How AI‑Powered VOC Recommendation Cut Customer Service Time by 38%

1. Introduction

As the business expands rapidly, the volume of inbound customer calls grows, making it critical to enable agents to handle calls and create tickets efficiently with limited human resources. This article explains how AI assists agents in two key areas to reduce their workload.

2. VOC Recommendation

2.1 VOC Concept and Business Pain Points

VOC Definition : Voice of Customer captures the core intent expressed in inbound conversations (e.g., after‑sale requests, product inquiries, complaints) and must be accurately classified when creating tickets.

Business Value : VOC data is the primary source for customer demand analysis, service optimization, and product improvement.

Process : After a call, the agent selects the appropriate VOC tag to create a ticket.

2.2 Iteration: From Static to Dynamic AI Recommendation

The recommendation evolved from a static, channel‑based top‑VOC list to a real‑time AI model that parses conversation intent and returns the top‑5 suggestions. The static approach ignored conversation content, leading to frequent mismatches, while the AI approach reduced selection time from 14.5 seconds to 9 seconds.

2.3 AI Implementation

Architecture :

Training & Iteration : Positive examples are derived from agents' final VOC selections, manually labeled as ground truth. New data triggers model retraining, validation, and deployment in a continuous loop.

Model Choice : A supervised classification model is used because the task is well‑defined and simple, mapping input conversation data to discrete VOC tags.

2.4 Quantitative Results

Key metrics before and after optimization:

Single VOC selection time: 14.5 s → 9 s (‑37.9%)

Model recommendation accuracy: 14% → 74% (+428%)

Business impact includes saving over 10,000 hours of labor annually and improving VOC data quality, which drives product enhancements.

3. Session Summary

3.1 Scenario and Pain Points

Agents must review historical call records and manually summarize each conversation into tickets, which is time‑consuming and leads to inconsistent quality.

3.2 How It Works

The system uses different prompts for various scenarios, tests large‑model outputs with real data, and iterates through multi‑round labeling to finalize the solution.

3.3 Business Metrics and UI Presentation

Ticket creation time: 148 s → 131 s (‑11%)

Conversation summary feature adoption: 80% after rollout

4. Conclusion

Providing a more efficient and user‑friendly system for agents is a continuous goal; AI opens new development directions, and future work will aim for >90% recommendation accuracy and fully automated VOC matching without manual intervention.