Aivis: Pioneering Autonomous Agents for Alibaba Cloud’s Next‑Gen Intelligent Services

Why : In Alibaba Cloud’s service domain the authors identify a fundamental “impossible triangle” – the need to deliver ultra‑high user experience, keep costs low, and handle extremely complex service scenarios. Customers now expect instant responses, rapid problem resolution, precise answers, and even emotional empathy, while the organization processes hundreds of thousands of complex scenarios, millions of tickets, tens of millions of dialogs, and serves over five million customers.

What : The service evolution is described in three stages. Tool 1.0 (human‑AI collaboration) relied on BERT‑style understanding models and a simple “ask‑answer” interaction, focusing on self‑service rate. Partner 2.0 emerged with GPT‑style generative models, shifting to a “I do, you help” copilot model and emphasizing reduced human handling time. Teammate 3.0 introduces autonomous agents that “you do, I manage”, driven by Agent technology that expands AI decision‑making and execution, with the north‑star metric moving to per‑person service volume.

How – Architecture : A four‑layer “human‑machine collaborative intelligent service system” is presented.

Data Capability Layer : Massive domain‑specific service data is consolidated as the foundation for intelligence.

Model Capability Layer : A hybrid strategy combines a large “service‑domain model” with specialized “scenario‑specific small models” to balance generality and expertise.

Platform Capability Layer : An Agent platform enables large‑scale creation, deployment, and management of autonomous agents.

Service Form Layer : Built on the platform, multiple service forms – Chatbot, Copilot, and the new Aivis – can be flexibly delivered.

A closed‑loop feedback and evaluation mechanism continuously iterates the system.

Practice 1 – Enhancing Large‑Model Capability : The authors first identify gaps of generic LLMs in the service domain (lack of domain‑specific data, implicit expert experience, and reasoning ability) by mapping scenarios into three zones based on complexity and automation potential. They then design a domain‑model fine‑tuning pipeline that includes data synthesis, construction, varied training strategies, and rigorous evaluation. A key insight is that raw, noisy business data cannot be fed directly to the model; instead, high‑quality synthetic task data are generated to address cold‑start and generalization issues.

Practice 2 – Building Execution Ability : The next challenge is turning the “brain” into action. The proposed Agent architecture consists of three core modules: Memory (stores interaction context), Tool (invokes external APIs), and Planning (orchestrates steps). A concrete example shows how a password‑reset request is handled: Planning decides the sequence (identity verification → authorization → reset → notification), Tool calls the respective APIs, and Memory records the operation for future context.

Practice 3 – Organizational Integration : To embed these capabilities, the service organization introduces “digital employees” – the Aivis agents – that work alongside L1‑L3 human experts. New roles are defined: Cloud Assistant Designer (designs and composes agents), Cloud Assistant Trainer (optimizes agent performance), and Cloud Assistant Supervisor (monitors execution). A simplified ticket‑flow diagram illustrates how inbound requests are first handled by self‑service tools and chatbots, then routed to the new human‑machine collaborative service team when needed.

Action Blueprint : Based on the exploration, three practical steps are recommended for peers building enterprise AI services.

Step 1 – Single‑Point Breakthrough : Target high‑value, high‑frequency, low‑complexity scenarios to deliver quick wins and build confidence.

Step 2 – Process Integration : Deeply embed AI capabilities into business systems, creating a virtuous “human‑AI collaboration flywheel” that encourages experts to both use and teach the AI.

Step 3 – Scale Human‑Machine Collaboration : Define the digital employee’s role, design layered collaboration architectures, and establish flow mechanisms to construct an efficient intelligent service organization.

The authors conclude that applying large‑model AI to enterprise services represents a fundamental transformation of service teams, echoing a quote from former Alibaba chief strategist Zeng Ming: “Two core AI metrics are the proportion of business run independently by AI and the proportion of employees who are silicon‑based.” They position Aivis as a concrete practice of the conference theme “Cloud‑Intelligence Integration, Carbon‑Silicon Co‑existence.”