How DiDi built intelligent contact center QA with Amazon Bedrock
Amazon Web Services Fei Huang, Raphael Hua
DiDi moved contact center QA onto Amazon Bedrock and ditched a black-box vendor. It now checks Spanish and Portuguese tickets with more traceable judgments and faster trend spotting.
Based on reporting by Amazon Web Services, Fei Huang, Raphael Hua — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DiDi has rebuilt its contact center quality assurance system on Amazon Bedrock, swapping out a third-party setup it couldn’t fully see into for one it owns itself. The system is for the company’s International Business Group customer experience team, and it handles Spanish and Portuguese tickets across ride-hailing, food delivery, and financial services.
The point of the overhaul is not just automation. It’s control. DiDi wanted QA decisions that could be traced, audited, and adjusted as its standards changed. The old setup was opaque, and manual spot-checking couldn’t keep up with the volume or the pace of change. So DiDi and AWS split the work into three pipelines: intent verification, compliance evaluation, and Voice of Customer analysis.
The intent pipeline checks whether a representative tagged a ticket with the right contact reason, and it also looks at tickets marked “Other” to find taxonomy gaps. A key lesson was that throwing the entire contact-reason tree at the model made it over-eager to find a “better” label. DiDi fixed that by isolating information in stages: first the model sees only the current label, and only if that fails does it get the full tree and the earlier reasoning. That shift pushed intent verification accuracy from 38 percent to 86 percent in production validation.
The evaluation pipeline takes a different tack. It uses a single prompt template that’s assembled at runtime from external configuration, so DiDi can cover multiple languages, business lines, and changing QA rules without rewriting everything each time. Amazon Bedrock’s Tool Use feature forces schema-valid JSON, and DiDi adds post-validation for rule-based checks, rather than trusting the model blindly. In production validation, compliance scoring went above 90 percent on average.
The third pipeline, VOC, is there for the operations team when they need to understand a surge in similar tickets. Instead of reading case after case by hand, the system groups them and produces structured analysis reports, turning hours of manual summarization into minutes. That’s the real shape of this project: not a flashy chatbot, but a QA machine built around context control, auditability, and less guesswork.
My take — AI-written commentary, not fact-checked reporting
This is the rare enterprise AI story that sounds sane: less magic, more containment. The industry keeps selling “automation,” while the useful breakthrough here is mostly refusing to let the model freeload on too much context. That’s not glamorous, but it’s how you get something audit-friendly instead of a very expensive confidence machine.
Read more about this at: Amazon Web Services