Many companies have found that well-trained AI customer service has become an “artificial intellectual disability” in practical applications, with poor user experience and greatly reduced project effectiveness. This paper deeply analyzes the dilemmas faced by AI customer service in enterprise-level applications, analyzes the root causes of problems from technology, knowledge base construction, project management to operation mechanism, and proposes strategies and methods to avoid the trap of “artificial intellectual disability”.
Since AI has become popular, all major enterprises do not want to catch up with the outlet, scrambling to promote various AI applications within the enterprise, the first is AI customer service, in various PPT or reports, AI can replace manual customer service in minutes, 7X24 hours of “cyber cattle and horses”, to reduce costs and increase efficiency. However, when it came to the actual use, the project team was busy in the dark, but the business department “did not squeak once used”, and did not know whether to reduce the cost or not, and the “laughter” was really “laughter”.
1. Why has “artificial intelligence” become “artificial intellectual disability”?
To explore this issue, we first need to have a basic understanding of the AI customer service workflow:
AI customer service is essentially doing three things:Semantic parsing, knowledge retrieval, and answer generation
- Semantic Parsing: AI agents use natural language processing (NLP) technology to break down the text input by users into keywords, grammatical structures, and semantic intent.
- Knowledge retrieval: The system converts the parsed semantics into query instructions and searches for matching content in the structured knowledge base.
- Answer Generation: The AI integrates the retrieved information into natural language responses.
Users and the blogger service “chicken and duck talk” can be summarized into three major reasons:“AI can’t understand human language”, “The knowledge base can’t find the answer”, “AI doesn’t speak human language”
To achieve these three challenges, product managers will only continue to appreciate
Good product managers are very scarce, and product managers who understand users, business, and data are still in demand when they go out of the Internet. On the contrary, if you only do simple communication, inefficient execution, and shallow thinking, I am afraid that you will not be able to go through the torrent of the next 3-5 years.
View details >
1. Why can’t AI understand human speech?
In real business scenarios, the questions asked by users are always “varied”, and the same question, due to the user’s different educational qualifications, different regions, different dialects, and different speaking habits, will produce “countless” ways to ask questions. The following figure roughly summarizes 7 common situations, in one word –The vast majority of real users don’t know how to ask questions effectively。
I believe that many products will have such an experience, in daily business communication, customers or business departments often lose a screenshot or a video, asking you what is wrong? How to solve it? And this often requires you to peel off the cocoon, to rule out the possibility one by one, sometimes it may not be just caused by a single problem, sometimes after a long time of investigation may find that there is no problem at all. This is still the case with human judgment, not to mention that AI is trained from existing data.AI customer service cannot go beyond its training scope and creatively solve user problems.
2. How important is a structured knowledge base?
The most difficult thing for enterprise-level AI customer service is not technology, but building a knowledge base for AI customer service training.
Now major manufacturers are rolling up technology and pushing their own applications, AI customer service technology actually has no essential gap, it is nothing more than a question of who to buy and how much to spend. The “soul” of enterprise-level AI customer service is the knowledge base, and the essence of building a knowledge base is “Translate the complex business logic, data, and experience of enterprises into a structured language that machines can understand—— This is more dependent on business insights, organizational collaboration and continuous operation than technology implementation.
The following figure lists 7 major difficulties in the construction of a structured knowledge base in detail:
In the cognition of many people, basically all large models now have “multi-modal processing capabilities” (that is, large models can read ready-made PDF files, pictures, Word documents, parse links, etc.), so isn’t the knowledge base of AI customer service very simple? Isn’t it okay to import existing data directly? Why do you need to do more structured sorting?
Multimodal processing capabilities essentially solve the problem of “input compatibility” (allowing AI to read various forms of information), while enterprise-level applications require “output controllability” (answering questions accurately, efficiently, and consistently).
Structured data transforms scattered information into “standardized parts” that can be directly called by machines through field-based and relational sorting, which will try to avoid AI “looking for a needle in a haystack” or taking it out of context in complex content.
For example, the maximum number of tokens in some models is 2048, assuming that each token corresponds to an average of 1.5 characters, then the number of characters that the model can process is about 3072. The character limit affects the model’s ability to process long text, and when the number of characters in the input text or generated output text exceeds the limit, the model may not be able to process the text completely, resulting in lost information or incomplete generated results.
However, the knowledge base that can ensure the normal operation of AI customer service is often not a “hammer deal”, but a long-term uninterrupted cyclical process, and this workload may not be less than the answer of manual customer service.
