AI智能体在运营商的应用与商机白皮书
The Underlying Drivers of Change in the 2.1 Industry
The Underlying Drivers of Change in the 2.1 Industry
The global telecommunications industry is undergoing a profound structural change. The full commercialization of 5G networks, the deep coverage of gigabit optical networks, and the acceleration of 6G technology research have brought unprecedented connectivity, while also making the network complexity exponentially increased. The traditional network operation and maintenance mode relying on artificial experience has been difficult to cope with the management requirements of ultra-large-scale networks-the introduction of network function virtualization (NFV), software defined network (SDN), multi-access edge computing (MEC) and other technologies, so that the network from hardware definition to software definition, the operation mode must also be from "manual drive" to "intelligent drive".
At the same time, AI technology itself is undergoing a qualitative leap. The breakthrough of the big language model represented by ChatGPT, GPT-4 and DeepSeek marks the AI from "perceptual intelligence" to "cognitive intelligence" and "action intelligence". AI evolved from passively answering questions to actively sensing the environment, making plans, invoking tools, and performing tasks-this is the core ability of AI agents.
From the perspective of technological evolution, the industry is experiencing three levels of deep integration: first, the integration of AI and communication networks, that is, the network itself has intelligence; The second is the integration of AI and business operations, that is, the use of intelligence to reconstruct the operation process; The third is the integration of AI and business models, that is, intelligence itself becomes a saleable service product. The integration of these three levels is superimposed on each other and is fundamentally changing the competition logic of operators.
Strategic Transformation Demands of 2.2 Operators
For operators, the maturity of AI agent technology is at the right time. The industry faces three structural pressures:
First, growth pressure. Revenue growth in traditional communications services has slowed and traffic dividends have peaked. According to the 2025 public financial reports of the three major operators, although the overall revenue remained stable, the growth rate of basic communication services has slowed down significantly, and the industry urgently needs to find new growth engines.
Second, cost pressure. 5G network construction and operation and maintenance costs continue to be high, and the traditional manual operation and maintenance model is inefficient in the face of ultra-large-scale networks. Without the introduction of AI automation, traditional network operating costs will be unsustainable. This cost pressure is not unique to Chinese operators-major operators around the world face similar OPEX control challenges, which makes the cost-effective value of AI intelligence more prominent.
**Third, competitive pressure. * Internet cloud service providers and AI native enterprises are eroding the traditional value space of operators. Operators must transform from "selling connections" to "selling capabilities"-computing power, models, and smart body services-or face the risk of being pipeline. It is worth noting that operators have unique advantages in this competition: network infrastructure across the country, mass user access, and reliable government-enterprise customer relationships, which are difficult for Internet companies to replicate.
2.3 the objectives and scope of this white paper
The purpose of this white paper is to provide a systematic reference framework for operator decision makers, technology managers and industry partners, covering the following core issues:
-The technical definition and classification system of AI intelligence.
-Operators in the field of AI intelligence strategic layout and the latest practice.
-Value analysis and business measurement of core application scenarios
-Technical architecture, implementation path and risk governance recommendations
-Evolution Outlook for Self-intelligent Network L4/L5
- AI agents: definition, classification and technological evolution.
3.1 Paradigm upgrade from AI models to AI intelligences
The traditional AI model is essentially an "input-output" system: instructions are received and results are returned. The AI agent (AI Agent) is an autonomous system that can:
-Perceive environment: Obtain external information through sensors, APIs, data streams, etc.
-Reasoning decision-making: using large language models for semantic understanding, logical reasoning, and scenario planning
-Execute action: Call tools, APIs, code executors, etc. to complete specific tasks
-Learning and Memory: Accumulate experience in the interactive process and continuously optimize behavioral strategies
-Collaborative communication: information exchange and task collaboration with other intelligences.
This paradigm upgrade can be summarized as a three-stage evolution from "co-pilot"(Copilot) to "agent"(Agent) to "multi-agent system"(Multi-Agent System). The core difference is that in the Copilot mode, the AI is only a consultant to assist human beings in making decisions, while in the Agent mode, the AI has the ability to complete tasks independently within the scope of authorization, while the multi-agent system further realizes the division of labor and cooperation between multiple Agents, thus solving complex cross-domain problems that are difficult for a single Agent to handle.
From the perspective of industrial practice, the AI intelligence in the operator scenario is rapidly crossing these three stages: the Copilot assistance in the field of customer service has become more mature, the single-Agent closed-loop optimization in the field of network operation and maintenance is landing on a large scale, and the multi-intelligent system of cross-domain collaboration represents the next stage of evolution.
