Gains:
- Ability to transform long and scattered customer requests into structured, actionable summaries
- Ability to classify requests according to category, urgency and customer sentiment with a fixed schema
- Ability to define a consistent output format (JSON/table) suitable for automation for bulk ticket processing
Imagine a support team's morning: 220 new tickets (tickets) have accumulated overnight. Some are a one-line "I forgot my password", some are an angry three-paragraph complaint, and some are actually a sales opportunity. Reading through this pile, assigning each one to the correct category, determining its urgency, and directing it to the right person (this is called triage; the same logic of sorting patients by priority in the emergency room) eats up the first two hours of the day.
Artificial intelligence (AI) can do this job in seconds and consistently. But the magic isn't in saying "summarize this request"; It imposes a fixed list of categories, clear urgency levels, and an immutable output format on the model. In this unit, we will establish a triage system that goes from processing a single request to labeling hundreds of requests in an automation-ready way.
Note: The category and urgency labels generated by AI are a preliminary screening tool. In particular, requests labeled "urgent" and "complaint" must be confirmed by a human before being processed.
Why Structured Summary?
A free summary (“customer is having problems with their shipment”) cannot be searched, sorted, or automated. However, the need of the support manager is clear to the following questions:
- Which category does this request fall into? (Shipping, Return, Payment, Technical, Product Information, Complaint, Sales Opportunity)
- How urgent is it? (Critical / High / Medium / Low)
- What is the customer's emotional state? (Angry / Disappointed / Neutral / Satisfied)
- What is its one-sentence essence?
- What should be the next step?
Once you define these questions in advance and give them to the model as a schema (constant fields and possible values), all 220 requests become comparable and filterable in the same format.
Step by Step: Establishing a Triage Scheme
- Pin the category list. Don't let the model fit; Give a closed list.
- Define the criterion of urgency. Concrete what “critical” means: service completely stopped, payment loss, security risk.
- Identify emotion labels. Use a limited and clear set.
- Import the output format. For batch processing, JSON (machine-read data format consisting of field-value pairs) is suitable, for single request, table is suitable.
- Make a "tick if not sure" rule. If the model is unsure of the category, let it say uncertain and the human will look.
- Verify. In the first batch, manually check the accuracy of the labels and set the prompt.
Copiable Prompts
Basic prompt that converts a single request into a structured summary:
Role: You are an experienced support triage specialist. Analyze the customer request below. Add a comment; just rely on what's in the text.Fill in the following fields:- summary: (max 1 sentence)- category: [Shipping | Return | Payment | Technical | Product Information | Complaint | Sales Opportunity]- urgency: [Critical | High | Medium | Low]- emotion: [Angry | Disappointment | Neutral | Satisfied]- next_step: (single sentence, concrete action)- unsure: ("yes" if category/urgency is unclear, otherwise "no") Request:"""{{ request_text }}"""
For batch processing, the prompt converts multiple requests into a JSON array at once:
Process the numbered requests below. Generate a JSON object for each with the following schema and return them all as a JSON array. Going outside the scheme: { "id": "", "summary": "", "category": "", "urgency": "", "emotion": "", "next_step": "", "I'm not sure": "" }Categories only: Shipping, Return, Payment, Technical, Product Information, Complaint, Sales Opportunity. Requests: {{ numbered_request_list }}
The prompt that clarifies the urgency criterion and teaches the model the definition of "Critical":
Determine urgency according to the following rule:- Critical: service completely unavailable, loss of payment, security/data risk, legal threat.- High: important function is broken but workaround exists; angry customer.- Medium: singular issue, not stopping workflow.- Low: request for information, suggestion, general question. Write the reason for your decision in one sentence in the "urgency_reason" field.
Prompt that captures the sales opportunity and establishes a support/sales bridge:
When processing the request, if the customer shows interest in purchasing a new product/package/supplement (e.g. "do you have a larger package", "how many users does it take"), make the category "Sales Opportunity" and add a one-sentence tip for the sales team in the "sales_note" field.
