0

Moving up the automation ladder ultimately to closed-loop autonomy begins with data control, advances use case by use case, and is monetized when the network becomes a source of trusted signals

The vendor-side story coming out of DTW Ignite last week in Copenhagen was relatively clear. Agentic AI needs architecture, governed data, orchestration, APIs, observability, policy, security, domain-specific models and closed-loop control. It needs, in other words, a deterministic envelope around probabilistic systems. For more on that nascent convergence, and how Danish philosopher Soren Kierkeegard fits in, read this research note. 

The operator-side story is more interesting because it is less abstract. Operators are not asking whether AI can act. They are asking what it should act on, where it should be allowed to act, who remains accountable when it acts and what measurable value comes back to the business.

TM Forum organized the event around three major mission areas — AI and Data, Autonomous Networks, and Composable IT and Ecosystems — with a stated emphasis on moving from vision to real-world delivery and outcomes. The Autonomous Networks track included sessions on scaling Level 4 networks, the viability of the “Dark NOC” model and the workforce implications of low-touch operations. But the carrier interviews suggest the real industry pivot is  from general enthusiasm about the latest and greatest in AI to disciplined decisions about delegation.

Boost Mobile, Rakuten Mobile and Deutsche Telekom each approach this concept of the agentic network from a different starting point. Boost starts with data ownership and operational control. Rakuten starts with use-case discipline and closed-loop execution. Deutsche Telekom starts with trusted network signals as an enterprise product. Together, they show that the agentic network is an operating model, a data architecture and, increasingly, a monetization strategy.

The first lesson is that data readiness is not the same as data accumulation.

Boost Mobile’s position is unusually direct because of how its network was designed. As Executive Vice President Jeff McSchooler put it, the company made a foundational decision early in the build; the company had to own the data. That meant not allowing critical network information to be trapped inside vendor systems or fragmented across domains. In a cloud-native 5G network, that decision now becomes a strategic asset. The RAN, core and other domains can be correlated because the data was architected with that possibility in mind.

But Boost’s more important point is not that more data is always better. Senior Vice President Dawood Shahdad’s framing is more mature. The company built the “piping” to collect data broadly, then learned that efficiency depends on selecting the right data at the function and domain level. That is the operator problem in miniature. AI-native networking cannot mean shipping everything everywhere, then hoping a model finds the answer. It has to mean understanding which data matters for which decision.

This is where agentic AI becomes less fashionable and more useful. An agent is only as valuable as the decision boundary around it. What data can it see? What action can it take? What does it optimize for? When does it escalate? When does it stop? Those are all operating model questions.

Boost’s north star is a network that “runs itself,” with Shahdad pointing to software management as a practical example. Today, software upgrades still require significant manual choreography. In a more AI-native future, network software should update more like a smartphone, with features pushed automatically, safely and with far less human intervention. The ambition is autonomy, but the route is a pragmatic progression designed to reduce operational drag, free scarce technical talent and redirect human attention toward customer experience, service creation and monetization.

McSchooler’s caveat is the one operators should keep pinned above every AI strategy deck: Where does AI bring value to the business, and where does it not? The industry has a habit of confusing technical possibility with commercial relevance. Boost is making the opposite argument. AI is interesting only where it helps acquire customers, retain customers, improve the network, lower cost, reduce risk or create new revenue.

Rakuten Mobile pushes the same idea further by showing what happens when data becomes action.

Chief AI and Data Officer Sachin Verma’s framing is that valuable AI begins with the use case. That sounds obvious, but it is a needed correction to the way telecom often talks about data. Data is not inherently strategic. Data becomes strategic when it changes a decision. Rakuten’s examples make that concrete. Network-level signals helped Rakuten Card identify potentially fraudulent transaction patterns earlier than the card team could see them, reducing riskier transactions by 33% in a two-week pilot. Aggregated, privacy-preserving mobile payment insights helped Rakuten Pay identify areas where payment activity was happening but merchant coverage was weak, creating a path to more targeted merchant onboarding.

Those are not conventional network operations use cases, but that is why they stand out. Rakuten is not just a mobile operator. It sits inside a broader digital services ecosystem. In that context, the network is not only a connectivity asset and an intelligence layer that can improve adjacent businesses when data moves from detection to decision to action. This is what a platform company is and does. 

The same logic applies inside the network. Rakuten Mobile’s autonomous RAN energy efficiency work, developed with Rakuten Symphony, was validated by TM Forum at Level 4 for the RAN Energy Efficiency Optimization scenario in a live Open RAN environment. Rakuten said the solution uses cloud-native Open RAN architecture and AI to deliver around 20% RAN energy conservation through autonomous operations, and TM Forum characterized Level 4 autonomy as intent-driven, cross-domain, AI/ML-enabled closed-loop management with minimal human oversight.

