Scenario
About 1 minutesLet's look at a scenario firstFictional example, numbers are illustrative
Approximately 90 minutes per day are used for ticket classification and first responses.
Approximately 600 tickets per week, with about 10% misrouted, and an average first response time of about 3 hours.
The knowledge base is updated only once a month, leading to repeated responses for similar issues.
Approach
About 2 minutesApproach: Build a dossier first, then have it answer based on the knowledge base
In these cases, always escalate to a human agent; do not allow the AI to reply.
Prompts and Examples
About 4 minutesPrompt 1: Ticket Processing
Suggested timeYou are a technical support team leader with six years of experience serving enterprise clients. Below are the Q&A library and a ticket (anonymized). Please output: 1. Category: Choose one from Inquiry, Fault, Request, or Billing; 2. Referenced entry number; if no matching entry exists, write 'No matching entry, requires human escalation'; 3. Draft reply, where every sentence must be supported by the referenced entry; 4. Whether to escalate to a human agent, specifying to whom and why. Rules: Prices, deadlines, and frequencies must be copied exactly from the original text; do not promise timelines or compensation, do not speculate on causes, and do not request passwords, card numbers, or addresses. Q&A Library: [Paste] Ticket: [Paste]
Example: Three fictional tickets, three handling methodsFictional example, numbers are illustrative
Tickets 1 and 2: Reply according to the library
- Ticket 1: Inquiry, referencing KB-01, reply stating the export range is up to 90 days
- Ticket 2: Fault, referencing KB-02, stating the junk folder has been checked according to steps, escalated to technical support, with no promise of handling time
Ticket 3: Escalated to human agent
- The customer requests compensation, involving a major performance issue, with no matching entry in the Q&A library
- Only write 'Escalate to human agent,' escalate to the account manager and technical support lead, without providing a reply
Common deviations and remedies
Inventing non-existent entry numbers
Remedy: Spot-check numbers against the original Q&A library text, and record such errors in the acceptance checklist.
Overpromising to customers
Remedy: Copy that sentence, add it to the 'transfer to human agent' list, and ask the same question again.
Merging similar questions that have different answers
Remedy: Instruct it not to merge questions with different answers.
Responses full of clichés, or the assistant gets agitated when the customer is
Remedy: Answer the question first, then provide next steps, keeping it under 80 characters; acknowledge the customer's situation first, then state facts, without making excuses or judging the customer.
Quiz and Takeaways
About 3 minutesThree quick questions to check yourself
Question 1: A customer reports a swollen battery. Can the assistant reply?
Question 2: The item number it provides does not exist in the Q&A library. What should you do?
Key takeaways from this lesson
- First, build the file: clearly write item numbers in the Q&A library, using numbers exactly as they appear in the original text
- It can only reply based on the library; if it cannot find an answer, state that transfer to a human agent is required
- Refunds, compensation, safety, privacy, and major failures must all be transferred to human agents; do not provide replies
- Do not promise timeframes or compensation in replies, and do not request passwords or card numbers
Next step: Organize your own Q&A library
Select the most frequently asked questions and write one response for each
- Select the eight to ten most frequently asked questions from this week, and number each response
- Copy the numbers exactly as they appear in the original text; if a number is unclear, mark it as 'to be supplemented'—do not let the AI fabricate it
Show the full transcript (video narration)
Customer Tickets and Knowledge Base
This lesson covers customer tickets and the knowledge base. We will have it handle three tasks at once: classifying tickets, drafting the first response, and updating the knowledge base; at the same time, we will set a threshold for when to hand back to a human. Let's look at a scenario first.
Let's look at a scenario first
Let's look at a scenario, which is fictional, and the numbers are just illustrative. Luoxiaotang, Customer Success and Technical Support at Qilan Intelligence, spends approximately 90 minutes per day on ticket classification and first responses. There are approximately 600 tickets per week, with about 10% misrouted, and an average first response time of about 3 hours. The knowledge base is updated only once a month, leading to repeated responses for similar issues. The Dragon Warrior drafts the first response quickly, but there are two risks: first, fabricating non-existent clauses, and second, exceeding authority by making promises to customers.
