Autonomous networks bring new AI opex costs for telcos

Telecom operators deploying AI agents across network operations face a new operating-cost challenge as automated systems take on more network management tasks, according to research from Bain & Company.

Bain’s warning comes as operators target higher levels of network autonomy. TM Forum said in June that 81% of 80 operators it surveyed are targeting Level 4 autonomous networks or above by 2030, while 20% expect to reach that level by 2027.

At Level 4, networks move beyond predefined automation towards autonomous decision-making within specific network domains. TM Forum’s work around Level 4 includes closed-loop operations in live networks, agent-based architectures, and frameworks for measuring the value generated by autonomous-network deployments.

The cost of higher network autonomy

Bain outlined a scenario in which AI agents and token consumption account for 20% to 30% of a telco’s operating cost base over the next three to five years. Traditional operating expenses would account for the remaining 70% to 80%.

The issue is not simply the price of running AI models. Bain said operators risk creating a dual operating model if AI is added to existing processes without removing the costs associated with those processes.

Traditional teams, software licences, outsourced operations, and existing infrastructure can remain in place while AI compute becomes another operating expense. Bain said avoiding that outcome requires operators to redesign workflows around the work AI takes over, rather than automate individual steps within existing processes.

Its recommendations include reducing manual monitoring and operational hand-offs, adjusting workforce requirements, and reviewing software contracts where agents begin performing tasks previously handled through user-facing applications.

Network operations are among the areas particularly exposed to higher AI usage because of the volume of tasks involved. Bain said increased adoption, more complex reasoning, and the use of newer models can cause overall token consumption to rise even as the price of individual models falls.

Bain said operators should assess AI costs against operational outcomes rather than token consumption alone. In network operations, that means measuring the cost of resolving a network incident rather than focusing only on the cost of model inference.

Autonomous workflows can continuously monitor network conditions, correlate events, diagnose faults, select remediation actions, and verify whether those actions worked. These processes can run across large volumes of network events, adding costs beyond model inference alone.

Inference is only one part of the expense. Agent-based workflows can also require tool calls, orchestration systems, runtime evaluation, observability, storage, and human oversight.

Bain therefore recommends measuring the cost of the completed operational outcome. In a network environment, that can include the total cost of triaging or resolving an incident after inference, orchestration, supporting infrastructure, and human intervention are taken into account.

TM Forum has also introduced metrics aimed at measuring the business value of network autonomy. It approved version 2.0 of its Autonomous Networks High-Value Scenarios Effectiveness Indicators guide in July, with the framework designed to help operators quantify the value of Level 4 autonomous-network scenarios.

Opex savings from closed-loop operations

Bain pointed to Vivo in Brazil as an example of redesigning network processes around AI rather than adding automation to individual tasks. The operator has implemented a self-healing mechanism for its virtualised standalone 5G core as part of Telefónica’s Autonomous Network Journey programme.

The system monitors network function performance, identifies anomalies, diagnoses their root cause, and automatically applies corrective actions. It also checks whether an action resolved the problem and can escalate to another level of remediation when required.

Telefónica said the system correlates events across logical and physical infrastructure and operates without human intervention throughout the process. The company reported that the implementation reduced mean time to resolution for the targeted incidents by 30 minutes.

Telefónica also said the deployment reduces repetitive work and manual intervention and contributes to lower operating costs through more efficient use of computational resources. Telefónica’s published account does not provide a monetary figure for the operating-cost savings associated with the deployment.

Bain said operators should account for AI systems, supporting infrastructure, and remaining human intervention when assessing the cost of an operational outcome. It cited Vivo as an example of redesigning a complete detect-to-resolve workflow rather than automating isolated network tasks.

Bain said operators need to remove or modify existing processes if AI is to replace, rather than supplement, parts of the operating-cost base.

Other operator deployments provide examples of the conventional operating savings associated with higher levels of network autonomy. In a TM Forum case study, China Mobile reported that intelligent agents helped its network operations centre reach Level 4 autonomy, based on a self-assessment using TM Forum’s AN Levels framework, while reducing backend operations and maintenance manpower by more than 30%.

China Mobile also reported savings of more than 5% in frontline installation and maintenance manpower and an average 30% reduction in mean time to repair for faults and customer complaints. The figures were reported by the operator through the TM Forum case study.

An earlier China Mobile autonomous-network case study published by TM Forum reported O&M efficiency improvements of between 10% and 20%, alongside a 30% to 50% reduction in service-opening times. It also reported energy-consumption reductions of between 3% and 5% across internet data centres and base stations covered by the programme.

Bain’s research adds the cost of AI agents, model inference, orchestration, and supporting infrastructure to the calculation operators make when assessing autonomous network operations.

Managing AI costs at telecom scale

AT&T provides an indication of the volume of AI consumption that a large telecom operator can generate, although its figures cover AI workloads across the business rather than autonomous network operations specifically. The operator said in July that it processes an average of 45 billion tokens per day.

AT&T uses an AI gateway that routes tasks between models according to factors including cost, speed, and expected output quality. The operator said the system can change models during multi-turn interactions and has reduced some AI costs by as much as 90%, generating millions of dollars in savings.

The company said only a relatively small share of tasks require its most capable models. Using smaller models and appropriately sized hardware for less complex workloads is one way it has sought to control inference costs.

Bain similarly recommended matching model capability to task complexity rather than using the same model across different workloads. It also said operators should include compute spending in dedicated budgets and track the economics of individual workflows.

Agent behaviour creates another area for cost control. Bain said agents can consume unnecessary resources by repeating context, looping without reaching an outcome, or running overlapping checks.

Bain recommended limits on AI spending and agent runtime, including how long an agent can operate before handing a task to a person. It also advised operators to remove repeated context and consolidate overlapping checks that consume tokens without adding value.

Bain also called for named ownership and financial accountability for individual AI agents. It said autonomous decision-making requires operators to monitor areas including model drift, token spending, and the outcomes produced by deployed agents.

(Photo by Growtika)

See also: ZTE and XLSMART test autonomous network optimisation AI agent

Want to discover how IoT is transforming telecoms and connectivity? Join the IoT Tech Expo in Amsterdam, California, and London. Explore how innovations in 5G, edge computing, and IoT are shaping the future of networks and services. The event is part of TechEx and co-located with other leading technology conferences, click here for more information.

Telecoms Tech News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.