3.2 Classification system for AI intelligences
Based on industry practices and functional characteristics, AI agents can be classified from the following dimensions:
Classified by degree of autonomy:
| Category | Feature | Typical Scenario |
|---|
The auxiliary intelligent body runs within the system's preset rule framework and authorization boundaries, provides analysis suggestions, decision support and operation assistance for human operators, and does not directly perform high-risk operations. Copilot-assisted operation and maintenance, standardized customer service response, and intelligent work order dispatch.
Autonomous agent has independent perception reasoning and decision execution capabilities, can complete tasks in a closed loop within a given authorization range, and can dynamically adjust strategies according to environmental changes. Self-intelligent network L4 level closed-loop optimization, dynamic network slicing arrangement, automatic fault repair.
The core dividing line of this classification lies in the "attribution of decision-making power"-the auxiliary intelligent body makes the final decision by the person, and the autonomous intelligent body makes its own decision within the preset safety fence. In terms of evolutionary trends, as technology maturity and trust increase, more and more scenarios will migrate from auxiliary to autonomous, but critical network operations and scenarios involving significant business impact will retain human approval for the long term.
Classified by functional level:
| Hierarchy | Responsibilities | Examples |
|---|
Planning and Scheduling Layer, Global Task Understanding and Decomposition, Multi-Agent Scheduling, Resource Allocation, Intent Resolution and Policy Generation Agent.
Collaboration layer, communication coordination and resource convergence between cross-domain intelligences, to ensure the exchange of information and consistent action of different domain intelligences, cross-domain fault location intelligences, end-to-end business support intelligences.
Execution layer, specific task execution, status collection and result feedback in a single domain, base station parameter tuning intelligence, complaint analysis intelligence, alarm processing intelligence.
The design logic of this three-level architecture lies in "layered decoupling and synergy"-the planning and scheduling layer is responsible for the decision of "what to do", the coordination layer is responsible for the coordination of "how to cooperate", and the execution layer is responsible for the action of "concrete. The layers are docked through standardized communication interfaces, so that each layer can evolve and optimize independently, while ensuring the flexibility and scalability of the overall system.
3.3 key technology enable
The implementation of AI intelligences relies on the convergence of a number of core technologies:
-Large Language Model (LLM): Provides basic capabilities for natural language understanding and reasoning planning.
-Retrieval Enhancement Generation (RAG): Combine corporate private knowledge bases with LLM to improve the accuracy of answers in specialized areas
-Tool call (Function Calling / Tool Use): Enables the smart body to call external APIs, databases, code executors.
-Chain-of-Thought: Solve complex problems through multi-step reasoning
-Model Context Protocol (MCP): Standardize the interaction interface between the intelligent body and the tool/resource.
-Inter-intelligent communication (A2A): to achieve standardized collaboration between multi-intelligent bodies.
3.4 AI agent vs traditional automation
Compared to traditional rule-based automation (RPA, script automation), AI intelligences are fundamentally different:
| Dimension | Traditional Automation | AI Intelligence |
|---|---|---|
| Trigger method | Fixed rule trigger | Autonomous perception and judgment |
| Adaptability | Scenario change is failure | Dynamic learning and adaptation |
| Processing Scope | Structured, deterministic tasks | Unstructured, nondeterministic tasks |
Collaboration capabilities, isolated execution, multi-intelligence collaboration.
| Optimization method | Manual adjustment rules | Autonomous learning optimization |
##4. operator AI strategic layout panorama
4.1 AI Strategic Positioning of China's Three Major Operators
2025-2026 is the key window period for the three major operators to upgrade their AI strategies. From the annual work meeting and public financial disclosure, AI has become the core of their respective strategies.
China Telecom: "Cloud Change Number to Wisdom" Strategic Upgrade
China Telecom upgraded the development strategy of "cloud to number" to "cloud to number to wisdom" and clearly put forward "embracing artificial intelligence in an all-round way". Key initiatives include:
-Build a five-in-one intelligent cloud system of "computing power, platform, data, model and application"
-The total scale of self-owned and access intelligence reached 91 EFLOPS (according to public financial reports)
-Build more than 100 industry models and more than 300 industry intelligences.
-Central enterprises AI penetration rate of 85%
-Jointly launched with Huawei the "Network Excellence Task Model" and deployed more than 900 AI intelligences in 31 provinces.
-According to public reports, it has won TM Forum's "Self-intelligence Network Excellence Award" for two consecutive years (the world's comprehensive score is the first)
-Launch of the Star Super Agent, integrating large models, tool invocation, RAG, multimodal perception and autonomous planning capabilities.
-Released the industry's first AI intelligence security governance white paper.
