Analyze Customer Service performance

#Use performance to answer two questions

Open Customer Service → Performance. This page is for analyzing Customer Service operations and outcomes. It answers two main questions: whether work or human-review pressure is building now, and whether service quality is improving over time.

APP CAPTUREThe English light-theme Customer Service performance page with real-time operating metrics, trends, and Shop filtering.

The page has two primary analysis tabs:

Tab	Best for	Time range
Realtime	Is AI handling work, are buyers waiting, and are escalations being resolved?	Last 1, 6, 12, or 24 hours
History	Are volume, response speed, escalation handling, satisfaction, and Guided GMV improving?	Last 7, 30, or 90 days

The Unpaid reachout tab belongs to a separate business workflow. See proactive outreach for unpaid orders for setup, delivery conditions, and experiment analysis.

#Realtime: find problems while they are happening

Realtime data refreshes every minute. Select all Shops or one Shop, then review four groups of signals:

1. Active / pending / escalated: active conversations show current workload; a growing pending line means buyer messages are queuing; a growing escalated line means more conversations need a human decision or takeover.
2. Escalation handling: compare new escalations with resolved escalations. When new work remains above resolved work, the human queue is accumulating.
3. Waiting buyers: separate counts show conversations waiting longer than 5, 15, and 30 minutes. This reveals a small set of long waits that an average can hide.
4. AI handling rounds and ended sessions: handling rounds count Agent work that actually started; ended sessions show conversations completed and closed.

Do not interpret one number in isolation. If pending work rises without a matching increase in AI handling rounds, check the responsible Desktop, Shop device assignment, and model connection. If AI continues working but escalations accumulate, confirm that recipients receive escalation delivery and that people complete the required decisions.

#Use Realtime to locate pressure

Follow this sequence:

1. select the affected Shop and an appropriate hourly range;
2. find when the pending line began to rise;
3. compare AI handling rounds during the same period to confirm that the Agent is still working;
4. check whether buyers waiting over 15 or 30 minutes are increasing;
5. compare new and resolved escalations to determine whether pressure is in the AI queue or the human queue;
6. open the conversation inbox or escalation and human review to inspect the underlying cases.

Realtime is for monitoring and investigation. A few hours of data is not enough to decide whether a model or business-instruction change is better.

#History: compare durable outcomes

History aggregates results by day for the last 7, 30, or 90 days. Its summary and charts include:

• New and ended conversations: whether incoming work is being completed;
• Escalated conversations, resolved escalations, and resolve rate: whether people keep up with cases that require judgment;
• Satisfaction and the 7-day weighted average: daily satisfaction is sensitive to rating volume, so read it with rated-conversation count and the longer trend;
• First response P50: half of measured conversations received a first response at or below this time; it describes the typical experience but does not replace the long-wait buckets in Realtime;
• Errors per conversation: a signal for changes in operating stability;
• Customer Service Guided GMV: TikTok attributes orders placed within seven days after a Customer Service reply. The page shows only complete 7-day rolling averages, so the latest available point is at least seven days behind today.

Promotions, traffic mix, complex aftersales cases, and small rating samples can all move one day’s result. Compare similar weekdays and business periods before and after changing a model, instruction, or operating process. Do not call a winner from one unusually high or low day.

#Review multiple Shops and export data

Start with All Shops to see whether a change is account-wide or concentrated in one Shop. Then filter to the affected Shop so volume from a large Shop does not hide a smaller Shop’s problem.

History supports CSV export. Use the daily file for weekly reporting, staffing comparisons, or a saved baseline around a configuration change. It includes conversation, escalation, satisfaction, first-response, error, and Guided-GMV fields.

#Recommended analysis sequence

1. Confirm health in Realtime: rule out an offline Desktop, unavailable model, or growing queue.
2. Find a trend in History: compare at least one complete and equivalent business period.
3. Read the underlying conversations: inspect the cases behind escalations, poor ratings, or long waits.
4. Change one primary factor: for example, an instruction boundary, model, or human-escalation process.
5. Record when the change happened: then review the same metrics and time range again.

Good Customer Service performance does not simply mean fewer escalations. It means routine work is completed reliably, authorization decisions reach people promptly, buyer waits stay controlled, and service quality can be verified over time.