The AI telecom operating model

Prove the value across interacting control loops

An energy controller puts a radio resource to sleep while a service controller adds capacity nearby. Each can satisfy its own objective while the combined network becomes more expensive or less reliable. This illustrative conflict captures the investment challenge for AI in telecoms: value must survive interaction across the operating boundary, not merely appear inside one control loop.

Start with a domain where an accountable owner can connect an action to an observable outcome. Specify the baseline, controllable decision and service constraints. A prediction creates value only if someone can act on it in time; an automated action creates value only if its wider effects remain acceptable. These links should define the investment case.

The industry's standards work supports a systems perspective. ETSI's zero-touch network and service management programme addresses end-to-end automation across domains and layers. ¹ It describes an architectural direction, not evidence that any particular operator has achieved reliable full autonomy. The CEO should therefore ask how the proposed investment changes an operating decision today and what additional conditions must hold before that decision can be delegated more broadly.

Separate three sources of economic value

Network efficiency concerns capacity, energy and fault response within service obligations. Define the domain and observation period, recording traffic, weather, configuration and service conditions. Lower energy expenditure may reflect tariffs or demand rather than better optimization. Test transfer to different equipment and operating conditions before multiplying a local result across the estate.

The second source is service productivity. AI may help staff diagnose issues, find policy information or coordinate a field visit. Measure resolution and repeat contact, not just the speed of the first response. A chatbot that reduces contact-centre volume by making escalation difficult can look efficient while worsening customer outcomes. Similarly, a faster field schedule can fail economically if it increases repeat visits or allocates the wrong skills. The complete service chain matters more than the local task metric.

The third source is revenue and investment quality. Better demand forecasting, enterprise solution design or retention targeting may improve commercial decisions. These claims need careful attribution because pricing, competition and product changes occur simultaneously. Where controlled trials are feasible and appropriate, use them. Where they are not, disclose the weaker inference. The enterprise should avoid combining speculative revenue uplift with well-measured cost savings in one headline figure that conceals the different confidence levels.

Each operating mode needs different evidence

mayden proposes four operating modes. Observation establishes evidence and a baseline. Recommendation tests whether advice changes useful decisions. Bounded execution delegates a narrow action with exposure limits and rollback. Cross-domain coordination governs interactions among loops. Moving between them changes the proof obligation; it is not a prescribed climb toward maximum autonomy.

ETSI's January 2026 study of agents in autonomous networks distinguishes resource, policy and goal conflicts, and discusses delegation, escalation and arbitration. ⁸ This is a standards-group study, not evidence that multi-agent networks deliver a particular saving. Its practical contribution is to make coordination a separate design obligation: capable agents can still pursue incompatible objectives on the same network.

For the energy-and-service example, define which constraints prevail, who arbitrates conflicting proposals and how the combined action is reversed. Test that conflict before connecting autonomous domains. Measure the shared outcome, including neighboring cells: adding local improvements does not establish network value.

ETSI's 2026 report on closed-loop security examines risks across monitoring, analysis, decision and execution, including interactions among loops. ² The 3GPP Release 18 description documents AI/ML and management-related work within the wider mobile-system evolution. ³ Neither source is a benchmark of commercial deployment quality. Their practical implication is architectural: procurement should examine interfaces, lifecycle management and control interaction, not merely model accuracy or a vendor's claim to autonomy.

Keep standards, research and vendor evidence in their proper roles

Standards provide common concepts and interfaces. Research identifies possibilities and failure modes. A supplier demonstration shows what occurred under specified conditions. An operator's measured production trial provides evidence about its own environment. These categories complement one another, but they should not be substituted. A claim that a product follows a standard does not establish its financial return, and a strong trial result does not establish interoperability with every future system.

The ITU focus group on AI-native telecommunications networks was established to explore architectural requirements, challenges and opportunities. ⁴ Its existence is evidence that the architecture remains an active area of work. It is not proof that the industry has settled one complete operating model. Operators should preserve the ability to evolve components and avoid committing core control authority to interfaces they cannot inspect, test or replace. The cost of integration and exit belongs in the investment case.

Require vendors to identify the baseline, population, operating conditions and period behind performance claims. Separate simulation from live trials and customer reports from independent verification. Ask for logs and reproducible failure conditions. Where disclosure is constrained, narrow the claim and the delegation it can support.

Energy is a real opportunity, but emissions are not an AI score

The GSMA's 2026 report records a 13% reduction in operators' operational emissions between 2019 and 2024 while connections grew 10% and traffic more than quadrupled. Its coverage includes 116 operators representing 85% of mobile connections. ⁵ These figures describe an industry trend assembled from disclosures. They do not isolate the contribution of AI. Changes in energy sourcing, equipment, traffic mix and other operating practices can contribute to the observed result.