China Mobile: "Three Business Layout" and AI Terminal Strategy
China Mobile for the first time clear "communication services, computing services, intelligent services" three major business layout:
-The total size of intelligence reached 92.5 EFLOPS (according to public financial reports)
-Upgrade the nine-day model 3.0 and launch the "rhinoceros intelligence"
-Revenue from smart computing services grew by 279 per cent (according to public financial reports)
-Joint industry partners incubate more than 50 AI terminals
-Plan to promote 15 million AI terminals on a large scale in 2026
-Based on the open source Hongmeng to build the "China Mobile Zhihong" operating system, launch the "Ten Million" action plan.
-Jointly launched an intelligent complaint analysis agent with ZTE to shorten the fault identification time to less than 30 seconds
-Launched 1 yuan 400000 Tokens universal AI service in Shanghai to realize "one universal, cross-platform use and telephone fee payment"
China Unicom: "Four Racetracks" and Self-Intelligence Network
China Unicom focuses on the four major tracks of "connection, computing power, service and safety:
-Intelligence scale reaches 45 EFLOPS (according to public financial reports)
-AI revenue increased by more than 140 percent year-on-year (according to public financial reports)
-Construction of national artificial intelligence application pilot base
-Build data, model, intelligent body platform, the formation of large-scale industry service capabilities.
-Released the "China Unicom Self-Intelligence Network White Paper (2025)", the system proposed to "intelligent body" as the core towards L4 high-level self-intelligence network.
-Launch the "Unicom Cloud Rhinoceros" AI assistant to realize intelligent call transcription and summary generation
-In Sichuan, Shanghai, Hubei and other places to launch Token Plan personal version and team version
4.2 Global Operator AI Agent Layout Overview
| Carrier | Country | Core Initiatives |
|---|---|---|
| AT&T | United States | Cooperate with Microsoft Azure to deploy AI operations intelligence, focusing on network failure prediction and self-healing. |
| Vodafone | Europe | Collaborate with Google Cloud to deploy AI agents at scale in customer service |
| SK Telecom | South Korea | Launches "A." AI Personal Intelligence Service, Covering Communications, Finance, Life Scenarios |
| NTT Docomo | Japan | Development of a large network operation and maintenance model for base station fault diagnosis and parameter optimization |
| Singtel | Singapore | Launches enterprise-oriented AI intelligence service platform RE:AI |
4.3 Looking at Operators' AI Business Paradigm Innovation from "Token Package"
According to public reports, in May 2026, the three major operators will simultaneously launch the AI Token package, which will clearly mark the price of AI computing power like a "traffic package". This is a milestone event in the commercialization of operators' AI:
- China Telecom : Launches three packages for individual and family customers, with a minimum of 9.9 yuan/10 million Tokens; The maximum 0.15 billion of enterprise package is Tokens/39.9 yuan/month
- China Mobile : Launches Universal Token Service, "No.1 Universal, Cross-platform Use, Telephone Fee Payment"; Cooperate with Tencent to Launch AI Native Workspace
- China Unicom : Launches 6 million/12 million/18 million Tokens with prices ranging from 7.5 yuan/month to 359 yuan/month, integrating with mobile phone communication and gigabit broadband
underlying logic:Token is becoming the "traffic unit" of the AI era ". Operators package AI computing power as standardized goods, taking advantage of their user reach and billing capabilities to sink AI services to the billion-user market. The essence of this model is the transition of operators from "network pipeline providers" to "AI service providers. At a deeper level, the innovative value of the Token package lies in the fact that it expands AI consumption from the technology circle to the national market, enabling ordinary users to consume AI power like using traffic, which provides a key infrastructure for the large-scale popularization of the entire AI industry.
