DiDi replaced a third-party QA solution with an in-house system on Amazon Bedrock
DiDi International Business Group collaborated with AWS to build an in-house QA system for contact centers on Amazon Bedrock with three pipelines (intent verification, compliance, VOC analysis); intent verification accuracy increased from 38 % to 86 %.
DiDi International Business Group, the international division of DiDi Global operating in 14 countries across three business areas (transportation, food delivery, financial services), collaborated with AWS to build an in-house intelligent quality control (QA) system for contact centers on the Amazon Bedrock platform. The system replaced the previous opaque third-party solution and covers tickets in Spanish and Portuguese across both chat and phone calls.
The system consists of three separate pipelines: customer intent verification, compliance evaluation, and Voice of Customer (VOC) analysis. According to DiDi, in production deployment, intent verification accuracy improved from 38 % to 86 %, compliance evaluation accuracy exceeded 90 %, and VOC analysis reduced manual trend summarization from hours to minutes.
DiDi and AWS chose Amazon Bedrock for access to multiple foundation models through a single API, built-in security and governance controls (private connectivity through AWS PrivateLink, encryption, access management through IAM), and the Amazon Bedrock Guardrails tool for content filtering and masking sensitive data. According to the description, a key principle was precise context management, meaning control over what information the model sees in each call – the authors state that when the model was given the complete list of categories at once, it automatically compared them and often suggested a “more precise" alternative even for a correctly categorized ticket, reducing accuracy.
You can find details in the source article.
Why it matters
The case shows a concrete, measurable benefit of switching from an opaque external tool to an in-house system built on foundation models for customer support – higher accuracy of quality checks and significant savings in manual work when analyzing feedback. It also demonstrates a practical problem in LLM pipeline design (distorted decision-making when the entire list of categories is presented at once) and a solution through targeted context management.
Two audiences, two different impacts
What this means
For individuals
Developers of similar systems can learn from the observed effect that when an LLM receives the complete list of categories at once, it automatically compares them and often finds a “more precise" alternative even when the category has been assigned correctly, leading to incorrect reassignment.
For a business
Companies operating contact centers have a documented example showing that switching from a third-party solution to an in-house system built on foundation models can significantly improve the accuracy of quality checks and reduce manual analysis from hours to minutes.
ProcessesCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.