Understanding Intelligent Customer Service in One Article: Development History, System Construction, and Market Promotion
-
In the field of artificial intelligence, intelligent customer service is one of the most easily implemented and technically mature application practices. This article focuses on intelligent customer service, outlining its development history, system construction, and market promotion. Enjoy~
At the 2018 I/O Developer Conference, Google demonstrated its conversational AI, Duplex.
Duplex accomplished two tasks:
In reality, Duplex played the role of an intelligent customer service agent.
In the field of artificial intelligence, intelligent customer service is relatively easy to implement and technically mature because the scenarios in customer service have relatively clear characteristics. This makes AI, which relies on large-scale data and high-concurrency demand processing, highly promising in the customer service domain.
Currently, leveraging advanced AI technologies such as big data, cloud computing, and deep learning, intelligent customer service can perform a series of complex operations, including autonomous Q&A, business processing, and fault diagnosis, meeting most response needs in the customer service industry and efficiently resolving user issues.
According to the "China Intelligent Customer Service Industry Research Report" released in May 2018, there are approximately 5 million full-time customer service agents in China. With an average annual salary of 60,000 RMB, plus hardware and infrastructure costs, the total market size is around 400 billion RMB.
Such a massive market naturally attracts many companies to intelligent customer service. But why hasn't a unicorn company emerged in this space yet?
Although this is one of the easiest AI projects to implement with relatively mature technology, how much cost would a company need to invest to develop and build an AI-powered customer service system?
Should a company build its own intelligent customer service system or partner with a suitable intelligent customer service platform provider, leveraging their capabilities to create a tailored solution?
Today, we'll explore this in detail.
The Chinese customer service software market has roughly gone through three development stages: traditional call center software, PC web-based online customer service + traditional customer service software, and the intelligent customer service stage featuring cloud-based customer service + AI chatbots.
From the current composition of the customer service industry chain, the upstream infrastructure segment is already mature, with a few giants monopolizing the market. In the future, they will continue to expand downstream, building an enterprise service ecosystem.
Among midstream customer service product providers, cloud-based customer service vendors have seen a few leaders emerge after years of competition, but no giants have yet appeared, and competition remains fierce. Product features have become more diverse, and application scenarios have expanded from customer service to sales, marketing, and other areas. On the other hand, AI chatbots, by assisting human agents and handling simple, repetitive queries, have significantly improved the efficiency of human customer service. Meanwhile, AI is transforming the way businesses interact with customers across all touchpoints, accelerating the intelligent upgrade of both online and offline customer service.
(1) Intelligent Customer Service System Based on Knowledge Base Responses
The workflow for retrieval-based responses is as follows:
In practical applications, thresholds are set to ensure response accuracy. If the score for each question falls below the threshold, the top few questions are returned to the user in a list format. The user can then select the question they want to ask and receive a specific answer.
(2) Multi-Turn Dialogue System Based on Slot Filling
Building a slot-based dialogue system is a relatively specialized and complex process, typically divided into three main stages: requirement analysis, bot construction using a platform, and continuous optimization.
To understand this system, let's first familiarize ourselves with a few key terms:
1) Intent
Intent refers to the primary request or action expressed by the user during a voice interaction.
Example of intent:
2) Skill
A skill is an application designed to fulfill a specific user need. For example, when a user says, "Check where my shampoo delivery is," the system enters the package tracking skill.
3) Q&A Skill
By configuring Q (user questions) and A (bot responses), simple dialogues between users and bots can be implemented.
Task-based skill: Building on Q&A skills, advanced features like slot filling and API calls are added to enable functions such as information queries, searches, or other tasks.
4) Lexicon
A lexicon contains variable content for specific keywords, such as time or location lexicons.
Semantic slot: A semantic slot is a keyword in the user's utterance that helps the system accurately identify intent. For example, a zodiac semantic slot includes the names of the 12 zodiac signs. Semantic slots and lexicons are often used together, with slots typically referencing lexicons. One semantic slot can be bound to multiple lexicons, and one lexicon can be associated with multiple semantic slots.
5) Follow-up Question
When a user's utterance lacks a semantic slot value, the bot automatically asks a follow-up question.
For example, if a user asks, "What's the weather like?" and the system cannot determine the location, the bot might ask, "Which location's weather information would you like?" Multiple follow-up prompts are usually set and randomized.
In China, open bot systems like Baidu UNIT and WeChat's dialogue platform use this technical framework.
For a natural language dialogue system, the core tasks of understanding are intent parsing and slot recognition.
For example: "Book a train ticket from Beijing to Shijiazhuang tomorrow at 8 AM." In this case, the user's intent is to book a train ticket, and the slots include departure location, destination, and time. If there are multiple train options for that time, the system will ask the user to specify which one they want.
Using Baidu UNIT as an example, here's the process for building a ticket-booking intelligent response system.
Of course, this is just one scenario for train tickets. Other scenarios include refunds, rescheduling, and inquiries, all of which need to be identified during requirement analysis.
(1) Evaluation Based on Manual Annotation
For systems that rely on a Q&A knowledge base, response capability is limited by the richness of the knowledge base—specifically, its coverage of user questions. Higher coverage leads to higher accuracy.