Deep analysis of 5. core application scenarios
5.1 Scenario 1: Network Operation and Maintenance Automation-From Passive Repair to Active Self-healing
5.1.1 Current Situation and Pain Points
The core challenges faced by the traditional network operation and maintenance model:
-Alarm storm:5G network generates massive alarm data every day, and operation and maintenance personnel are submerged in the alarm ocean
-Slow fault location: The average fault location time for cross-domain faults can reach several hours, depending on expert experience
-Response lag: From user complaints to fault repair, the traditional mode is measured in "hours"
-Talent bottleneck: senior network engineers are scarce and experience is difficult to replicate on a large scale
5.1.2 AI Agent Solution
Intelligent fault diagnosis and self-healing:
According to public reports, the intelligent complaint analysis agent jointly launched by China Mobile Jiangsu and ZTE is a typical example:
-LLM real-time transcription of user voice complaints, NLP semantic understanding automatic classification of root causes
-Accurately map complaint work orders to specific network elements
-Average fault identification time reduced from 12 minutes to less than 45 seconds
-Network availability increased from 99.5 percent to 99.9 percent
-20% reduction in operation and maintenance labor costs
Intelligent alarm compression and correlation analysis:
AI intelligence can intelligently compress massive network alarms:
-Converting thousands of alarms into several root cause alarms based on spatio-temporal correlation analysis
-Automatically generate fault handling suggestions to guide front-line engineers to operate
-The filtering rate of false alarms and repeated alarms can reach more than 90%
Predictive maintenance:
Predictive models based on historical network data and AI time series:
-Predict the failure probability of base station equipment in advance
-Active scheduling and maintenance resources to avoid passive repair
-Reduce unplanned downtime by 30%-50%
5.1.3 China Telecom 900 Agent Practice
China Telecom's practice is currently the world's largest operator AI intelligence deployment case:
-Deployment scale:31 provinces (autonomous regions and municipalities), over 900 AI agents
-Coverage Scenarios: Cloud network security, operational efficiency improvement, talent team transformation, business process optimization
-Technical breakthrough: the first multi-modal domain data set construction method, to overcome the core technology of large model training and push integration; to achieve the "thinking chain" ability in the communication of complex scenes of the depth of internalization; build TB-level ultra-large-scale domain knowledge base.
- Actual Effect : The fault handling efficiency has been improved by magnitude, and more than 40000 operation and maintenance personnel have completed skill transformation and upgrading.
- Industry Recognition : According to public reports, 2025 and 2026 have won TM Forum's "Self-Intelligence Network Excellence Award" for two consecutive years"
5.2 Scenario 2: Intelligent Customer Service-From Call Center to Cognitive Services
5.2.1 Application Form
The application of AI intelligence in the field of operator customer service has evolved from simple FAQ questions to omni-channel, multi-modal cognitive services:
Intelligent customer service robot:
-Multi-round dialogue, intention recognition, sentiment analysis based on LLM
-7 x 24 hours uninterrupted service, second response
-Shanghai Telecom is based on the rising self-research system, and the accuracy of large-model customer service has been increased from 80% to 90%.
Call Intelligence Enhancements:
-Huawei AI quiet calls: integrated AI noise reduction model on the network side, reducing noise from 80 decibels to 40 decibels, and improving human voice clarity by 80%
-China Unicom's "cloud rhinoceros" AI assistant: automatically converts call content into structured minutes and to-do items
Intelligent Complaint Handling:
-Automatic transcription, classification and routing of voice complaints
-Prejudge problems based on user portraits and historical data
-Automatic generation of processing suggestions and speech recommendations
Proactive Care and Marketing:
-Predict potential user needs based on user behavior and network usage data
-AI intelligence actively initiates personalized service recommendations (renewal reminders, package optimization recommendations, etc.)
-Upgrade customer service from "Passive Response" to "Active Care"
5.3 Scene 3: AI computing service -- Token's new track for commercialization
5.3.1 Business Model Innovation
The launch of the operator's AI Token service marks the entry of AI computing power consumption into the "popular" era:
Mode 1: Standard Token Package
-Package AI computing power as standardized products, similar to mobile phone traffic packages
-Support multi-model access (100 mainstream models such as DeepSeek, QWen and GLM)
-Billing by token, lowering the user's AI usage threshold
Model 2: Integrated AI Service Portal
-China Telecom "Zhiyun Shanghai AI STORE": Convergence Computing Supermarket, Model Supermarket, Application Supermarket and Skill Square
-Users only need to enter business scenarios to get personalized solution recommendations
-With the off-line empowerment of engineers, escort the whole process from consultation to landing.
Mode 3: Smart Body as a Service (Agent-as-a-Service, AaaS)
-Preset in-depth research, contract review, graphic processing and other office skills
-Supports enterprise-level out-of-the-box and custom capability precipitation
-Build industry vertical intelligence (government, medical, education, industry, etc.)
5.3.2 Market Prospects
-As of March 2026, China's average daily Token consumption exceeded 140 trillion, up more than 1000 times from early 2024.
-China Mobile smart computing service revenue growth rate of up to 279 (according to public financial reports)
-China Unicom's AI revenue grew by more than 140 percent year-on-year (according to public financial reports)
-Operator AI service market is expected to exceed 100 billion in 2027
5.4 Scenario 4: Network Planning and Optimization Intelligence
Intent-Based Networking:
-Humans only need to express business intentions (such as "guarantee VIP user video experience rate> 10Mbps")
-AI intelligences automatically translate intent into network configuration policies.
-Intent management intelligence is responsible for closed-loop monitoring and strategy adjustment.
-Industry standards organizations have incorporated intent management functionality into the core architectural framework of self-intelligent networks
Dynamic network slicing orchestration:
-AI intelligence to analyze business needs and usage trends.
