From Manual to AI‑Driven: How a Multi‑Agent Autonomous Platform Revolutionizes Rights Operations

Background

Rapid growth of coupon marketing on AliExpress increased the complexity of operations across 20+ key countries and 200+ regions. Manual configuration caused high cost, error‑prone processes, slow communication, and insufficient data analysis.

Four Core Pain Points

High communication cost between the platform team and country teams.

Configuration errors due to manual, high‑frequency adjustments.

Slow abnormal response because of manual monitoring.

Lack of granular data analysis per country.

Breakthrough Idea: AI‑Co‑ordinated Decision‑Making

Shift from “platform team defines → country team executes” to “platform team defines → system automatically executes”, reducing cost, improving efficiency, and enabling single‑person full‑process management.

Product Capability Thinking – Five Core Abilities

Automated configuration execution across multiple countries and tiers.

Intelligent Q&A and validation via knowledge base and diagnostic model.

Smart effect analysis that auto‑generates reports from usage, conversion and behavior data.

Intelligent strategy generation that proposes optimization suggestions.

Data‑analysis monitoring with 24/7 anomaly detection, rule‑based alerts and visual dashboards.

Why a Digital‑Human Interface?

Compared with traditional form‑based UI, a conversational “digital person” lowers the usage threshold, speeds up work, adapts to unknown scenarios, and improves user experience.

Technical Selection – Agentic Agent

Evaluated traditional coding, Workflow Agent, and Agentic Agent. Chose Agentic Agent for higher business adaptability, stronger evolution capability, better UX, and long‑term cost advantage.

Multi‑Agent Detailed Design

The system comprises Activity‑Creation Agent, Placement‑Creation Agent, Data‑Analysis & Alert Agent, Preview & Acceptance Agent, Intent‑Recognition Agent, Memory‑Management Agent, and Stability Infrastructure. Each Agent encapsulates tools and can plan, execute, and reflect.

Automation of Activity & Placement Creation

Natural‑language commands create 20‑country campaigns in one sentence; shared templates reduce rule duplication; digital‑human interaction replaces high‑barrier form UI.

Human‑Machine Collaboration for Interrupt‑Recovery

Agents pause when information is incomplete, request clarification, then resume precisely, avoiding long dialogues and token waste.

Template Execution with ReAct + planAct

Workflow templates guarantee deterministic high‑risk writes while allowing Agents to perform dynamic reasoning within the template.

Data Analysis & Alert Agent

Uses a master‑sub Agent ReAct architecture to decide between anomaly investigation and report generation, enabling minute‑level alerts.

Intent Recognition via Slot‑Filling

Consolidates thousands of query patterns into 19 core slots, reducing knowledge‑base maintenance and raising intent accuracy from ~60% to 100%.

Layered Memory Management

Short‑term, mid‑term, and long‑term memories isolate context for single‑step inference, full‑process tracking, and knowledge retrieval, eliminating hallucinations.

Stability Foundations

Business‑identity virtual threads isolate resources; full‑link tracing and semantic metrics pinpoint errors and performance bottlenecks.

Future Outlook

Rights Agent will focus on personalized intent‑driven coupon generation, effect analysis, and end‑to‑end automated execution, aiming for single‑operator full‑process management during major promotions.