in the intelligent customer service scheme both have obvious hallucination problems. We can use the form of "customer service advice" to avoid the risk of hallucination.
The so-called "customer service suggestion" means that in the process of customer reception, the output results of "customer service AI" and "RAG + LLM" are displayed in the form of suggestions, and the customer service staff decides whether to adopt them. This process is equivalent to a backup solution for manual review, and the results are presented to the user only after the manual review is passed. This approach is perfect for preventing AI illusions from negatively impacting the business, while also improving efficiency by using the output of AI.
In addition, we can also make the AI combine the human design to give corresponding reassurance suggestions, rather than based on the knowledge base. This can also enrich the diversity of customer service skills to a certain extent.
4. Summary of customer complaints
In customer service work, when the number of customer complaints is huge and the content is complex, it is difficult for humans to quickly identify the key content, and it takes a lot of time to read the context, which seriously affects the efficiency of customer complaint processing.
In order to solve this problem, we can use AI to summarize and refine content, quickly assist customer service personnel to refine customer demands, emotions, and intentions, and output them in a certain format, so that customer service personnel can quickly grasp user demands and customize response strategies.
In addition, the results of this refinement and summary can also provide decision-making basis for functions such as "customer service suggestions" and "intelligent question and answer", so as to further improve the efficiency and quality of customer service work.
5. Refinement of public opinion
In customer service work, the value of customer service is not only to answer users' questions or reassure users, but also to be able to extract valuable information from a large number of customer complaints, and feed this information back into operations, R&D, sales and other businesses, so as to improve the overall business effect.
I remembered that I saw an article before, which mentioned a mechanism within Tencent - the 10/100/1000 rule. The product manager conducts 10 user surveys every month, follows 100 user blogs, and collects feedback on 1,000 user experiences. While I'm not sure if this rule is still in place, it does highlight the importance of "listening to user feedback".
However, allowing front-line customer service personnel to manually summarize valuable information greatly depends on their professionalism and ability to classify problems, and is also susceptible to the influence of individual subjective factors, resulting in distorted summary results. In addition, the manual collation of the front-line cannot completely cover all the customer complaint cases online, so the comprehensiveness of the manual summary is insufficient.
If it's up to the manager to look up it in person, it's likely that you'll get lost in the massive amount of information and won't be able to find what you're looking for.
Therefore, we can use the power of LLM to classify public opinion for us, extract valuable information from it, and make it possible to manually access thousands of customer complaint information.
In this process, we can use the prompt word project to classify the topic by preset, and classify the corresponding content into specific topics, so that it can be consulted in the future according to the needs. This scheme can avoid the existence of "multiple approximate classifications" caused by multiple requests to LLMs, so that the distribution of public opinion cannot be effectively counted.
Compared with the traditional word segmentation scheme, the LLM-based public opinion analysis can analyze the connotation of public opinion more accurately, rather than simply segmenting the word.
6. AI voice chat/outbound calls
Due to the limited information conveyed by text, multimodal content forms, such as voice, pictures, videos, etc., can be considered for higher reception effect, which can be quickly generated in combination with existing AIGC solutions.
At present, the more mature is voice, with the help of the Wensheng sound model, we can give the customer service a voice, which is used to convey the corresponding content to the user. Through voice, emotions can be better conveyed, so as to emotionally comfort users. At the same time, based on AI voice, we can also realize intelligent outbound calls, which can be used for active marketing, user return visits and other purposes.
Combined with the crowd segmentation strategy, it realizes large-scale automatic user reach for thousands of people.
7. Typo Recognition
Using large language models to identify typos can reduce low-level errors of customer service personnel and improve the professionalism of service. Although this technology is relatively simple, its practical effect is remarkable.
8. Speech polishing
LLM is used to optimize the customer service skills, combined with specific personality settings, so that the words are more appropriate and professional, so as to assist the customer service to carry out more efficient reception. Through the optimization of dialogue, the service level of customer service can be improved to better meet the needs of users.
4. AI in the management service link
1. AI quality inspection
In customer service management, there are various problems, such as laziness, cheating, abusive users, and theft of corporate assets. Therefore, identifying, warning and responding to these problems has become one of the important tasks of managers.
If you only rely on manual inspection, it is easy to have problems such as "incomplete view" and "untimely view". Therefore, we can use large language models (LLMs) to conduct AI quality inspections, combined with prompt word engineering, sort out the dimensions of quality inspection (such as "friendly attitude", "emotional stability", "user feedback", etc.), and score them on different dimensions to evaluate the performance of customer service within a specified date.
Although this approach may have hallucinatory problems, it can greatly improve the timeliness and efficiency of management in identifying problems. At the same time, based on the results of AI quality inspection, we can also implement the early warning push notification function to ensure that the results are synchronized to the relevant members in the first time.
summary
These are some of the practical applications of AI in customer service systems. Overall, for SMEs, AI does not completely replace human customer service, because in many cases, the "cost of customization" is greater than the "labor savings". However, in some specific scenarios, AI can effectively empower businesses and improve efficiency. I believe that one day in the future, AI can bring greater changes to the customer service industry.
Columnist
The lemon cake is clean and hygienic, the official account: The lemon cake is clean and hygienic, and everyone is a product manager columnist. A B-end product in the game industry, responsible for CRM, risk control, BI, SDK, and AI-related content in the game industry, and regularly output personal thoughts or summary articles~
This article was originally published by Everyone is a Product Manager and is prohibited from reprinting without permission
The title image is from Unsplash and is licensed under CC0
The views in this article only represent the author's own, everyone is a product manager, and the platform only provides information storage space services.