-Automatically create, tune and optimize network slices
-Ensure quality of service for different services (eMBB/uRLLC/mMTC)
Wireless network parameter self-optimization:
-Coverage and Capacity Optimization (CCO)
-Mobility Load Balancing (MLB)
-Energy saving strategy intelligent scheduling
-AI the original wireless air interface high-precision positioning (China Telecom field test accuracy of 13 meters, more than 10 times higher than the traditional)
5.5 Scenario 5: AI Enabling Communication Security
Intelligent Fraud Identification and Interception:
-China Mobile AI Intelligent Fraud Identification Function: Relying on Super SIM and New Call Technology, China Mobile Initiated Pre-call Notification Reminder and Real-time Voice Early Warning During Call
-Protective call restriction technology: After analyzing the risk of fraud, let the fraud "come to an end"
-Served over 600000 users
AI Security Intelligence:
-China Telecom released the industry's first AI intelligence security governance white paper.
-Launch of carrier-grade smart body security solutions.
-Build a three-tier protection system covering "data security-model security-application security".
Quantum AI Security Fusion:
-China Telecom "Tianyan" quantum computing cloud platform, the world's first open "quantum superiority" cloud services
Quantum security infrastructure has been deployed in 40 key cities across the country, serving more than 5000 important customers.
5.6 Scenario 6: Smart Family and Industry Empowerment
Smart Home AI Intelligence:
-China Telecom Tianyi Smart Screen: integrates home control, call and monitoring, and supports multi-dialect recognition.
-Voice "voice-to-eye" service for visually impaired people
-Suitable for aging transformation: AI monitoring of the elderly heart rate, activity trajectory, home safety
Vertical industry intelligences:
-Low-altitude economy: China Telecom's plan has landed in more than 160 cities.
-Smart Health Care: China Mobile's "Family Bed" Elderly Living Alone "Worry-free Living" Service
-Industrial Internet: China Unicom serves nearly 400000 corporate customers
##6. business opportunity analysis: from cost reduction and efficiency enhancement to industrial restructuring
Three-tier model of 6.1 value creation
The value created by AI agents for operators can be summarized in three levels:
┌─────────────────────────────┐
│ 第三层：产业生态重构 │
│ 平台化运营 · 能力开放 · │
│ 生态赋能 · 新商业模式 │
├─────────────────────────────┤
│ 第二层：收入增长引擎 │
│ AI Token服务 · 行业解决方案 │
│ · 智能体平台 · 数据变现 │
├─────────────────────────────┤
│ 第一层：降本增效基石 │
│ 运维自动化 · 智能客服 · │
│ 能耗优化 · 人力转型 │
└─────────────────────────────┘6.2 Layer 1: Commercial Quantification of Cost Reduction and Efficiency Increase
| Scenario | Cost reduction and efficiency estimation | Calculation basis |
|---|---|---|
| Network Operation and Maintenance Automation | Operation and Maintenance Labor Cost Reduced by 15%-30% | China Mobile Agent Practice: Labor Cost Reduced by 20%, Fault Handling Time Reduced by 90% + |
| Intelligent Customer Service | Customer Service Cost Reduced by 25%-40% | Intelligent Customer Service Can Handle 70%-85% of Routine Consultation, Diverting Labor Seat Pressure |
| Intelligent Optimization of Energy Consumption | Base Station Energy Consumption Reduced by 10%-20% | AI Dynamic Adjustment of Base Station Sleep/Wake Strategy to Optimize Transmission Power |
| Predictive maintenance | 30%-50% reduction in unplanned downtime | Predict equipment failures in advance and actively schedule maintenance |
Based on the annual operating cost of 10 billion yuan for medium and large operators, the full deployment of AI intelligence can bring 15-3 billion yuan in cost savings.
6.3 Tier 2: Revenue Growth Engine
AI Token and computing power service
Here's the operator's most imaginative new revenue stream:
-China Mobile's smart computing service revenue is nearly 90 billion yuan (according to public financial reports), a growth rate of 279
-China Unicom's AI revenue grew by more than 140 percent year-on-year (according to public financial reports)
-Token package has opened up hundreds of millions of existing users, realizing AI consumption from "developer exclusive" to "universal benefit"
-It is estimated that the AI service revenue of operators will exceed 100 billion in 2027
Industry Solutions
-Digital Revenue of China Telecom Industry 147.3 billion Yuan (According to Public Financial Report)
-China Unicom's strategic emerging industries account for more than 80% of revenue (according to public financial reports)
-Low-altitude economy, smart city, industrial Internet and other scenarios continue to expand
Intelligent body platform service
-Preset the knowledge market for industry Skill
-Enterprise smart body custom development.