Thus, the system cannot answer all user questions. The ideal state is to accurately answer all questions it can handle and reject those it cannot.
Evaluation metrics include problem resolution rate, rejection rate, recall rate, and accuracy rate. The goal is to align the system's response rate as closely as possible with the actual data response rate while maximizing recall and accuracy.
A small dataset is sampled from the daily full dataset, ensuring its distribution matches the full dataset. An annotation team then labels the dataset, marking the actual answers for each question. After annotation, a quality check is performed to ensure accuracy, resulting in a standardized evaluation set for daily data.
This evaluation set is used to assess system performance. Any new model iterations are tested against this set, and only models meeting certain metrics are deployed.
(2) Evaluation Based on User Feedback
Manual evaluation can assess the accuracy of an intelligent customer service system, but whether the answers are reasonable and solve user problems requires user feedback. The ultimate goal of the system is to help users resolve issues.
Evaluation features are designed into the product, such as allowing users to rate each answer or session. After chatting with a human agent, users may receive a satisfaction survey, as shown below.
Metrics like participation rate and satisfaction are compiled to gauge system performance. In practice, low participation rates are common, so various methods are used to encourage feedback.
Intelligent customer service systems are primarily B2B. Generic systems target a broader market with more users, while vertical-specific systems serve fewer users.
For example, in the insurance industry, there are only about 100 insurance companies nationwide. Developing vertical-specific intelligent customer service systems requires AI teams to deeply understand the industry, business needs, workflows, and cross-departmental communication.
Developing intelligent customer service systems for vertical industries often leads to being trapped in one or two major projects, constantly catering to users' personalized needs. The result is a highly "customized" system with a very limited market. After completing a few projects, companies quickly hit a transparent ceiling.
While general-purpose intelligent customer service systems have a vast market, they lose their competitive edge against vertical-focused teams. With minimal technological differences between providers currently, small companies can offer customization to users—something general systems cannot—ultimately leading to a large market being gradually eroded by emerging vertical-focused small firms.
So what's the solution?
In the early days of the internet, portal sites emerged first, serving the needs of the majority. Later, WeChat public accounts allowed subscription-based content, creating personalized information platforms for each user. From the user perspective, customization is the evolutionary direction, and eventually general-purpose customer service will be replaced by vertical industry solutions.
Traditional industries like banking, insurance, securities, and real estate—large enterprises with strong customer service demands—are highly motivated to adopt intelligent systems but also have stringent data security requirements, insisting only on locally deployed solutions.
For these major clients, locally deployed solutions mean project-based business models—charging per project. The advantage is generating millions in revenue from a single project, ensuring profitability early on. The downside is that private deployments require extensive customization, consuming significant manpower and making scaling difficult, limiting long-term growth potential.
So what's the solution?
From a data security standpoint, technological advancements will address these concerns. When mobile payments first emerged, people feared their bank cards might be stolen or their Alipay accounts hacked. Now, such worries seem unfounded. Sufficient investment drives technological development. SaaS services, serving more users, expose vulnerabilities faster, accelerating system security evolution. Private deployment isn't the optimal choice.
At inception, all intelligent customer service startups face a critical decision: target large enterprises or focus on SMEs first?
Focusing on SMEs allows standardized SaaS products to meet their needs. This lightweight model reduces manpower costs, enables scaling, ensures recurring annual revenue, and builds technical barriers through continuous data feedback.
The downside is the high difficulty in early customer acquisition, requiring extensive market education. Additionally, high SME mortality rates make sustained renewals challenging, making profitability difficult initially.
Targeting large clients presents different challenges: meeting customization demands is tough, sales cycles are long, and established vendors with mature products dominate. Relying on a few major clients poses significant risks for startups.
So what's the solution?
Vertical SaaS systems need more users to iterate faster. With only one or two major clients, constructive improvement suggestions are scarce. Therefore, targeting SMEs to quickly onboard initial users, launch the system, and continuously optimize is crucial.
Many cite traditional customer service industry pain points that intelligent systems can solve, such as:
(1) High labor costs
With demographic dividends fading, employer costs keep rising.
Is this a genuine need? Customer service isn't a core department for most companies, often receiving little attention. SMEs have few service staff, offering limited cost savings, reducing motivation to switch. For large enterprises, while labor costs are significant, they can afford to develop in-house systems where ROI makes sense (e.g., Didi with its large service team prefers building its own).
(2) Decision-making paradox
Intelligent systems replace human tasks, implying department downsizing.
While cost-saving for companies, this reduces department heads' organizational influence. Though inevitable long-term, current sales processes remain top-down rather than driven by urgent departmental needs with sustained follow-up.
Nothing is new under the sun. Large companies adopting foundational tech to build their own systems may become a trend, balancing data security and cost control. For SaaS providers lacking technological oligopoly advantages, product promotion and service capabilities become critical.
Do intelligent customer service companies have moats? What constitutes their moat?
Usage habits, data accumulation, and knowledge base refinement form industry barriers. High switching costs deter users from changing providers.
Thus, rapidly expanding the user base creates the moat. Solely focusing on intelligent customer service limits growth potential; finding a second growth curve determines how far these companies can go.