At the Amazon Web Services Summit China, the Global Vice President and APAC Co-President of Amazon Web Services stated that AI Agents are set to fundamentally reshape collaborative relationships, driving a paradigm shift in enterprise organization and value creation. He noted that AI is evolving from an auxiliary tool into a genuine productive force, with its unit of measurement for value consequently changing. AI is no longer just about answering questions or improving efficiency; it is beginning to deliver measurable business outcomes directly for enterprises.
However, faced with a constant stream of dazzling AI innovations, businesses require a clear roadmap to identify their specific needs. To address this, he outlined a five-layer AI technology stack map necessary for enterprises to achieve Agentic business transformation.
Foundational Infrastructure Layer
The first layer is the AI infrastructure layer, encompassing GPUs, AI accelerator chips, and the associated networking and storage. This serves as the foundational bedrock for all technological implementation. The business value of this layer lies in providing ample computing power for models. For most enterprises, there is no need to purchase chips or build data centers themselves; these underlying infrastructure concerns are the domain of model service providers and cloud vendors.
Model Layer
The second layer is the model layer. Innovation in large language models remains in a phase of rapid iteration with alternating leaders. The frontier includes various top-tier commercial models as well as numerous powerful open-source models, a significant portion of which originate from China. The value of this layer is providing the "brainpower" for agents. He likened models to talent, pointing out that different roles and tasks require people with different backgrounds and capabilities. Model selection is crucial for enterprise applications. Enterprise Agentic AI applications should choose the most suitable model based on actual needs regarding intelligence level, speed, and cost, and crucially, should avoid self-imposed limitations by not locking into a single vendor's model.
Data and Knowledge Layer
The third layer is the data and knowledge layer. The core value here is to provide agents with relevant, accurate, fresh, and well-governed high-quality data support. He candidly stated that the failure of the vast majority of enterprise AI projects can be attributed to data not being ready. For Agentic AI applications to create truly unique and non-replicable competitive barriers for a business, they must deeply integrate and effectively utilize the proprietary data and knowledge accumulated by the enterprise over many years. High-quality data forms the barrier for enterprise differentiation: because a company's data is its own, accumulated over time and cannot be quickly replicated by others.
Agentic Platform Layer
The fourth layer is the Agentic platform layer. When an enterprise scales from a few pilot agents to hundreds or thousands of agents working in coordination, a platform is needed to empower and manage them. This is analogous to a startup establishing management systems and HR mechanisms as it grows. The business value of this layer lies in providing a unified runtime environment and development tools for agents, along with management capabilities such as rules, evaluation, and governance. The key is whether it can effectively support the development, deployment, management, and iteration of agents at scale. This layer also represents the dividing line between moving agent applications from proof-of-concept to production.
Agents and Application Layer
The fifth layer is the agents and application layer. This is where Agentic AI truly creates value and delivers business outcomes for the enterprise. These applications include cross-industry general-purpose scenarios such as software development, IT operations, knowledge work, and customer service, as well as industry-specific or customized applications that can deliver the greatest business value. Agents function like a company's digital employees, working for the enterprise and delivering tangible business results.
Cross-Cutting Considerations
He further pointed out that beyond these five technology stack layers, four dimensions—security, effectiveness, performance, and cost—are critical throughout, requiring consideration at every layer from top to bottom. First is security, covering model safety, data security, and the permission boundaries of agents; this is the baseline for agent implementation. Second is effectiveness, referring to the quality of the agent's output. Every layer, from model selection and data quality to platform guardrails and agent design/implementation, influences the final outcome. Third is performance, including metrics like response speed, throughput, and latency, which are essential for agents serving a massive user base in a production environment. Finally, there is cost. Enterprises must be able to trace costs from the agent all the way down through the platform, data, and model tokens. If the actual cost of an agent completing a single task cannot be clarified, what the enterprise is doing is not deployment, but an experiment.
He emphasized that the ultimate purpose of all five technology stack layers is to serve the business output at the very top. The sole criterion that truly determines the success or failure of an Agentic project is measurable business output. Enterprises must quantify the business value delivered by Agents, whether it's task completion throughput, output quality, cost per task, delivery cycle time, or more macro-level metrics like human-equivalent effort, customer satisfaction, and revenue growth.
Comments