The Well-Architected Framework AI Lens is an extension of the Alibaba Cloud Well-Architected Framework for artificial intelligence. It systematically explains how to design, build, run, and continuously optimize AI workloads on Alibaba Cloud. It covers the entire technology stack, from model selection, inference architecture, and data pipelines to agent orchestration. This lens also provides specialized guidance for AI scenarios based on the five pillars: Security, Reliability, Cost Optimization, Operational Excellence, and Performance Efficiency. The content covers core architectural patterns such as Large Language Model (LLM) inference services, Retrieval-Augmented Generation (RAG) systems, multi-agent collaboration, model fine-tuning and evaluation, and AI data engineering, supplemented with actionable decision frameworks, design checklists, and best practices for Alibaba Cloud products.
The unique nature of AI workloads
Traditional cloud workloads—whether web applications, databases, or big data pipelines—are built on a deterministic foundation: given the same input, the system produces the same output; performance can be improved near-linearly through horizontal scaling; cost is directly tied to computing resources; and network isolation and access control clearly define security boundaries.
AI workloads break these assumptions. A model's output is probabilistic, meaning the same question can yield different answers. Quality is no longer a binary judgment (correct or incorrect) but a multi-dimensional spectrum. Cost depends not only on computing power but also on the product of model size, context length, and call frequency. The security boundary extends from the network layer to the semantic layer, where a carefully crafted natural language input, such as a prompt injection, can bypass traditional defenses.
Understanding these fundamental differences is the starting point for designing a well-architected AI system. Although the examples in this framework primarily focus on text generation (LLM) scenarios, the architectural principles—including the design concepts of the five pillars, decision frameworks, cost models, and governance mechanisms—also apply to multimodal AI systems, such as visual understanding, voice interaction, video generation, and mixed-modality inference. The differences in multimodal scenarios primarily lie in the implementation of the input preprocessing pipeline (like video frame extraction and audio segmentation), inference resource requirements (multimodal models often require larger VRAM and more computation), and output moderation strategies (image and video content moderation requires specialized models). However, the high-level architectural decision logic remains the same. When applying the architectural recommendations in this document, "token" can be generalized to "inference unit" (which includes image patches, audio frames, etc.), and "prompt" to "multimodal input orchestration." The underlying principles can then be applied directly.
Relationship with the Alibaba Cloud Well-Architected Framework
The Alibaba Cloud Well-Architected Framework defines five design pillars: Security, Reliability, Cost Optimization, Operational Excellence, and Performance Efficiency. This framework provides general architectural design principles and review methodologies for all cloud workloads and is used by thousands of enterprises for cloud architecture governance. This whitepaper does not start from scratch; instead, it adds a layer of specialized guidance for AI workloads on top of the five pillars. The unique nature of AI systems is that while traditional pillars focus on "Are resources sufficient and correctly configured?", the AI perspective requires architects to further scrutinize "Is the model trustworthy, is inference cost-effective, and are agents controllable?". When discussing Security, for example, a traditional cloud workload focuses on network isolation and access control. In an AI scenario, you must also answer whether model weights have been verified for provenance and whether prompt injection can bypass existing defenses. Similarly, when discussing Cost Optimization, a traditional workload measures resource utilization, while an AI workload must determine if each token generates equivalent business value and whether model routing minimizes inference costs without sacrificing quality.
The core "Learn → Measure → Improve" methodology of the Alibaba Cloud Well-Architected Framework also applies to AI scenarios, but you must expand the measurement dimensions. In addition to traditional metrics like availability, latency, and throughput, you must also measure model quality (accuracy, hallucination rate, consistency), token economics (token consumption per unit of business value), guardrail effectiveness (interception rate and false positive rate), and compliance coverage (algorithm filing completion rate and content moderation pass rate).
The Agentic Cloud Governance Center is the operational hub for implementing this cycle. By extending the Agentic Cloud Governance Center with AI-specific governance baselines and evaluation rules, enterprises can transform the architectural recommendations in this whitepaper into automatable governance metrics. Enterprises can then continuously track AI architectural maturity as it evolves from a foundational to an optimized state.