For an operator, the relevant trial should distinguish electricity saved from emissions accounting and financial savings. Electricity is a physical quantity; the bill depends on tariffs and contracts; reported emissions depend on the accounting method and energy sources. A reduction in one does not establish the same percentage reduction in the others. Keep them in separate lines of the business case. Where a network intervention shifts demand to another site or increases computing elsewhere, include those effects in the chosen boundary.

mayden recommends an energy-and-service acceptance test. Specify the permitted change in service quality before running the optimization. Compare energy, coverage, latency, reliability and customer complaints over representative operating conditions. Include the additional inference and orchestration cost. An intervention that saves energy while degrading a critical service is not automatically a good trade. The operator must decide the trade-off explicitly, within its obligations, rather than letting an optimization objective make that decision invisibly.

Resilience is the condition for economic scale

More automation can reduce some human errors while increasing the speed and reach of a mistaken action. That does not justify avoiding automation; it changes the requirements for containment. A control loop should have bounded permissions, observable actions and a clear way to suspend or revert changes. The operator also needs staff who understand the fallback. A manual procedure that has not been exercised may not work when many sites or customers are affected at once.

Ofcom's network and service resilience guidance provides a UK-specific regulatory context for communications providers. ⁶ Its relevance here is the focus on resilient operation, not a universal rulebook for every telecom market. The NCSC's AI guidance separately reinforces lifecycle security. ⁷ mayden's operational extension is to test the full recovery path: detect anomalous behavior, stop further actions, restore an acceptable configuration and communicate with the affected service teams. A model rollback alone may not reverse changes already executed elsewhere.

Dependency concentration deserves board attention. Several domains may depend on the same data platform, orchestration layer or external model. Individually reasonable project choices can create an estate-wide dependency that no project owner sees. Maintain a portfolio view of shared services and correlated failure. Test degraded modes under constrained capacity, including the increased workload on people. The resilience case should describe what service remains available and for how long, rather than offer an undifferentiated claim that a backup exists.

Design the trial so neighboring cells cannot hide the cost

In an illustrative radio-cluster energy trial, include affected neighboring cells in the measurement boundary. A sleeping resource can move traffic and energy use elsewhere. Define the cluster, adjacent resources, observation windows and service stop conditions before intervention. Otherwise the trial can reward cost displacement rather than improvement.

Make the interaction concrete: an energy controller requests that a resource sleep while a service-quality controller requests additional capacity in an adjacent cell. Neither local objective should silently override the other. A named coordinating function applies the agreed service constraints and permits, defers or reverses the action. Record both proposals, the governing rule and the observed result across the shared boundary. If coordination is unavailable, use the previously approved restricted mode rather than let competing loops improvise a new priority. This is an illustrative test condition, not a claim about a deployed operator architecture.

Where feasible, compare intervention and reference periods or clusters with similar traffic and equipment. Record events that break comparability, including maintenance, special events, software upgrades and unusual weather. A perfectly matched control may be unavailable, in which case the report should state the remaining attribution limits. Do not discard adverse periods merely because they complicate the result. Network leaders need to know whether an optimization is useful under the demand conditions that make their service difficult to operate.

The acceptance record should then separate three decisions. First, whether the energy effect is credible within the tested boundary. Second, whether service performance remains acceptable, including at the distribution's extremes. Third, whether the financial return survives implementation, computing and support costs. Passing one test does not imply passing the others. Expansion to another equipment family or traffic pattern is a transfer question that requires additional evidence. This approach gives the CTO and CFO a shared way to challenge a vendor's savings claim while preserving a route to scale when the measured result is persuasive.

Tie the next delegation to the next investment

Some value may require simultaneous change across several domains. Preserve that strategic hypothesis while testing the next commitment. Prove local control, interfaces and arbitration before connecting their authority. Premature integration can make both attribution and recovery harder.

Choose one material service or cost problem and an accountable domain owner. Agree the baseline, intervention and acceptance criteria with finance, operations, security and customer teams. Where appropriate, start in recommendation mode and record rejected advice. Grant limited execution when the evidence supports it, rather than treating advisory accuracy as permission to act.

At the investment gate, reconcile observed benefit with implementation, computing, support, exceptions and recovery capacity. Test sensitivity to traffic, supplier price and equipment differences. Then scale, redesign, retain supervision or stop. The chosen delegation should follow demonstrated service and economic performance.