Support triage and weekly reporting

Background

The client is an online retailer selling fashion and home goods, taking inquiries through five channels: its own web store, Naver Smart Store, Coupang, KakaoTalk Channel, and email. That is 6,000 inquiries a month, 9,000 during sales, handled by four support agents.

Each channel kept its own pile. Every morning the agents opened the channels one by one, copied new inquiries into a spreadsheet, and typed in a type and an owner by hand. That alone took six person-hours a day. Urgent cases such as delivery incidents and payment errors sat among routine questions and sometimes crossed a weekend unnoticed. The team lead spent every Friday afternoon assembling channel figures into a report. Why a certain kind of inquiry was rising was known only as a hunch: “exchange requests seem to be up lately.”

What we built

The goal was not to add a chatbot. It was to remove the time agents spent finding, moving, and sorting inquiries and leave only the time spent answering and deciding. We built five things.

1. One queue

Inquiries from all five channels are pulled every five minutes into a single queue. Smart Store and Coupang connect through their seller APIs, KakaoTalk Channel through its consultation API, and email through inbox parsing. Messages from the same customer on two channels are linked by order number and contact details and shown as one. The client kept its existing help desk tool; nothing new was introduced.

2. Classification agent

Every incoming inquiry gets three labels.

• 12 types: delivery tracking, delivery delay or incident, exchange and return, payment error, product question, restock, coupon and points, account and login, review and rating, bulk purchase, partner proposal, other
• 3 urgency levels: delivery incidents, payment errors, and anything involving personal data are marked urgent immediately
• Owning team: support, logistics, payments and settlement, or merchandising

The agent finds the order number in the message, looks it up in the order system, and attaches the order and delivery status to the inquiry, so nobody has to search. Low-confidence cases are not forced into a category; they are flagged “needs review” for a person to assign.

3. Draft replies for routine inquiries

For six types, delivery tracking, restock notices, coupons and points, exchange and return procedures, accounts, and receipts, a reply is drafted in advance with the order details and delivery status filled in. Agents read, adjust, and send. No draft is ever sent before a person has reviewed it.

4. Escalation

If an urgent case has no reply within 30 minutes, or a routine case within 4 business hours, the team lead gets a messenger notification with a summary and a direct link.

5. Weekly report

Every Monday at 08:30, a report on the previous week goes to the team lead and the merchandising team: volume and trend by type, response time by channel, the top five recurring issues, and for each issue the likely cause (which product page or promotion the inquiries clustered around) with links to the related cases.

How it was rolled out

Weeks	Work
1–2	Analyzed 20,000 inquiries from the previous three months to define the 12 types and urgency rules, together with the four agents
3–6	Connected the channels and built the classification agent. Ran two weeks in shadow mode alongside human classification
7–10	Added draft replies, escalation, and the weekly report. Revised the type criteria three times based on agent feedback
11–12	Handover. Agent training, monitoring dashboard, and a guide for adjusting the classification rules

During the two shadow weeks, agreement with human classification rose from 87% to 94% as the criteria were refined. The remaining 6% are flagged for a person. After launch, the types and rules are reviewed together once a month.

How an agent’s day changed

Before, an agent’s morning began with an hour and a half of walking through five channels and copying inquiries. Now the queue opens already sorted by priority, each inquiry labeled with type and urgency and carrying its order details. Routine inquiries are handled by checking a draft and sending it, and the time saved goes to work that needs judgment, such as exchange disputes and bulk purchase consultations.

The team lead got Friday afternoons back and, on Monday, sees causes directly in the report: “31 of last week’s 40 exchange requests came from one product’s size-chart error.” After the page was fixed, those requests dropped to five a week.

The table below compares each task before and after.