Autonomous networks have been an industry ambition for the better part of a decade. What has changed in 2026 is that operators can now measure progress in a common way, and the measurements show both genuine advance and a clear ceiling. The ceiling is not the algorithms. It is the state of operators’ data, processes and organisational boundaries.
A common yardstick
TM Forum’s framework defines six levels of autonomy, from fully manual (Level 0) to fully autonomous (Level 5)1. Level 4 is the threshold that matters commercially. TM Forum describes it as the point at which networks move from human-defined automation processes to autonomous decision-making, and operators are increasingly declaring and validating it in specific domains2. At Level 4, systems act on intent expressed by people rather than on scripts written by them.
Since 2024, 45 operators have self-assessed at least one network scenario with TM Forum’s Autonomous Network Levels Evaluation Tool (ANLET), and around 40 evaluations by 13 operators have been validated and published1. Bain’s analysis of the results finds that about 55% of assessments focus on fault management across the core, RAN or IP network, and that average scores for the use cases tested rose by about 1.1 points to 3.7 between 2024 and 20251. Operators in Asia lead, with an average score of 3.561.
Operators automate operations far more readily than change
Average ANLET autonomy score by network scenario, 0 (manual) to 5 (fully autonomous), 2024–26
Note: Self-assessed scores from 45 operators; network optimisation has a small sample.
Source: Bain & Company, “Is Your Telco Winning the Race to Autonomous Networks?” (2026)
Where autonomy stalls
The averages flatter the industry. Bain is explicit that most operators remain in the Level 1–2 range once the full scope of network domains and applications is considered1, and NVIDIA’s 2026 industry survey finds 88% of respondents placing their organisations between Levels 1 and 36. The spread across scenarios is revealing: energy-efficiency optimisation, service assurance and network optimisation average 3.9, fault management 2.7 and network change management just 1.91. Change, meaning provisioning, configuration and upgrades across multi-vendor, multi-domain estates, is where legacy OSS, fragmented data and organisational silos bite hardest1.
Omdia’s survey of 80 service providers on transport-network automation tells a similar story. Only 12% have already integrated AI into daily operations and planning, a further 24% expect to do so within 12 months and 45% within one to three years5. The benefits they expect are fewer human errors and faster troubleshooting (48% each) and opex savings (40%); only 16% expect AI to reduce capex5.
From targets to operating models
Group-level targets are now appearing. One large European operator group closed 2025 with 12 Level 4 use cases in operation across three countries, from autonomous remediation of IP-network weaknesses to in-service software upgrades of its 5G core, and has committed to an average autonomy level of 3.75 by 2028 and Level 4 by 20303. Its method is instructive: it measures autonomy by domain (fixed and mobile access, transport, IP, core and telco cloud) and by process (planning, testing, deployment and operations), then rolls the results into a weighted group indicator3.
Agents and intent: the route to Level 4
Agentic AI is the technology most likely to move operators from Level 3 to Level 4, because it can reason across domains rather than execute single-domain playbooks. A TM Forum Catalyst that won the Autonomy Accelerator category of the 2026 Open Innovation Catalyst Awards combined specialised agents for fault correlation, anomaly detection and service and customer impact on a shared data foundation, using business and technical intents such as SLAs, policies and cost and energy targets as the control inputs4. On champion operator data, the team reports up to 60% reductions in mean time to detect and to repair, up to 70% faster fault identification, around 40% fewer incidents escalated to people and 15–20% radio energy savings4.
Cross-domain agents target the hand-offs that slow operations
Improvements reported by a multi-agent NOC Catalyst on operator data, %, ‘up to’ values (%)
Note: Proof-of-concept results reported by the project team; escalation reduction is described as ‘around’ 40%.
Source: TM Forum Inform, “Agentic NOC: AI-native operations for the autonomous telco” (2026)
These are proof-of-concept figures, not audited production results. But they point to where the value sits: in removing cross-domain hand-offs and bridge calls, not in automating individual alarms. Bain’s interviews with leading operators make the complementary point that senior sponsorship, a centralised scaling model and ROI-led metrics matter more than the choice of tools1.
Digital twins as the safety net
The main objection to agentic operations is trust. In Omdia’s transport survey, operators saw strong promise in AI decision-making but were concerned about the reliability and accuracy of agentic AI output5. Network digital twins are emerging as the answer: operators rank network optimisation and traffic engineering (46%) and performance monitoring (43%) as the most suitable use cases for twins, which can validate AI recommendations before they are applied to the live network5. A twin turns an agent’s proposed change into a tested change, which is exactly what the weakest ANLET scenario, change management, lacks.
The investment appetite is there. In NVIDIA’s 2026 survey, AI for autonomous networks was the use case most often cited for return on investment, by 50% of respondents, and 77% expect AI-native networks to launch before 6G is deployed6. The risk is that spending runs ahead of the operating model.
What changes for people
Level 4 does not remove engineers from operations; it changes what they do. In the Catalyst described above, the human role moves from finding and fixing faults to setting intent and governing: people define the policies, risk appetite and guard rails that agents work within, and concentrate on ambiguous cases and new services4. That requires new roles, such as intent owners, agent supervisors and data stewards for inventory and topology, and new performance measures for the teams that hold them. Operators that plan the workforce transition alongside the technology will move faster than those that treat autonomy as a tooling upgrade.
Higher levels of autonomy require more than tools; they demand shifts in process, governance, and culture.