PS: Here’s a digression, in the process of landing the project, I suddenly had an idea, maybe “AI customer service cost reduction and efficiency increase” is a false proposition from the beginning, which may also follow the conservation of energy theorem, it seems that it may replace the workload of manual customer service answering questions, but in fact, it is just transferred to the work of sorting and updating the knowledge base.
3. Why do users always feel that AI customer service doesn’t speak human language?
Although large language models are constantly iteratively tuned to make AI conversations more and more “anthropomorphic”, this does not mean that it has been solved”The conflict between machine logic and human language habits”.
AI’s responses rely on keyword matching and templated answers, lacking natural tone, emotional resonance, and contextual understanding, leading to cold responses, repeated questioning of provided information, or misuse of jargon, while unable to handle vague expressions and actively soothe emotions, like an “intern who only memorizes process manuals.”
2. How to avoid the trap of “artificial intellectual disability”?
A formula”Choose reasonable scenarios + scientific project management + long-term operation mechanism“, summarized into four major items with mind maps:
1. Identify the scene: high-value single point breakthrough
I believe that if you read this from the beginning, you have a certain understanding of the capability boundaries of AI customer service, and you can fully accept this reality: AI just can’t adapt to all customer service scenarios. It is hoped that AI customer service can replace taking over all the work of manual customer service, which in itself violates objective laws. But this does not mean that doing AI customer service projects is meaningless or completely impossible to do, but it is necessary to focus more on high-value scenarios and avoid blindly rolling out across the board.
Scenario screening criteria: high frequency, standardization, low risk
1) High-frequency standardized query
Typical scenarios:
- Order status inquiry (logistics tracking, return and exchange progress);
- Account information retrieval (balance, points, bill date);
- Policy clause answers (e.g. “What is the return period?”) ”）。
This type of scenario usually has clear answers and a fixed process, and the user’s questions are highly repetitive. You only need to connect to the existing business system within the enterprise through the API interface to find the specific parameters, and then convert the parameters into more vivid “human words” through the large model language to achieve reply, which can really replace manual customer service to query operations.
2) Rule-driven services
Typical scenarios:
- Reservation/Cancellation Service (Restaurant Reservation, Course Reservation, Event Guarantee);
- Form filling guidance (data upload, member registration information entry);
- Process operations (account password reset, service ticket submission).
AI can gradually guide users to complete operations through preset multi-round dialogue logic, connect with existing business systems within the enterprise, and reduce the workload of manual customer service operations.
3) 7×24 hours basic response
Typical scenarios: basic information: customer service phone number, official website link, working hours;
These scenarios include FAQs, how-to guides, and more. Preset AI reply rules and call logic. AI can respond to basic user needs around the clock, avoiding the “service vacuum” during off-hours.
2. Manage well: Balance anticipation and execution
Management here includes [upward management], [project management], and [risk management]
1) Upward management: cool down AI expectations for leaders
Upward management, in fact, is that as the project leader, it is necessary to brainwash the leader in reverse, which is of great significance to the promotion of the project. The core value is to protect the entire project team and avoid deviating from realistic high expectations such as “full business and full scenario coverage” and “comprehensive replacement of manual customer service within half a year”, resulting in the team’s exhaustion in the landing stage, consuming a lot of resources, with little effect and consuming trust and support for the team.
2) Project management: Scientifically plan the project rhythm
Project management mainly includes dissecting the entire project, delineating the plan and work objectives of each phase, and the delivery time nodes of each department. It must not be done here, without delineating the standard or scope, the business department will provide it by itself, and the database will be imported without brains. In the enterprise, everyone is very realistic, this work increases the workload of the business department, but it cannot reflect the achievements of others, there is a high probability that it is just a reluctant individual to cooperate, directly brainlessly import these into the knowledge base, and finally a lot of rework will be required, and finally slow down the project progress.
3) Risk management
In the process of AI customer service project execution, risk management is a key link that should not be underestimated, especially content compliance and information permission management. Once there is a leak in external content compliance, AI customer service may output content that violates laws and regulations, industry norms, or corporate policies, triggering regulatory penalties and public opinion crises.
Information permission management is also crucial, as incorrect permission configuration can lead to the leakage of sensitive information, threatening user privacy and core data security. If customer service personnel and technicians access users’ personal information and transaction data beyond their authority, or if the AI system maliciously tampered with or illegally obtained data due to improper permission settings, it will bring huge legal risks and economic losses to enterprises.
