How AI-Powered Minos Transforms Customer Service Quality Inspection

1. Challenges and Breakthrough

ZhiZhi insists on strict control of customer service quality, assigning staff to inspect a portion of daily interactions. However, the sheer volume means many non‑compliant conversations are missed, and manual reviewers are slow and error‑prone.

Traditional inspection is like a leaky sieve, unable to keep pace.

Amid the AI wave, ZhiZhi launched the AI quality inspection system “Minos” (named after the Greek judge), giving customer service quality management a “sharp eye”.

2. Abstract Modeling and Process Support

The Minos core enables efficient, accurate post‑hoc checks of service workflows.

It is built on three core concepts: inspection items, inspection plans, and inspection tasks, which work together to support end‑to‑end intelligent inspection.

Inspection Items

The smallest unit, focusing on specific service dimensions such as “use of polite language”. Items can enable AI inspection and select models (large‑model semantic understanding, programmatic logic checks, or keyword regex), and multiple models can be applied simultaneously.

Inspection Plans

Templates composed of multiple items, tailored to business scenarios. Plans define a passing score; conversations scoring below trigger automatic alerts.

Inspection Tasks

Execution units that automatically sample service data based on conditions (e.g., time window, low user rating) and apply the corresponding plan. After inspection, results are generated and non‑compliant conversations are pushed to supervisors, closing the loop.

Inspection Configuration

Design Idea

By abstracting the three concepts, different items can be assembled into scenario‑specific plans, achieving fully configurable inspection across all use cases.

Inspection Process

3. Evolution of Inspection, Continuous Exploration

Initially, Minos relied solely on large‑model inspection via prompts. Real‑world use revealed that some items (e.g., follow‑up commitments) cannot be judged by semantics alone.

Therefore, regex matching and programmatic judgment were added to complement the model, improving accuracy and diversity.

Large Model

Acts as a “semantic interpreter”, evaluating soft service standards such as friendliness, complete answers, timely responses, and emotional soothing.

Keyword Inspection

Regex serves as a “keyword sentinel”, catching prohibited phrases like “I can’t help” or “I have no authority”, either independently or alongside the model.

Programmatic Judgment

Functions as a “commitment auditor”, verifying that promised actions (e.g., ticket creation, escalation) are actually executed by cross‑checking conversation content with business system data.

4. Building Custom Models for Targeted Training

Because generic models cannot cover domain‑specific standards (e.g., “device inspection report clarity”), ZhiZhi trained a proprietary model to classify abnormal conversations and hand them to human reviewers, boosting detection of complex items.

Project Results

For the “lack of soothing and apology” item, the custom model achieved a weekly precision rate 8.6 percentage points higher than the baseline, raising detection rate by 26 times compared with a generic model, and now supports full‑volume online data.

5. Current Usage

Minos now supports 10 inspection scenarios, each with a dedicated plan, including eight AI‑driven items. It processes over 20,000 service interactions daily, delivering a two‑order‑of‑magnitude increase in inspection volume and meeting the original efficiency goals.

6. Future Plans

Refine Inspection Items

Split complex items into finer dimensions to let the model perform more granular checks, improving accuracy.

Optimize Inspection Workflow

Replace hard‑coded combinations with a workflow engine that allows users to define model sequencing (e.g., regex → large model → programmatic check) and conditional triggers (e.g., activate model B only when model A flags potential violation).