Delivery and mobility / Ride-hailing and delivery
DiDi: Verifying multilingual support classifications, scoring compliance and analyzing complaint trends
- Company
- DiDi
- Country
- China
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
DiDi International Business Group checks service quality across ride-hailing, food delivery and financial services in Spanish and Portuguese. Its previous external QA tool did not expose the reasoning behind judgments and was difficult to update as language and business-specific standards changed. Operations staff had to read and tally conversations manually to identify causes of surging complaints.
Technology and data
The team normalizes chat and phone transcripts and feeds them into separate intent-verification, compliance-evaluation and customer-feedback pipelines on Amazon Bedrock. Verification sees only the current label; the full taxonomy is supplied for alternatives only after the label fails. Compliance rules come from configuration, while computable facts such as waiting times are calculated in code. Sensitive information is masked and rule-based checks are applied. Scores retain reasoning for human review. Customer feedback passes through per-conversation extraction, embedding-based clustering and report generation.
Results
AWS and DiDi reported that production validation raised intent-verification accuracy from 38% to 86% and achieved average compliance-scoring accuracy above 90%. In an analysis of cancellation-fee complaints in Latin America, they said manual reading and summarization that previously took hours was compressed into minutes. Operations teams receive reports identifying recurring causes and recommended actions.
Limits and open questions
The source does not disclose the accuracy-validation sample size, share of all conversations processed or monetary savings. Expansion to additional languages and business lines remains a follow-up plan.
Sources
Compiled from public sources. These are not results from ATF Works customers.