AI Agent Stack Industry Chain

Enablement Layer（应用使能层（上游））Sector — Everything an AI application buys or depends on: model access, agent infrastructure, data and memory, and engineering tooling. This layer sets an application's cost structure and capability ceiling.

• Model Access & Inference（模型接入与推理）Sector — How applications obtain model capability, and the largest variable cost for most AI products.

  • Proprietary Model APIs（闭源模型 API）ServiceBooming — Calling flagship models from the leading labs: strongest capability, fastest iteration, billed per token.🔒5

  • Open Models & Fine-tuning（开源模型与微调）SoftwareImproving — Domain fine-tuning and on-premise deployment of open-weight models, trading peak capability for cost control and data residency.🔒4

  • Model Routing & Gateways（模型路由与网关）SoftwareImproving — Routing requests across models by cost, latency and capability - the most direct lever applications have on unit economics.🔒3

  • Inference Providers（推理服务商）ServiceImproving — Hosted inference for open-weight models, billed by usage and competing on throughput and price-performance.🔒4

  • On-device Inference（端侧推理）SoftwareImproving — Running compact models locally on phones and PCs to address privacy, latency and offline scenarios.🔒4

• Agent Infrastructure（Agent 基础设施）Sector — The substrate that lets models call tools, run code and operate software - the fastest-growing part of this cycle.

  • Agent Frameworks（Agent 框架）SoftwareBooming — Orchestrating multi-step reasoning, tool calls and state transitions - the ceiling on how reliable an agent can be.🔒4

  • Model Context Protocol（工具调用协议 MCP）TechnologyBooming — An open protocol connecting models to external tools and data sources; now the de facto standard, collapsing integration cost.🔒3

  • Execution Sandboxes（沙箱执行环境）ServiceBooming — Isolated runtimes where agents execute code - the precondition for models actually doing work rather than just answering.🔒3

  • Browser & Computer Use（浏览器与计算机操作）SoftwareImproving — Letting agents drive web pages and desktop software like a person, reaching systems that expose no API.🔒3

  • Multi-agent Orchestration（多 Agent 编排）SoftwareImproving — Splitting work across specialised agents running in parallel, for complex and long-horizon workflows.🔒3

• Data & Memory（数据与记忆层）Sector — Connecting private data to models and letting agents retain context across sessions.

  • Vector Databases（向量数据库）SoftwareStable — Storing and retrieving vectorised semantic data; the base component of retrieval-augmented generation.🔒4

  • RAG Frameworks（RAG 框架）SoftwareStable — Assembling retrieved results into usable context while controlling recall quality and citation accuracy.🔒3

  • Document Parsing（文档解析与清洗）SoftwareImproving — Turning PDFs, tables and scans into structured text - the first gate for getting private data into a model.🔒4

  • Memory & Context Management（长期记忆与上下文管理）SoftwareBooming — Letting agents remember users and task state across sessions - the step from tool to genuine assistant.🔒3

  • Labeling & Synthetic Data（数据标注与合成数据）ServiceImproving — Producing high-quality samples for fine-tuning and evaluation, combining expert labeling with synthetic generation.🔒3

• Engineering & Observability（工程化与可观测）Sector — Turning non-deterministic model output into a measurable, operable production system - the real blocker for enterprise rollout.

  • Evaluation Systems（评测体系）SoftwareBooming — Quantifying model and agent performance to drive iteration and release decisions.🔒3

  • Tracing & Observability（追踪与可观测）SoftwareImproving — Recording the trace, cost and latency of every call to diagnose why an agent failed.🔒3

  • Prompt Management（提示管理与版本控制）SoftwareStable — Treating prompts as code with versioning, staged rollout and rollback.🔒3

  • Guardrails & Safety（护栏与内容安全）SoftwareImproving — Blocking harmful output, unauthorised tool calls and prompt injection - a compliance precondition for public-facing applications.🔒3

• Build & Delivery Platforms（开发与交付平台）Sector — Lowering build and billing complexity so non-specialist teams can ship AI applications.

  • AI App Platforms（AI 应用开发平台）SoftwareStable — Low-code orchestration of models, knowledge bases and workflows to assemble a usable application quickly.🔒4

  • Usage Metering & Billing（用量计量与计费）ServiceImproving — Infrastructure for billing by token, call or task outcome - the backbone of AI-native pricing.🔒3

Application Products（应用形态（中游））Sector — The AI products that deliver value to users, grouped by product form rather than industry.

• General Assistants（通用助手）Software — Consumer-facing conversational assistants, the first AI touchpoint for most people.

