How Alibaba’s Aivis Agent Is Transforming Cloud Customer Support

Introduction

The talk introduces "Aivis", Alibaba Cloud’s digital employee that assists customers under the supervision of real technical experts, aiming to improve post‑sale service efficiency.

Why Build Aivis

Traditional expert‑driven support faces challenges: deep technical breadth, fragmented tools, and ambiguous customer queries, leading to long response times and heavy expert workload. Aivis is designed to break the human‑capacity ceiling.

Overall Architecture

The system is organized into four layers: Data Capability, Model Capability, Platform Capability, and Service Form Layer, each handling data processing, model training, tool integration, and user‑facing services respectively.

Core Module Best Practices

Accurate intent recognition and fluent answer generation.

Efficient knowledge retrieval and inference.

Autonomous execution of predefined business processes.

Agent Mechanism

Aivis uses a three‑component agent: Planner (splits complex queries into sub‑tasks), Reasoner (the brain that decides next actions), and Executor (calls specific tools). Multiple specialized agents (Function Code Agent, RAG Agent, etc.) cooperate to handle diverse scenarios.

Data Layer

Historical tickets are parsed into customer behavior, service behavior, and environment state, then labeled by complexity, frequency, and scenario type (diagnostic, operational, consultative). This creates a structured, searchable knowledge base for training and inference.

Model Training

Domain‑specific models are fine‑tuned using large‑scale multi‑turn dialogues and task trajectories. Techniques include Prompt Engineering, Context Engineering, supervised fine‑tuning, reinforcement learning with reward models, and DPO for aligning with expert preferences.

Autonomous Reasoning

Aivis adopts a dynamic workflow (Agentic Reasoner) rather than static if‑else flows, allowing it to compose atomic modules on‑the‑fly for complex, high‑frequency scenarios such as ECS connectivity issues.

Trust and Human‑Machine Collaboration

Transparency is achieved through visualizing actions (screenshots, videos) and the reasoning process. A "human‑guides‑machine" workflow hands over unresolved tickets to experts, who then feed back their decisions to continuously improve the agent.

Q&A Highlights

Key metrics include independent ticket resolution rate, adoption rate, action accuracy, and tool‑call accuracy. The three‑dimensional classification (complexity, frequency, scenario) drives routing and solution strategies.