1. Introduction

In late 2025, TM Forum presented 37 independent validations of progress toward Level 4 autonomous networks, and Ericsson — working with TDC NET in Denmark — took the first validated Level 4 result under the new assessment-validation scheme. The scenario was narrow and concrete: a Predictive Cell Energy Management solution that has run in production since May 2024, automating radio-access energy-efficiency optimisation by reading per-site traffic profiles and trading energy savings against measured customer experience. It earned full marks on the assessment. It also covers exactly one operation flow in one network domain.

Read past the certifications and the wider picture is sober. A TM Forum regional benchmark covering 2025–2026 placed only about 4% of operators at Level 4 overall, 17% at Level 3 and 31% at Level 2. A separate Accenture study found 79% of telcos still at Level 0 or Level 1, with only 22% expecting to reach Level 4 by 2030. Both readings hold at once because they measure different objects: a validated Level 4 result is a single scenario, while the maturity surveys score an operator’s network as a whole.

That gap — between certified islands of autonomy and end-to-end self-management — is what this article is about. TM Forum frames the move to Level 4 as a shift from “systems assist humans” to “humans assist machines,” and the barrier is rarely the algorithm. It is data quality, trust and assurance, multi-vendor brownfield integration, and the reorganisation of the people who run the network. What follows is the level model and the cognitive handover that defines it, the layered closed-loop architecture and the data-centric core beneath it, how autonomy is scored and why a score stalls one step short, the concrete barriers, and what the leaders actually did. If you want the level taxonomy on its own first, the companion guide to the six levels of autonomous networks, L0 to L5, walks each level in detail.

2. Fundamentals: the level model and what Level 4 demands

TM Forum’s Autonomous Networks Project, formed in 2019, defines a six-level taxonomy from Level 0 (manual operation) to Level 5 (full autonomy), modelled loosely on driving-automation levels but scored per scenario rather than per network. A scenario is one operation flow on one network domain — fault management on the radio access network, quality optimisation on the IP backbone, energy management on a base-station fleet. Each level is set by which actor performs five cognitive activities: Intent and experience, Awareness, Analysis, Decision and Execution, abbreviated IAADE. The level rises as each activity moves from people (P) to system (S).

The handover is the whole point. At Level 2 the system executes repetitive operations inside statically configured rules. At Level 3 it owns awareness and analysis and proposes an action, but a human still approves the decision. At Level 4 the system owns awareness, analysis, decision and execution for the chosen scenarios, and the human retains only intent and exception handling. At Level 5 the system owns intent as well, across every scenario and domain. The table below shows the canonical handover.

The difference between Level 3 and Level 4 is the difference between automation and autonomy. Automation follows rules a person wrote; autonomy infers. At Level 4 the system generates the rules — predictive models read the situation and generative models translate intent and synthesise configuration. That is also where the new failure mode lives: a model that drifts on stale data, or hallucinates a configuration, is acting without the human checkpoint that Level 3 deliberately keeps. The capability and the risk are the same mechanism.

Level 4 runs on intent, and intent is the “what,” not the “how.” It expresses a goal and its constraints, abstracted from the network’s internal workings, and is carried into the architecture by the TM Forum Intent Management API (TMF921). The decoupling is what lets the system choose resources and innovate behind a stable boundary — the same idea explored in intent-based networking for optical systems. The boundary is also a hazard at Level 4: an ambiguous or malformed intent has no operator to catch it before the loop acts.

Above this sits the vision TM Forum calls Zero-X and Self-X — zero-touch, zero-wait, zero-trouble services delivered by self-configuring, self-healing, self-optimising operations. The structural unit that makes it tractable is the autonomous domain: a slice of the estate — access, metro, core, edge, or a service such as SD-WAN or VoLTE — that runs its own closed loop behind an intent-based interface and collaborates with its neighbours.

No single body owns this. TM Forum writes the domain-agnostic layer — the reference architecture (IG1251), the levels evaluation methodology (IG1252), intent management (IG1253 with the TMF921 API), and the effectiveness indicators (IG1256). 3GPP writes the technology-specific layer for mobile, including the levels of autonomous network (TS 28.100) and intent-driven management services (TS 28.312). ETSI’s Zero-touch network and Service Management group defines the closed-loop framework (the GS ZSM 002 reference architecture and the GS ZSM 009-1 closed-loop enablers), and ETSI F5G extends autonomy-level work to the transport layer. ITU-T Study Group 13 publishes an autonomous-networks architecture framework (Y.3061), and CCSA aligns its definitions to the same level scale. The split across bodies is itself a challenge: a network-wide “level” is only ever a weighted blend of many separately scored scenarios.