  • Consumer Chatbots（消费级对话助手）SoftwareSoftening — Mostly free general-purpose chat products monetised through scale.🔒5

  • AI Companions（AI 陪伴与角色）SoftwareStable — Long-running interaction products built around companionship and role-play.🔒3

• AI Coding（AI 编程）Software — The most commercially proven AI application category, with the strongest willingness to pay.

  • AI-native IDEs（AI 编程 IDE）SoftwareBooming — Development environments with completion, refactoring and chat built deep into the editor.🔒4

  • Coding Agents（编程 Agent）SoftwareBooming — Agents that read code, edit files, run tests and submit changes autonomously once given a task.🔒4

  • AI Code Review & Testing（代码审查与测试）SoftwareImproving — Automatically finding defects, generating test cases and gating pull requests.🔒3

• AI Search（AI 搜索问答）Software — Information retrieval that returns generated answers instead of a list of links.

  • AI Search Engines（AI 搜索引擎）SoftwareImproving — Live retrieval plus synthesis, returning a cited answer directly.🔒3

  • Enterprise Search（企业内部搜索）SoftwareImproving — Unified semantic search across internal documents, email and tickets.🔒3

• Generative Content（生成式内容）Sector — Applications that directly produce text, image, video and audio assets.

  • Text & Writing（文本与写作）SoftwareSoftening — Tools for marketing copy, long-form writing and rewriting.🔒3

  • Image Generation（图像生成）SoftwareStable — Text-to-image and image-to-image generation serving design, e-commerce and advertising production.🔒4

  • Video Generation（视频生成）SoftwareBooming — Text-to-video and image-to-video generation, reshaping how short video and ad creative get produced.🔒5

  • Voice & Music（语音与音乐）SoftwareImproving — Speech synthesis, voice cloning and music generation.🔒3

  • 3D & Digital Humans（3D 与数字人）SoftwareImproving — Digital-human livestreaming, virtual avatars and 3D asset generation.🔒3

• Agent Products（Agent 产品）Software — Products that plan and execute multi-step tasks autonomously - the main direction of travel for the application layer.

  • General-purpose Agents（通用 Agent）SoftwareBooming — Assistants that take an open-ended goal, decompose it and call tools to complete it.🔒3

  • Vertical Agents（垂直 Agent）SoftwareBooming — Agents that deeply automate a single function such as support, sales, recruiting or finance.🔒3

  • AI-augmented RPA（流程自动化 RPA+AI）SoftwareImproving — Classic process automation augmented with model understanding to handle unstructured input.🔒3

• Enterprise Knowledge & Productivity（企业知识与办公）Software — AI embedded in everyday enterprise workflows: the broadest reach, though shallower value per touchpoint.

  • Office Copilots（办公 Copilot）SoftwareImproving — AI assistants embedded in documents, spreadsheets, email and meetings.🔒4

  • Enterprise Knowledge Bases（企业知识库）SoftwareImproving — Consolidating scattered internal material into a queryable knowledge hub.🔒3

  • Meeting & Collaboration AI（会议纪要与协作）SoftwareStable — Meeting transcription, summary generation and action-item extraction.🔒3

Adoption & Distribution（落地与分发（下游））Sector — Which industries ultimately buy AI applications, and the channels through which they reach customers.

• Industry Adoption（行业落地）Sector — Adoption depth varies sharply by industry, driven by data availability and tolerance for error.

  • Financial Services（金融）ServiceImproving — Automating research, risk control, customer service and compliance review.🔒4

  • Healthcare（医疗）ServiceImproving — Imaging-assisted diagnosis, clinical documentation, drug discovery and trials.🔒4

  • Legal（法律）ServiceBooming — Contract review, case research, due diligence and drafting.🔒3

  • Education（教育）ServiceStable — Personalised tutoring, grading and course content generation.🔒3

  • E-commerce & Retail（电商零售）ServiceImproving — Support automation, product creative generation, personalised recommendation and guided selling.🔒3

  • Manufacturing（制造工业）ServiceStable — Inspection, scheduling, predictive maintenance and process knowledge capture.🔒3

  • Marketing & Advertising（营销广告）ServiceBooming — Bulk creative generation, campaign optimisation and content localisation.🔒3

  • Gaming & Entertainment（游戏娱乐）ServiceImproving — Art asset production, NPC dialogue and narrative generation.🔒3

• Distribution Channels（分发渠道）Service — The routes by which AI applications reach end customers, which shape acquisition economics.

  • App & Plugin Marketplaces（插件与应用市场）ServiceImproving — Secondary distribution inside assistants, IDEs and office suites.🔒3

  • Cloud Marketplaces（云市场）ServiceStable — Selling through hyperscaler marketplaces, drawing down a customer's committed cloud spend.🔒3

  • System Integrators（系统集成商）ServiceStable — Integrators handling delivery and customisation for large enterprise and public-sector buyers.🔒3