DiDi has built its own AI quality-assurance system for Spanish and Portuguese customer support across ride-hailing, food delivery and financial services. The Amazon Bedrock deployment replaces an opaque third-party tool with an architecture that exposes reasoning and can be updated as internal audit rules change.
Three pipelines divide the work. One verifies the reason assigned to each customer contact, another evaluates compliance, and a Voice of Customer pipeline groups similar tickets to find emerging trends. DiDi reports that production validation increased intent-verification accuracy from 38% to 86%, pushed compliance-scoring accuracy above 90%, and reduced trend-analysis work from hours to minutes.
The system normalizes chat and phone transcripts before processing them. Bedrock Guardrails masks personally identifiable information and checks grounding, while deterministic code recalculates facts such as wait times and verifies rule-based judgments rather than accepting every model output. Each score retains a reasoning trail for human review.
These figures come from DiDi and AWS’s own case study, so they are not an independent benchmark. Still, the design shows a practical pattern for using language models in audits: isolate context by task, preserve reviewable evidence and keep calculations outside the model when ordinary code can do them reliably.