3. The Level 4 architecture: layered closed loops over a data-centric core

The TM Forum target architecture stacks three operations layers — business, service and resource — each running its own closed loop, with a cross-layer loop binding them end to end. Intent flows down (business intent becomes service intent becomes resource intent) and observed state flows back up. AI runs at every layer rather than bolted on top, an arrangement TM Forum calls full-stack AI.

Each layer applies AI to its own work. The business layer translates customer intent and shapes offers and experience; the service layer translates service intent, orchestrates resources end to end, creates SLA policy and runs fault and performance management; the resource layer does local inference, perception and control down at the element-management systems and network elements. Generative AI handles configuration synthesis and operator question-answering while traditional machine learning handles pattern detection and prediction — the two run together.

The loop itself is the IAADE engine: awareness collects and correlates multi-layer state, analysis finds root cause and predicts impact, decision chooses an action (at Level 4 the AI generates the rule), and execution applies it. The next diagram traces that loop and the Level 4 handover. Execution is mechanically an API call from a controller to a network element — over TL1, NETCONF, RESTCONF or gNMI — and the multi-vendor model that carries it is increasingly OpenConfig, the subject of the guide to northbound and southbound interface protocols.

Underneath the layers sits the TM Forum Open Digital Architecture, the cloud-native core that makes self-management buildable. ODA rests on five principles: it is AI-capable and autonomous, data-centric rather than process-centric, microservice-based over the TM Forum Open API suite, real-time at every layer, and built for platform and cloud-native operation. Each component belongs to one function block, is meant to be intent-based, and is secure by design. Operators already on ODA can recast it into composable, self-managing autonomous domains rather than starting over.

The plane that feeds every loop is data and knowledge: streaming telemetry, a digital twin, and increasingly a telecom foundation model. The twin matters because it gives the loop somewhere safe to test a decision before it touches production. In one published case, China Telecom ran a high-precision IP-network twin that lifted the rate of preventing change-induced risk by more than 80% in pilots, with a projected avoidance of roughly USD 140 million a year if applied nationally. The boundary is plain: a twin is only as accurate as the data feeding it.

That dependence is the first hard barrier. A closed loop acts on telemetry, and brownfield estates run 2G through 5G in vertical silos with vendor-locked element managers, inconsistent schemas and inventory that is frequently wrong. Without a common data spine, automation stays trapped inside each silo and never reaches the cross-domain orchestration that on-demand services need — the central argument for migrating legacy OSS and BSS toward a data-centric core. The first project on most Level 4 roadmaps is therefore not an AI model; it is a unified data layer.

The second barrier is multi-vendor integration. Open APIs were meant to give zero-touch integration, but adoption is uneven, and the brownfield carries proprietary management for years. The southbound is converging on OpenConfig and OpenROADM over NETCONF and gNMI, yet vendors implement the same models differently, so each integration still carries cost. Analysys Mason notes that network-automation software revenue in the radio domain dipped in 2024 precisely because of integration complexity in disaggregated, multi-vendor environments — the topic of the network-as-code control loop for keeping that integration testable.

The third barrier is the one the framework itself flags: at Level 4 several closed loops run at once, and the system “may not know how to resolve conflicts between multiple automatic closed-loop objectives” — an energy loop and an experience loop pulling opposite ways on the same cluster. ETSI ZSM answers with hierarchy: a superior loop sets non-conflicting goals for subordinate loops by delegating policy and intent, and a subordinate escalates when it cannot meet a goal it was given. That is the mechanism, and it breaks where the objectives cannot be expressed as a clean hierarchy of goals.

This is why Level 4 is being pursued scenario by scenario. TM Forum has published a set of high-value-scenario solution packs — RAN fault management, core fault management including network stability, IP-network optimisation, IP-network fault management, change management, network planning and energy-efficiency optimisation — that define where the industry attacks first, behind interfaces such as TMF921 for intent, ZSM for closed loops, and OpenConfig or OpenROADM with gNMI streaming for the resource layer.

4. Measuring autonomy and where the score stalls

The level is not declared; it is evaluated. The first step is to fix an evaluation object, a point in three dimensions: a services domain (for example mobile B2B), a network domain (RAN, core, IP, optical transport) and an operation flow (planning, deployment, provisioning, maintenance, optimisation, inventory). For that object, TM Forum’s methodology runs an IAADE questionnaire whose tasks are weighted and scored, with weights drawn from the professional grading in 3GPP’s autonomy-level specification. An AN Level Evaluation Tool produces the self-assessment, and an AN Level Assessment Validation service audits it independently — only evaluators who have completed the skills path may submit. More than 30 operators have run the evaluation; the late-2025 wave produced 37 validated results.

