How AliExpress Created the First AI‑Powered Cross‑Border Voice NPS Agent

Problem

Cross‑border commerce receives tens of thousands of multilingual raw‑voice feedback daily from social media, app stores and support channels. Manual processing is slow and cannot capture real‑time user needs.

Solution Overview (Global Experience Platform)

GEP builds a data pipeline that collects raw‑voice from internal sources (MyAE, search, app market, CCO feedback, user‑research questionnaires) and external social media (Reddit, X, Facebook). Offline ODPS batch jobs and real‑time MetaQ streams form a hybrid batch‑plus‑stream architecture. An AI‑augmented refinement workflow cleans, annotates and standardises the voice data into a reusable asset.

Agent Engine

The core cognition hub integrates multimodal large models Qwen‑3, Deepseek‑R1 and Claude‑4 with a Retrieval‑Augmented Generation (RAG) knowledge‑search framework. Custom knowledge graphs and semantic‑matching strategies can be added via the AppBuilder visual orchestration platform. In a test set of 1 000 samples the engine achieved 99 % classification accuracy and 99 % consistency.

Data Sources

In‑site channels – thousands of items per day, text + partial screenshots, integrated.

Social‑media (Reddit, X, Facebook) – hundreds of thousands per day, text + images/video, integrated.

User‑research questionnaires – thousands per day, text + images, integrated.

App‑store feedback – thousands per day, text, integrated.

CCO user feedback – hundreds of thousands per day, conversational text + images, integrated.

Core Capabilities

Multilingual processing.

Accurate voice classification.

Automatic alert generation and diagnosis.

Problem‑resolution tracking.

Multimodal Translation Agent

Combines the MCP image‑to‑text translator with Qwen‑3 and Deepseek consistency checks. Supports translation from 17 source languages for text and 11 source languages for images into English.

Voice Classification Agent

Taxonomy

Three‑level hierarchy: primary business domains (logistics, after‑sale, order, payment); secondary groups (search issues, coupon problems, address validation, etc.); fine‑grained sub‑categories (e.g., “order limit‑purchase”, “high shipping fee”).

Challenge 1 – Precision

Feedback often mismatches the actual problem. Mitigations: add prompt constraints, lower model temperature, inject domain‑specific examples, run parallel models to reduce hallucinations.

Challenge 2 – Completeness

Three‑step taxonomy expansion over two months:

Manually preset >200 third‑level categories.

AI agents expand to >1 000 categories based on inference.

Iterative pruning reduces the set to ~400 categories while preserving granularity.

Evaluation

Three‑layer verification: business domain 99 %, scenario group 99 %, fine‑grained class 97 %.

Human‑in‑the‑loop correction: initial AI judgment → expert calibration (≈1 %) → model reinforcement (3‑day cycle), yielding ~300 edge‑case optimisations per month.

Confidence gating: cases with confidence ≥80 % are automated (≈3 % of total); lower‑confidence cases are routed to experts for review and model fine‑tuning.

Alert Platform

Design

More than 350 alert rules are assigned unique TaskIDs. A modulo‑60 hash (TaskID % 60) distributes checks across 60 time slots, preventing simultaneous triggers.

Performance Impact

Peak alerts per second reduced from >350 to ≈6.

Database connection‑pool utilisation lowered from 98 % to below 15 %.

Server CPU variance narrowed to ±5 %.

Attribution Agent

Based on the MCP protocol, expert troubleshooting knowledge (error‑code analysis, API checks, business‑log inspection, change impact analysis) is stored in a knowledge base. The agent selects appropriate diagnostic tools and generates root‑cause reports with confidence and impact assessments.

Results

GEP aggregates feedback from personal centre, search and experience questionnaires, providing multi‑dimensional top‑issue views with translated voice details. Classification feeds trigger alerts that are deployed in social‑interaction and app‑usability domains; AI diagnosis is live in the transaction domain. Hundreds of tickets have been auto‑generated, achieving a 90 % resolution rate and improving user experience.