DiDi Enhances Contact Center QA with Amazon Bedrock

DiDi's International Business Group (IBG) collaborated with AWS to build an advanced contact center quality assurance (QA) system utilizing Amazon Bedrock. This new system replaces an opaque third-party solution, offering enhanced transparency and control. Covering Spanish and Portuguese across ride-hailing, food delivery, and financial services, the system comprises three core pipelines: intent verification, compliance evaluation, and Voice of Customer (VOC) analysis.

The intent pipeline verifies whether customer service representatives correctly assign contact reasons, improving accuracy from 38% to 86%. The evaluation pipeline assesses compliance and extracts business insights, achieving over 90% accuracy. The VOC pipeline aggregates large volumes of multilingual tickets, enabling rapid trend detection that previously took hours of manual effort.

DiDi chose Amazon Bedrock for its model-agnostic API, built-in governance, and security controls, which ensure data privacy and responsible AI practices. The system architecture involves dual-channel data ingestion, preprocessing, and parallel pipelines powered by Bedrock, each designed with specific principles such as information isolation and dynamic prompt assembly.

The intent verification pipeline employs a two-step process to improve classification accuracy by isolating tasks and controlling context, leading to a significant performance boost. The evaluation pipeline uses external configurations to dynamically assemble prompts, allowing coverage of multiple languages and business lines with a single template. The VOC pipeline employs a multi-stage process—parallel extraction, issue clustering, and report generation—to efficiently surface systemic issues.

Overall, the project demonstrates that precise context management and data control are crucial for reliable large language model applications. DiDi plans to expand this system to additional languages and business lines, further integrating insights and analytical capabilities.