The gating term is the whole story. A scenario can have three of five dimensions sitting at system level and still certify a level lower, because the minimum across dimensions decides. The dimension that lags is almost always Decision, and the reason is almost always trust rather than capability: operators will let the system see and analyse, but they hold approval for any change that could hurt a revenue-bearing cluster.

Value is measured separately, through a three-tier index of key business, effectiveness and capability indicators feeding a business-value model. TM Forum’s Level 4 target-state definitions put concrete numbers on the goal: for fault management, an automation rate of at least 90% of sub-scenarios, a sharp cut in mean time to repair, and at most one engineer site visit; for wireless quality optimisation, an automatic-closure rate above 60%, optimisation time down about 15% and issue-handling time down about 25%, with most optimisation needing no site visit. These are targets, not measured outcomes — the bar a Level 4 scenario is expected to clear.

The measured outcomes that operators report are encouraging where the data and the scenario are clean. China Mobile, nearing Level 4 across IP backhaul and RAN fault management, reports an 80% reduction in major network faults and more than 3,200 person-years of labour saved. Thailand’s AIS, evaluated at Level 3 in complaint and fault management, reports a 10% improvement in customer experience, a 90% reduction in traffic loss in extreme events and a 20% gain in efficiency. The Ericsson and TDC NET radio-energy scenario earned full marks on its Level 4 assessment. The two charts below show where operators actually sit and how they are applying generative AI.

The design trade-off follows from per-scenario scoring. End-to-end Level 4 is not the goal everywhere; in deterministic domains such as IP and optical transport, Level 2 or Level 3 already returns strong value without the cost of pushing all the way. The right move is to apply Level 4 where effort and value are highest — fault management on the expensive radio access network — and to accept lower levels where the pain is smaller. Operators rate operational-cost reduction as very important far more often than they rate the jump from Level 3 to Level 4, which tells you where the real prize sits.

5. Deployment: who has reached Level 4, and how

The validated results to date are single-domain scenarios, not whole networks. Ericsson and TDC NET took the first independently validated Level 4 in Denmark for radio-access energy optimisation, in production since 2024. Telefónica reports its first two Level 4 milestones — one in Brazil and one in Germany — for network optimisation. Orange, which began its programme in 2023, has reached high autonomy across a spread of scenarios: RAN fault management and 5G service assurance in Romania, RAN energy management in Slovakia, an externally certified Level 4 result for IP-network optimisation through its Spanish operation, and RAN performance, capacity and provisioning work in France. China Mobile is nearing Level 4 across IP backhaul and RAN fault management. Thailand’s AIS, running a “cognitive techco” strategy, sits at Level 3 in complaint and fault management. T-Mobile used its autonomous network during a January 2026 winter storm to make roughly 30,000 antenna adjustments to keep customers connected, and frames its target as an intent-based, agent-driven platform.

The configuration pattern is consistent across these programmes: domain-scoped closed loops rather than one monolith, a digital-twin sandbox to validate changes, operation-scenario agents and copilots for analysis and recommendation, and humans on the loop for intent and exceptions. Single-domain autonomy comes first; cross-domain collaboration follows. Transport often reaches higher autonomy earlier than other domains because it is more deterministic, which is part of why a structured approach to optical network automation pays off.

On the supplier side, Analysys Mason has Huawei, Nokia and Ericsson leading network automation and orchestration on the strength of combined AI, intent-driven networking and domain-level control. Independent software vendors are gaining with modular, cloud-native stacks that integrate across vendors and domains more freely than monolithic platforms — the case for consolidating multi-vendor operations behind common interfaces. Among optical-line suppliers, Ribbon’s Muse domain orchestration drives multi-vendor optical through NETCONF/YANG with OpenROADM interworking, alongside Ciena, Cisco with Acacia coherent optics, Nokia and Adtran.

The troubleshooting checklist for why a Level 4 ambition stalls is short and repeats across operators. Inaccurate inventory or siloed data caps autonomy at Level 2. Concurrent closed loops with conflicting objectives need a goal hierarchy that may not exist. Brownfield multi-vendor silos make each integration expensive. The workforce lacks AIOps, MLOps and data-engineering skills at scale. And trust in handing over decisions for revenue-bearing scenarios lags the technical capability — which, as the gating formula shows, is exactly what keeps the Decision dimension below system level.