This is the right kind of proof point because it is bounded. It does not claim the whole network is autonomous. It claims a specific domain, a specific scenario and a specific business outcome. Verma described the practical mechanism as RIC-hosted rApps that use machine learning to predict radio site behavior, decide when the network should be on or off and execute without a human taking action. That is autonomy as an operating condition.

The boundary is just as important as the achievement. Verma was explicit that Rakuten is still working toward broader end-to-end autonomy. Detection and root cause analysis are relatively mature. Remedial action remains harder. That is the operator reality. Level 4 in one use case does not mean Level 4 everywhere. The agentic network will be assembled closed loop by closed loop, with measurable value as the forcing function.

Deutsche Telekom widens the frame from operating the network to monetizing what only the network knows.

SVP Magenta API Capability Exposure

Deutsche Telekom Senior Vice President of Magenta API Capability Exposure Chathurangi Wickramasinghe’s API argument starts with a blunt commercial observation: “No customer wakes up in the morning and asks for APIs.” Enterprises do not buy standardization. They buy less fraud, better authentication, lower integration friction, trusted customer engagement and more resilient digital workflows.

That reframes the network API discussion. The early industry work around CAMARA, GSMA Open Gateway and TM Forum was about exposure and standardization. That work remains necessary. GSMA’s Open Gateway API catalog includes capabilities such as Number Verification, which verifies the phone number of a connected device, and SIM Swap, which provides information on SIM pairing changes. Deutsche Telekom’s MagentaBusiness API launch in Germany, powered by Vonage, made APIs such as Quality-on-Demand, Device Status-Roaming and Device Location commercially available to developers and enterprises. In 2024, Deutsche Telekom, Vodafone and O2 Telefónica launched Number Verify in Germany under the GSMA Open Gateway initiative to help protect digital identities and reduce dependence on SMS-based two-factor authentication; Deutsche Telekom also positioned SIM Swap as a way for financial institutions to check whether a number has recently changed SIM cards before approving a transaction.

The AI connection is that autonomous enterprise workflows will need trusted inputs. If an AI agent is approving a transaction, changing a customer account, triggering a service interaction or orchestrating a business process, it needs to know whether the user is legitimate, whether the SIM has changed, whether the device is where it claims to be and whether the interaction can be trusted. The network can answer some of those questions in ways an application layer cannot.

That is a monetization thesis hiding inside a trust thesis. For years, operators have tried to move beyond connectivity by exposing capabilities to developers. The harder question has always been what developers and enterprises will actually pay for. Deutsche Telekom’s answer is that they will pay for outcomes. In an AI context, the outcome is not “an API.” It is a verified signal that makes an autonomous workflow safer, faster or more reliable.

That connects the three operator perspectives into one argument. Boost is focused on knowing what data matters and retaining control over it. Rakuten is focused on turning the right data into bounded autonomous action. Deutsche Telekom is focused on exposing trusted network intelligence into enterprise workflows. All three are, in different ways, moving away from AI as a horizontal capability and toward AI as a disciplined system of rights, responsibilities and outcomes.

This is also why the people issue keeps returning. Boost talked about trust inside the company and the need for employees to believe AI is the right thing for their future. Shahdad emphasized meaningful human work and oversight during the transition. Verma made a similar point, arguing that the diffusion of AI inside enterprises remains one of the biggest adoption barriers. The technology is moving quickly; enterprise absorption is largely not.

That may be the most important operator-side corrective to the hype cycle. The agentic network is not a “dark NOC” press release. It is a long negotiation between automation and accountability. Humans will not disappear from network operations. Their work will move toward defining intent, setting policy, supervising exceptions, validating outcomes, managing risk and deciding where autonomy is appropriate.

The deterministic telecom instinct is still there, and it should be. Networks carry emergency calls, financial transactions, industrial processes and customer trust. Operators cannot simply hand those systems over to probabilistic agents and hope for the best. But they also cannot ignore the operational complexity they have created. Cloud-native infrastructure, Open RAN, multi-vendor systems, network APIs, edge compute, private networks and AI-enabled services all increase the number of decisions that have to be made. At some point, human-speed operations become the constraint making machine-speed operations the imperative.

The operator answer is bounded delegation. Let agents handle the data they are allowed to see. Let them act in domains where the business case is measurable and the failure modes are understood. Let them recommend before they remediate, remediate before they optimize across domains and optimize across domains only when governance, observability and rollback mechanisms are mature. Let APIs expose network truth to enterprise systems, but make the value proposition about fraud prevention, authentication, quality assurance and trust rather than telecom complexity.

The user perspective from DTW is therefore less spectacular than the vendor perspective, but more useful. Vendors are building the machinery of agentic telecom. Operators are deciding where that machinery earns the right to run. The agentic network will not arrive as a single architecture or maturity level. It will arrive when operators can answer four questions with confidence. Do we control the data? Do we know which decision this data should inform? Can the system act safely without a human in the loop? Does the action create measurable value?