AI Customer Service Quality Inspection System

Problem

Manual review of customer-service chats is inconsistent when rules for needs discovery, recommendations, rapid-fire replies, and account-to-name mapping are scattered across spreadsheets and informal standards. Large session volumes make manual tallying and review hard to keep uniform.

How It Works

The skill takes a session Excel file as required input, optionally accepts a shift table in CSV or xlsx format, and outputs an HTML report plus 质检报告_客服绩效汇总.xlsx. It preprocesses sessions by splitting on human handoff, filtering template or placeholder messages, excluding robot accounts prefixed with jimi_vender_, and ignoring screenshot messages. It then checks three areas: performance, which requires both needs-discovery and recommendation after handoff and enforces the 3-second rapid-release rule; compliance, covering rapid sends, tone, wording, deflection, platform-sensitive terms, and zero-tolerance items; and negative-feedback review, which rechecks dissatisfied sessions. The shift table matches real names using date, human agent account, and online/offline time. The summary groups by person, caps each person at 200 counted sessions with sampling notes, and the report can be filtered by result, account, satisfaction, and name.

Limits

It assumes the documented column layout, Chinese scripts, and listed keyword rules; it does not infer unlisted policies or replace human review. Missing handoff markers, nonstandard account names, mismatched shift-table records, or uncertain first/repeat-contact and after-sales classification may require manual spot checks.