-Multi-intelligence collaborative orchestration service.
-Smart body effect evaluation and optimization services.
The third layer of 6.4: industrial ecological reconstruction
AI agents will fundamentally reshape the industrial role of operators:
From" pipeline vendor "to" platform vendor ":
Operators are no longer just providing connectivity, but becoming platforms for the aggregation and distribution of AI capabilities-aggregating computing power, models, applications, data, and exporting intelligent services to downstream businesses and consumers.
From" Closed Operation "to" Open Ecology ":
-Open network capabilities to developers through MCP (Model Context Protocol)
-Build a smart body app store (Agent Store)
-Joint ISV to build a vertical industry intelligent body ecology.
From Cost Center to Profit Center:
Under the support of AI intelligence, the traditional operation and maintenance department can be transformed from a cost center to a profit center that provides intelligent operation and maintenance services to the outside world.
6.5 Market Space Forecast
| Market Segments | Scale in 2025 | Forecast Scale in 2030 | CAGR |
|---|---|---|---|
| Global Self-Intelligence Network Market | ~ $12 billion | ~ $28 billion | ~ 18.5% |
| Operator AI Service Market (China) | ~ 50 billion yuan | ~ 250 billion yuan | ~ 38% |
| Telecom AI Agent Platform Market | ~ 3 billion yuan | ~ 30 billion yuan | ~ 58% |
| Operator AI Token Consumption | ~ 5 billion yuan | ~ 80 billion yuan | ~ 74% |
##7. technical architecture and implementation path
7.1 Reference Architecture of AI Agent in Telecommunication Network
Based on industry analysis and public information, the deployment architecture of AI intelligences in carrier networks can be understood from the following three planes:
Management plane smartlayer:
-Intent to receive and resolve intelligences: receive business intent expressions, translated into executable network policies.
-Service management and orchestration intelligences: responsible for task decomposition, resource allocation, and process orchestration.
-Cross-domain collaborative intelligence: to achieve coordination and linkage between different areas such as wireless domain, core domain, bearer domain, etc.
-Network Planning and Design Intelligence: Assist in network capacity planning, topology optimization, and new site design.
Control plane smartlayer:
-Policy Control Intelligence: Dynamically adjust QoS policies, access policies, and resource allocation policies.
-Network Data Analysis Intelligence: Aggregate and analyze network-wide data to provide decision-making basis for upper intelligence.
-Access and Mobility Management Intelligence: Optimize user access selection and switching strategies.
Application smart layer:
-User-oriented AI service intelligence: AI capabilities and applications that directly serve end users.
-Industry Solution Agent for Enterprise: Customized Intelligent Services for Vertical Industries
-Capability Open API Gateway Smart Body: Open network capabilities to third parties with standardized interfaces.
The core design idea of this hierarchical architecture is "separation of concerns": the management side focuses on network-wide decision-making and orchestration, the control side focuses on real-time policy execution, and the application side focuses on final business delivery. The three planes interoperate through standardized interfaces, forming a complete closed loop from business intent to network execution to service delivery.
Telecom Applications 7.2 Model Context Protocol (MCP)
The core values of MCP in telecommunications are:
- The standardization capability is open: the network API is encapsulated as an MCP tool for AI intelligence to call.
- Smart Body Friendly Layer: Add a conversational interaction layer to the traditional API, lowering the threshold for use
- Multi-tool orchestration : An MCP tool can aggregate multiple underlying API calls to realize advanced tasks
In terms of industry trends, MCP is becoming the de facto standard for smart bodies to interact with external tools. For operators, actively embracing the MCP standard means that on the one hand, they can export their network capabilities and data assets in a standardized way to build a developer ecosystem, and on the other hand, they can easily access third-party tools and services to enrich the capabilities of their own intelligences. This "two-way opening" capability is the key technical lever for operators to move from a closed system to an open platform.
7.3 implementation path proposal: five-stage evolution route
阶段一 阶段二 阶段三 阶段四 阶段五
(2025-2026) (2026-2027) (2027-2028) (2028-2029) (2029-2030+)
┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐
│ 单点 │ → │ 场景 │ → │ 域级 │ → │ 跨域 │ → │ 全自智 │
│ 智能体 │ │ 智能体 │ │ 自智 │ │ 协同 │ │ 网络 │
└────────┘ └────────┘ └────────┘ └────────┘ └────────┘
Copilot辅助 单域闭环 域内自治 跨域协同 零接触运营
人工决策 人机协同 机器自主 智能涌现 L5级自智Phase I (2025-2026): Single Point Agent Pilot
-Deploy Copilot intelligences in scenarios such as customer service, complaint handling, alarm analysis, etc.