Weak Prompt / Strong Prompt
Weak prompt
Powerful prompt
"Summarize and classify this request"
Closed category list + urgency definition + fixed JSON schema
Generates different labels each time
Always gives the same label to the same request
He uses the word "urgent" according to his own wishes.
Applies concrete criteria for "critical"
He makes up the vague
emin_degilim: say yes and leave it to the person
Consistency is the golden rule here: if the same complaint does not fall into the same category on two different days, no reporting and automation will be reliable.
Three Mini Cases
Case 1 — Confidential critic. In a SaaS (internet rented software) company, the message "I can't log in, the whole team is waiting for 40 people" seemed ordinary because it was short in length. The triage prompt marked it "Critical" thanks to the urgency rule ("service completely unavailable" criteria). The request was handled in 6 minutes instead of waiting 2 hours in the queue; an SLA (service level agreement, i.e. promised response time) violation has been prevented.
Case 2 — Anger prioritization. One day, when the AI tags of 180 requests were examined, it was seen that 14 requests with the emotion "Angry" were put into a separate queue. These requests were directed to experienced representatives, and the negative survey score (CSAT, i.e. customer satisfaction score) that week improved significantly compared to the previous week.
Case 3 — Bridge from support to sales. "My current package is for 5 users, I need to increase it to 20 people, is it possible?" AI tagged the message as "Sales Opportunity" and added a sales note. The request automatically fell to the sales team; An upsell opportunity that would have gone unnoticed if it had been lost in the standard support queue has become a gain.
Tip: Keep your category list as short and discrete as possible. 20 categories will confuse the model (and your team); 6-8 clear categories are labeled more consistently and are meaningful in reports. Combine two frequently confused categories.
Connecting to Automation
The real power of the structured JSON output is that it flows automatically to the next step: The request labeled “Critical” immediately notifies the manager, “Sales Opportunity” falls into the CRM (customer relationship management software), “Return” goes into the self-service flow. But the first rule of automation: high-impact actions (refund, account closure) are never triggered based on the AI tag alone; Sometimes there is human approval.
Caution: Sentiment analysis is a prediction, not an exact measurement. A customer whom the model calls "Neutral" may actually be quietly very angry. Use the emotion tag to prioritize; but don't rely on it alone to draw definitive conclusions like "this customer is already satisfied."
Common mistakes
- Leaving the category list to the model; getting different, incompatible labels every time.
- Leaving a relative word like "urgent" undefined; Everyone's request is urgent.
- Not fixing the output format; Sometimes paragraph, sometimes list appears instead of JSON.
- Not providing an exit door for uncertainty (I'm not sure).
- Linking high-impact transactions (refund, account closure) to the AI tag without human approval.
- Automating the entire flow without manually verifying the first batch.
In summary
- Triage is quickly sorting through the pile of incoming requests by category, urgency, and emotion.
- The key to consistency: closed category list, concrete urgency definition, and fixed output format (JSON).
- Urgency and emotion labels speed up prioritization; It brings forward critical and angry demands.
- Structured output can be directly linked to automation (notification, routing, CRM).
- High-impact actions and ambiguous labels should always undergo human verification.
Application task
Batch process the 5 different customer requests you have (or samples) with the JSON array prompt above. Then manually check the output: (1) Is each category correct? (2) Do those marked "critical" actually stop the service? (3) Did I sure say "yes" in the right places? Correct any tags that do not fit and update the prompt (especially category definitions and urgency rule) accordingly. This exercise builds the habit of calibrating the schema to your own reality.
checklist
- [ ] I have defined a closed and discrete list of categories.
- [ ] I described the levels of urgency with concrete measures.
- [ ] I fixed the output format (JSON/table).
- [ ] I added an exit door for uncertainty (I'm_unsure).
- [ ] I manually verified the first batch and calibrated the prompt.
- [ ] I put a layer of human approval on high-impact actions.