Approach: Build a dossier first, then have it answer based on the knowledge base
The process consists of five steps. Step 1: Create a Q&A library with 8 to 10 common entries, each numbered, keeping the numbers exactly as in the original text. Step 2: Anonymize tickets and import them in batches. Remove or replace names, contact information, account IDs, and serial numbers. Step 3: For each ticket, output the category, the referenced entry number, a draft reply, and whether it requires human escalation. Step 4: Create a list for human escalation. Issues involving refunds, compensation, legal matters, safety incidents, privacy, promised timelines, and major failures must always be escalated, specifying exactly who receives them. Step 5: Weekly extraction of candidate entries, which are added to the library only after human review. If no matching entry is found, write 'Requires human agent' instead of generating an answer yourself.
In these cases, always escalate to a human agent; do not allow the AI to reply.
This page defines the escalation rules, where the AI is strictly prohibited from replying. First, refunds, compensation, and legal threats. Second, safety incidents, including complaints about battery swelling or overheating involving personal safety. Third, privacy issues and suspected major failures. Fourth, replies must not promise timelines or compensation, nor request passwords or card numbers. When encountering these, it only writes 'Escalate to human agent,' without providing a reply or judging the cause.
Prompt 1: Ticket Processing
The first prompt is for ticket processing. It asks the AI to first categorize the ticket, then reference the entry number, draft a reply, and finally determine whether to escalate. The rules are specific: prices, deadlines, and frequencies must be copied exactly; do not promise timelines or compensation, do not speculate on causes, and do not request passwords, card numbers, or addresses. If no matching entry exists, write 'No matching entry, requires human escalation.'
Example: Three fictional tickets, three handling methods
Reviewing three fictional tickets, with numbers for illustration. Ticket 1 asks about the export range; it references KB-01, replying that exports can be up to 90 days, available on the records page. Ticket 2 states that the password reset email was never received; it references KB-02, stating the junk folder has been checked according to steps, escalated to technical support, with no promise of handling time. Ticket 3 requests compensation for one month of service fees and involves a major performance issue, with no matching entry in the Q&A library. It only writes 'Escalate to human agent,' escalating to the account manager and technical support lead, without providing a reply.
Common deviations and remedies
There are four common biases. First, inventing non-existent item numbers; the remedy is to spot-check numbers against the Q&A library's original text. Second, overpromising to customers; the remedy is to copy that sentence and add it to the 'transfer to human agent' list. Third, merging similar questions with different answers; the remedy is to instruct it not to merge. Fourth, responses full of clichés, or the assistant getting agitated when the customer is; the remedy is to answer the question first, then provide next steps; acknowledge the customer's situation first, then state facts, without making excuses. Test the seven types of tricky questions one by one after class. If any rule is not followed, add a more specific guideline.
Quiz Question 1
Let's do the quiz. Question 1: A customer reports a swollen battery. Can the assistant reply? A, Yes, reply according to the warranty terms in the library. B, No, transfer to a human agent; do not reply or determine the cause. C, Yes, but first have it determine if it is a quality issue. The answer is B. For issues involving personal safety, do not allow it to reply or determine the cause.
Quiz Question 2
Question 2: The item number it provides does not exist in the Q&A library. What should you do? A, Use this number directly; it looks professional. B, Base it on the original text in the knowledge library, and record this type of error in the acceptance checklist. C, Have it change the number until no error occurs. The answer is B. Item numbers must be based on the original text in the Q&A library, and this type of error should be recorded in the acceptance checklist.
Key takeaways from this lesson
Let's review the key takeaways from this lesson. First, build the file: clearly write item numbers in the Q&A library, using numbers exactly as they appear in the original text. Second, it can only reply based on the library; if it cannot find an answer, state that transfer to a human agent is required. Third, refunds, compensation, safety, privacy, and major failures must all be transferred to human agents; do not provide replies. Fourth, do not promise timeframes or compensation in replies, and do not request passwords or card numbers.
Next step: Organize your own Q&A library
Next step: Organize your own Q&A library. Select the eight to ten most frequently asked questions from this week, and number each response. Copy the numbers exactly as they appear in the original text; if a number is unclear, mark it as 'to be supplemented'—do not let the AI fabricate it. This video's narration is AI-generated. Thank you for watching.