-Verify the applicability and accuracy of LLM in telecom scenarios
-Accumulate domain knowledge and training data
Phase II (2026-2027): Scenario-level intelligence scale
-Extend the intelligence from pilot to full scenario (operations, customer service, optimization, security)
-Build domain knowledge bases and toolsets
-Establish an intelligent body development and management platform.
- Phase III (2027-2028): Domain-level Self-intelligence Network *
-Achieve closed-loop autonomy within a single domain (wireless domain, core domain, transmission domain)
-Intent management intelligence takes over the domain intent management and policy closed loop.
-Reach L4 self-intelligent network level
- Phase IV (2028-2029): Cross-Domain Collaborative Intelligence *
-Multi-domain intelligence collaborative orchestration.
-End-to-end business intent-driven network autonomy
-Initial formation of AI native network architecture
- Phase 5 (2029-2030 ): Fully Self-Intelligent Network *
-Zero-touch operation for 6G
-The intelligent body ecosystem is mature.
-Level L5 fully autonomous network
7.4 Critical Success Factors
- High-quality domain data: Build high-quality training data sets and knowledge bases in the field of communications.
- Robust evaluation system: Establishing criteria for evaluating the accuracy, safety and consistency of AI intelligences
- Human-Machine Cooperation Mechanism : Reasonable Design of "Human-in-the-Loop"(Human-in-the-Loop) Mechanism
- Organization and talent transformation: promote the transformation of operation and maintenance personnel to new roles such as AI trainer and intelligent body manager.
- Open Platform Ecology: Build an open ecosystem of tools and applications through standard protocols such as MCP
##8. Challenges and Risk Governance
8.1 Technical Challenges
Illusion and the Credibility Problem:
LLM's "illusion" (generation of inaccurate or fictitious content) can have serious consequences in telecommunications scenarios. A single incorrect network configuration recommendation can result in large service outages. Solving this problem requires a multi-pronged approach: RAG enhances domain knowledge accuracy, outputs a multi-layer verification mechanism, and high-risk operations force manual intervention.
Observability and interpretability:
The reasoning process of AI agents is often "black box. In critical network operations, it must be possible to trace and understand the decision logic of the agent. The industry is exploring techniques such as thinking chain visualization and decision path audit tracking to enhance the interpretability of intelligences.
Real-Time and Reliability:
Network operation and maintenance requires extremely high real-time performance (milliseconds), and the current LLM reasoning delay is still the bottleneck. Technical means such as model distillation and edge deployment are needed to reduce reasoning delay and ensure decision quality.
System complexity:
The complexity of the AI agent system itself (multi-model, multi-tool, multi-agent collaboration) brings new management and debugging challenges. When multiple agents work together, the abnormal behavior in the interaction may produce a cascade effect, which requires a well-designed abnormal isolation and fusing mechanism.
8.2 Security and Privacy Risks
-Data security: AI intelligences need to access large amounts of network data and user data, and the risk of data leakage cannot be ignored.
-Fighting attacks: Malicious input can induce AI intelligences to make wrong decisions.
-Out of control: Highly autonomous agents may perform dangerous operations without authorization
-Supply Chain Risk: Third-party models and tools may introduce backdoors or vulnerabilities.
8.3 governance recommendations
China Telecom's White Paper on AI Intelligence Security Governance proposes a useful governance framework that, combined with industry practice, recommends that operators:
- Establish a AI intelligent body hierarchical management system: according to the degree of autonomy and scope of influence of the intelligent body, divide the security level (L1-L5), implement differentiated control strategies.
- Implement the" people in the loop "mechanism: for high-risk operations, retain the manual approval process
- Deployment of architecture-level guard bars: including tool constraints, code sandboxes, thought chain supervision, output verification, etc.
- Establish full life cycle security assessment : cover the whole process of data collection, model training, deployment and online, operation monitoring
- Promote the formulation of industry standards : Actively participate in the formulation of AI safety standards by international standards organizations such as 3GPP and ITU, and promote the formation of a security framework of industry consensus
##9. Future Outlook: Towards L5 All-Self-Intelligence Network
9.1 6G Deep Fusion with AI Native
The 6G network will be designed with AI native (AI-Native) as the core concept. Unlike 5G's "AI-enabled network", 6G's "AI-native" means that AI is no longer a layer of functionality attached to the network, but rather a genetic component of the network:
-AI native air interface: AI algorithms deeply involved in physical layer signal processing
-AI native protocol stack: network protocol can be dynamically adjusted according to the scene
-AI native management plane: Intent-driven intelligent control becomes the core mechanism of management plane
-AI native service surface: Network native support for communication between AI intelligences.
