AI in telecommunications is used to run networks, serve customers and protect revenue: predicting faults, optimizing radio and transport networks, automating customer care, detecting fraud and planning capacity. In 2026 the big shift is agentic AI, where AI agents not only detect problems but decide and act within limits, moving operators toward Level 4 autonomous networks while humans set the goals.
Updated October 2026
What changed in 2026
- Autonomy got measured. In TM Forum's latest update, 81% of operators aim to reach Level 4 autonomous networks or higher by 2030, while most sit between Levels 2 and 3 today; NBN Co with Ericsson earned Australia's first Level 4 validation in June 2026 for one scenario (Telecoms Tech News).
- Network agents went live. Deutsche Telekom and Google Cloud built RAN Guardian, an agent on Gemini 2.0 that monitors radio network behavior, detects and classifies issues and recommends or applies fixes (Google Cloud).
- AI moved into the radio network. NVIDIA agreed to invest $1 billion in Nokia to build AI-RAN products on its ARC-Pro platform, with T-Mobile US trials expected to begin in 2026 (Nokia).
- Agent spending is rising. Deloitte expects as many as 75% of companies may invest in agentic AI in 2026 and sizes the agentic AI market at up to $45 billion by 2030 with proper orchestration (Deloitte TMT Predictions 2026).
How AI is used in telecommunications
| Domain | Use case | Traditional AI | Agentic AI in 2026 |
|---|---|---|---|
| Network operations | Fault detection and repair | Anomaly alerts to the NOC | Agent classifies the fault, finds root cause, applies an approved fix |
| Radio access network | Energy saving and optimization | Rules to sleep cells at night | Agent balances energy, traffic and quality targets continuously |
| Planning | Capacity and site planning | Traffic forecasts | Agent proposes upgrades and simulates impact on a network twin |
| Customer care | Support and troubleshooting | Scripted chatbots | Agent checks the line, runs diagnostics, books a technician |
| Field service | Dispatch and repair | Manual scheduling | Agent routes technicians and prepares parts and instructions |
| Fraud and security | SIM swap, IRSF, signaling attacks | Batch fraud models | Real-time scoring with automated blocking and case creation |
| Commercial | Offers, churn and billing | Churn scores in campaigns | Agent personalizes offers and resolves billing disputes within policy |
What makes AI "agentic" in telecom
A traditional model answers one question, such as "is this cell degrading?" An AI agent pursues a goal, such as "keep dropped calls in this cluster under target at the lowest energy cost." To do that it observes data, reasons with a model, calls tools (inventory, configuration, ticketing) and acts, then checks the result.
TM Forum's autonomy scale describes the same journey. At Level 4 the system handles awareness, analysis, decision and execution for defined scenarios, while humans still share the intent: people set the objective, and the system works out how to meet it.
A reference architecture for telecom AI agents
- Data layer: network counters, alarms, topology, inventory, customer and billing data, streamed in near real time. See real-time AI inference.
- Models: forecasting and anomaly models, graph models of the network topology (see GNN applications) and LLMs for reasoning and language.
- Agents: specialized agents for RAN, transport, core, care and fraud, coordinated by an orchestrator.
- Tools and actions: controlled APIs into OSS and BSS, SON, ticketing and field service.
- Guardrails: policies on what each agent may change, approval steps, rollback plans and full audit logs.
- Edge placement: latency-sensitive inference runs at edge sites or in the RAN itself; see edge computing applications.
Challenges operators face
- Data silos: network, IT and customer data sit in separate systems with different owners.
- Trust: engineers need to see why an agent proposes a change before they let it act.
- Safety: a wrong configuration change can affect thousands of customers, so blast-radius limits matter.
- Sovereignty and regulation: customer and network data often must stay in-country.
- Skills: teams need people who understand both networks and AI.
The World Economic Forum's report on AI in telecommunications and IBM's telecommunications in the AI era study are useful wider reading.
Where to start: three first agent use cases
- RAN energy saving. Clear goal, measurable result and easy rollback make it a safe first scenario for closed-loop automation.
- Alarm triage in the NOC. An agent groups related alarms, suggests root cause and drafts the ticket, while engineers keep the final say.
- Care-to-network handoff. When a customer reports a problem, an agent checks their line and local cell health before routing the case, aiming to reduce repeat contacts.
Start in "recommend" mode, measure accuracy against engineer decisions for a few weeks, then allow the agent to act on low-risk changes. Expand only when the audit trail shows consistent, explainable results.
Where it's heading: 2026 → 2030
| When | What it looks like |
|---|---|
| Around 2020 | SON features and ML anomaly detection in the NOC; scripted chatbots; network automation at early autonomy levels. |
| 2026 (now) | Network agents such as RAN Guardian; first Level 4 validations for single scenarios; AI-RAN trials; generative AI in customer care. |
| By 2028 (expected) | Our view: Level 4 becomes common for selected scenarios such as energy saving and service assurance; multi-agent systems coordinate across RAN, transport and care. |
| By 2030 (expected) | Our view: AI-native, 6G-ready networks run AI in the RAN by design, and intent-based operations are standard, with humans setting goals and policies. |
How NexaStack helps
NexaStack gives operators a platform to deploy and govern AI models and agents across central clouds, regional data centers and edge sites, keeping network and customer data inside their own boundary. Teams can set what each agent may change, roll out new model versions safely and audit every action. ElixirClaw, our AI agents computing platform, and ElixirData, the decision harness for enterprise AI agents, add agent runtime and decision governance where needed. See our Edge AI solution.
Frequently asked questions
How is AI used in telecommunications?
For network fault detection and self-healing, radio and energy optimization, capacity planning, customer care, field service, fraud detection and personalized offers.
What is agentic AI in telecom?
AI agents that pursue goals, such as network quality or customer resolution, by observing data, reasoning, calling tools and acting within set policies, rather than only producing alerts.
What are Level 4 autonomous networks?
In TM Forum's scale, Level 4 means the system handles awareness, analysis, decision and execution for defined scenarios, while humans set the intent.
Is AT&T using AI technology?
Yes. Like most large operators, AT&T uses AI in network operations and customer service. Check the operator's own announcements for current specifics.
What are the latest AI in telecom trends in 2026?
Agentic network operations, Level 4 autonomy validations, AI-RAN trials, generative AI customer care, real-time fraud prevention and sovereign AI infrastructure for operators.
Put telecom AI agents to work safely
Book a NexaStack demo to deploy governed AI agents across your network and edge →