Key Facts:
- Compliance pre-review: The legal and risk control team needs to intervene in advance to set up “whitelists” and “blacklists” for the answers to the knowledge base (such as prohibiting promised words such as “absolute guarantee” and “100% effective”, and prohibited words under the Advertising Law);
- Dynamic monitoring: Block high-risk replies in real time through sensitive word filtering (e.g., ROI, “Eradication”).
- Permission hierarchy: the three powers of editing, reviewing, and publishing are separated (for example, the customer service manager can only edit the speech, and the compliance team has the right to veto);
- Data isolation: Set field-level access control (such as ID number and bank card number can only be retrieved by the risk control department);
- Operation traces: All knowledge bases modify and record the operator, time, and content, and support full-link audit traceability.
3. Mechanism co-construction: a rule-based cooperation system
The biggest difficulty in enterprise-level AI customer service projects is the collaboration of “people”. A reasonable and efficient collaboration mechanism is an important guarantee to ensure the normal operation of the project and avoid mutual quarrels.
1) Knowledge base co-construction mechanism
Formulate unified knowledge base construction specifications and processes, and clarify the responsibilities of each department in knowledge base construction. The business department is responsible for providing business knowledge and practical cases, the technical department is responsible for the technical implementation and maintenance of the knowledge base, and the operation department is responsible for reviewing and updating knowledge. Regularly organize cross-departmental meetings to synchronize the progress of knowledge base construction and solve problems that arise in collaboration.
2) Content compliance review mechanism
Establish a strict content compliance review process to ensure that AI customer service response content complies with laws and regulations, industry norms, and corporate policies. The compliance department needs to participate in the review of the content of the knowledge base and strictly control sensitive information, policy interpretation, etc. At the same time, a multi-layer review mechanism is set up, such as the preliminary review of the business department, the review of the compliance department, and the final review of the technical department, to ensure the compliance and accuracy of the content.
3) Knowledge base update mechanism
Formulate a regular update plan for the knowledge base, and update the knowledge content in a timely manner according to business changes, policy adjustments, user feedback, etc. Establish a trigger mechanism for knowledge updates, such as evaluating and updating a knowledge point in a timely manner when the number of consultations suddenly increases or there are many user complaints. At the same time, clarify the person responsible for knowledge update and the process to ensure the timeliness and effectiveness of the update work.
4. Long-term operation: continuous iteration and monitoring
1) AI response effect monitoring
Establish a multi-dimensional effect monitoring index system, including the first resolution rate, user satisfaction, problem solving time, and reply accuracy.
2) AI reply strategy tuning
Based on the performance monitoring results, continuously optimize the response strategy of AI customer service. Regularly check and export the AI replies of the user’s “stepping”, and a special person will review and determine the problem points, and adjust the AI reply strategy. Or add content to your knowledge base? Or increase the problem of association?
3) Long-term operation and maintenance of the knowledge base
Establish a daily operation and maintenance mechanism for the knowledge base, including operations such as adding, modifying, deleting, and reviewing knowledge. Arrange special personnel to be responsible for the management of the knowledge base, regularly clean up and optimize the knowledge base, delete duplicate and invalid knowledge, merge similar knowledge points, and improve the retrieval efficiency of the knowledge base. At the same time, collect new problems and new needs from users, add them to the knowledge base in a timely manner, and continuously enrich the content of the knowledge base.
4) Computing power resource monitoring
AI customer service is not completely free, it needs to consume computing power, although it is only a text call, not particularly complex logical reasoning, relatively less resource consumption, but it also needs to do a good job in monitoring resource consumption, which is also an important reference indicator to judge the production ratio of this project.
epilogue
The sad point is that since the explosion of AI, not only AI customer service projects, but also many enterprise-level AI application projects have been reduced to “AI for AI” PPT special supply, gradually deviating from the original intention of “technology changes life”. No one cares whether it is really “reducing costs and increasing efficiency”, and no one cares about the user’s experience, many AI customer service “communicate like playing the piano to a cow” and “mechanically reply to answer non-questions”, which not only fails to solve the problem, but also “adds fuel to the fire” for users…
Of course, many things are not something that can be controlled as a project leader or product manager, as a passive role in undertaking tasks, we can only take as few detours as possible and try to avoid some pitfalls, which is also the original intention of writing this article.
Finally, thank you for reading this~