6. What is holding Level 4 back: the underlay first, then the rest

An autonomous network is a control overlay, and it can act only on what the network beneath it exposes. ITU-T Y.3061 makes the dependency explicit: the architecture is expected to discover the characteristics of the underlay network at runtime, and an optimisation decision frequently resolves to a configuration change pushed into that underlay. A loop that cannot observe a device at sub-second granularity, or cannot write to it through a model, has nothing to close the loop on. That is why the factor operators name first is rarely the model at the top of the stack — it is the state of the transport, IP and radio underlay at the bottom.

Brownfield estates run 2G through 5G as vertical silos, each with its own vendor-locked element manager, its own data schema, and a management plane built for human hands: SNMP polling on a multi-minute cycle, a command-line interface, or TL1. None of that feeds a closed loop. Awareness needs streaming telemetry — gNMI pushing state on change or on a sub-second sample — and execution needs model-driven write access over NETCONF or RESTCONF against a YANG model the controller understands. A device you can only poll every few minutes and reconfigure by hand caps the scenario above it at Level 2, no matter how capable the analytics are. The underlay has to become a programmable, observable substrate — the role of vendor-neutral OpenConfig and OpenROADM models — before the loop on top of it can certify past partial autonomy.

Modernising that underlay is the multi-year, capital-heavy programme that competes for the same budget as the automation itself. It means decommissioning 3G to recover spectrum and power, densifying 5G across non-standalone and standalone, migrating network functions onto telco cloud and hyperscaler infrastructure, and disaggregating transport so the line system answers to an API rather than a vendor element manager. The routed-optical, IP-over-DWDM model is part of that shift: collapsing the transponder shelf into a coherent pluggable in the router port removes a whole managed layer and its element-management system, which is as much an operations simplification as a capital one. Early adopters report meaningful capital and operations savings (a vendor and operator claim, not a measured industry average); the historical barrier was the faceplate and power penalty of line-side optics, now eased by 400G ZR and ZR+ pluggables.

Operators do not wait for the underlay to finish modernising. The working pattern is two parallel threads: the network modernises on its own multi-year schedule, while an operations-layer Level 4 programme automates high-value scenarios on whatever subset of the estate is already observable, widening the eligible scope as more of the underlay becomes model-driven. The corollary, which leading operators state plainly, is that they deliberately exclude the parts of the estate too legacy to instrument — there is no real automation case for a device that cannot expose its state — so the underlay does not stop the programme, it bounds which scenarios are eligible for it.

Underneath that headline sit the factors operators and analysts name consistently. Each one gates Level 4 through a specific mechanism, and each has a specific thing that unblocks it.

7. Future outlook

TM Forum splits the road to Level 4 into two phases. The first, running 2025 to 2027, focuses on roughly 20 high-value scenarios and applies single-domain agents for self-optimisation and self-healing. The second, 2028 to 2030, moves to multiple cross-layer and cross-domain agents that collaborate on harder, end-to-end problems. Level 5 — where the system generates its own intents and could trade customer experience against carbon on its own — remains a research target, and some operators say plainly they will stop short of it, unwilling to remove humans from assurance for revenue-bearing services.

The gating problem stays the trustworthiness of AI. The work that matters most is making intent understanding, inference, decision and execution reliable enough to cede control, and proving it — explainability, drift detection, and validation against a digital twin before a change reaches production. Telecom foundation models and agentic AI for network operations are the tools the industry is betting on, with security and trust as the explicit condition for scaling them.

For engineers, the skills that compound are intent modelling, AIOps and MLOps, data engineering, closed-loop assurance and digital-twin operation. The standards worth tracking are the maturing TM Forum evaluation and validation tooling, 3GPP’s continuing autonomy and intent work, ETSI F5G transport autonomy levels, and ITU-T’s autonomous-networks architecture framework. The honest near-term picture is incremental: most operators target Level 4 by 2030, scenario by scenario, and the ones who get there first will be the ones who fixed their data and their organisation, not just their models.

8. Reference

Standards and specifications

Glossary

References

• TM Forum, Autonomous Networks: Level 4 Industry Blueprint, TM Forum Autonomous Networks Project.
• TM Forum, Autonomous Network Levels Evaluation Methodology (IG1252), TM Forum.
• TM Forum, Intent Management (IG1253) and Intent Management API (TMF921), TM Forum.
• 3GPP, Levels of Autonomous Network (TS 28.100), 3rd Generation Partnership Project.
• ETSI, Zero-touch network and Service Management; Reference Architecture (GS ZSM 002) and Closed-Loop Automation Enablers (GS ZSM 009-1), ETSI Industry Specification Group ZSM.
• ITU-T, Autonomous Networks – Architecture Framework (Y.3061), ITU-T Study Group 13.

Sanjay Yadav, "Optical Network Communications: An Engineer's Perspective" — Bridge the Gap Between Theory and Practice in Optical Networking.