9.2 Self-intelligence Network L5 Vision
Industry standards organizations have defined a five-level maturity model for self-intelligent networks, a framework that is widely accepted among operators worldwide:
| Level | Definition | AI Agent Role | Expected Time |
|---|---|---|---|
| L0 | Manual Operations | None | Implemalized |
| L1 | Auxiliary Operations | Basic Automation Script | Implemtized |
| L2 | Partial Self-Intelligence | Single Scenario AI Assist | Implementally |
| L3 | Conditional Self-Intelligence | Single-Domain Copilot Intelligence | 2025-2026 |
| L4 | Highly self-intelligent | Domain-level Agent closed-loop autonomy | 2027-2029 |
| L5 | Fully Self-intelligent | Network-wide Multi-Agent Collaboration, Zero Contact | 2030 |
At present, China's head operators have reached the L3 level and are accelerating towards L4. China Unicom's White Paper on Self-Intelligence Network (2025) clearly proposes to move towards L4 high-level self-intelligence with "intelligent body" as the core, and China Telecom's 900 intelligent body deployment has shown some of the characteristics of L4.
9.3 Future Business Model Imagination
Smart Body as a Service (AaaS) Store:
Operators will build a smart body application market similar to App Store-developers upload industry smart bodies, enterprise users subscribe to them on demand, and operators will provide basic support for computing power, network and billing.
Multi-intelligent collaboration network:
As the inter-agent communication protocol (A2A) matures, the carrier network will become the infrastructure that supports the collaborative operation of billions of agents-similar to today's network connecting billions of mobile phone users, the network of the future will connect tens of billions of AI agents.
Intent Economy:
When the network realizes full self-intelligence, the core business interface of the operator will change from "purchasing bandwidth/traffic" to "expressing business intention"-the customer only needs to say "I want to ensure XX service experience in XX city", and the network will automatically complete resource arrangement and quality assurance.
##10. conclusions and recommendations
10.1 Core Conclusions
- AI intelligence is the core enabling technology for operators' digital transformation, which will unlock significant value in areas such as network operations, customer service, business innovation, and security management.
- China's three major operators have been at the forefront of global AI agent applications, China Telecom 900 agent deployment as the world's largest practice case, according to public reports for two consecutive years by TM Forum recognition.
- The commercialization of AI Token has opened a new growth curve for operators, and the business model transition from "selling connections" to "selling computing power/selling models/selling intelligences" is taking place.
- The evolution path of self-intelligence network is clear , currently in the critical period of L3 → L4 upgrade, and L5 full self-intelligence network will be ushered in around 2030
- Security governance is the premise of large-scale promotion, the need to establish a AI intelligent body classification management system and perfect security protection mechanism.
10.2 Recommendations for Operator Decision Makers
Strategic level:
-The AI intelligence will be upgraded to an enterprise-level strategy, with a clear short-to medium-term evolution roadmap.
-Build a complete AI capability stack of "computing power-model-platform-application".
-Promote organizational change and cultivate AI original talent team
Tactical Level:
-Give priority to landing in scenarios with clear value and controllable risks, such as operation and maintenance and customer service.
-Build domain knowledge base and high-quality data sets, which are core competencies
-Adopt the progressive path of "pilot-assessment-optimization-promotion"
-Actively participate in 3GPP, ITU and other international standards organizations, master the right to speak in the industry
Ecological level:
-Create an open smart body platform to attract ISVs and developers to build an ecosystem.
-Open network capabilities through standard protocols such as MCP
-Establish deep cooperation relationship with cloud vendors, model vendors and industry ISV
10.3 Recommendations for Industry Partners
-Equipment vendors: Embed AI intelligence capabilities into network equipment to achieve network-level intelligence.
-Software developer: Develop vertical intelligence applications for operator scenarios.
-System Integrator: Provides integration and customization services for AI intelligence platforms
-Security vendor: Build AI intelligence security assessment and protection solutions
References 1. Ericsson, "Application of AI agents in telecommunications network architecture" (public white paper), 2026. 2. China Unicom, "China Unicom Self-intelligence Network White Paper (2025)" 3. China Telecom, "AI Intelligence Security Governance White Paper", 2026. 4. Three major operators 2025 public performance report 5. people's post and telecommunications news, "AI computing power becomes the main engine of growth for the three major operators", 2026.03 6. people's post and telecommunications, "5 key words, perspective on the 2026 work meeting of 5 central communication enterprises", 2026.01 7. China Daily, "Telecom giants launch AI token services", 2026.05 8. The Next Gen Tech Insider, "China Mobile & ZTE Launch AI Complaint Analyzer", 2026.06 9. Solnix Media, "Telecom AI Agents: Automate Network Ops, Support & Churn Prediction", 2026.02